vf/vf_react
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update pay ui

Browse Source
This commit is contained in:
zzlgreat
2025-12-10 17:18:58 +08:00
parent beb349ac2f
commit 85c29483dd
25 changed files with 133 additions and 10246 deletions
Download Patch File Download Diff File Expand all files Collapse all files

112
ml/README.md
View File

@@ -1,112 +0,0 @@
# 概念异动检测 ML 模块
基于 Transformer Autoencoder 的概念异动检测系统。
## 环境要求
- Python 3.8+
- PyTorch 2.0+ (CUDA 12.x for 5090 GPU)
- ClickHouse, MySQL, Elasticsearch
## 数据库配置
当前配置（`prepare_data.py`）:
- MySQL: `192.168.1.5:3306`
- Elasticsearch: `127.0.0.1:9200`
- ClickHouse: `127.0.0.1:9000`
## 快速开始
```bash
# 1. 安装依赖
pip install -r ml/requirements.txt
# 2. 安装 PyTorch (5090 需要 CUDA 12.4)
pip install torch --index-url https://download.pytorch.org/whl/cu124
# 3. 运行训练
chmod +x ml/run_training.sh
./ml/run_training.sh
```
## 文件说明
| 文件 | 说明 |
|------|------|
| `model.py` | Transformer Autoencoder 模型定义 |
| `prepare_data.py` | 数据提取和特征计算 |
| `train.py` | 模型训练脚本 |
| `inference.py` | 推理服务 |
| `enhanced_detector.py` | 增强版检测器（融合 Alpha + ML） |
## 训练参数
```bash
# 完整参数
./ml/run_training.sh --start 2022-01-01 --end 2024-12-01 --epochs 100 --batch_size 256
# 只准备数据
python ml/prepare_data.py --start 2022-01-01
# 只训练（数据已准备好）
python ml/train.py --epochs 100 --batch_size 256 --lr 1e-4
```
## 模型架构
```
输入: (batch, 30, 6) # 30分钟序列，6个特征
↓
Positional Encoding
↓
Transformer Encoder (4层, 8头, d=128)
↓
Bottleneck (压缩到 32 维)
↓
Transformer Decoder (4层)
↓
输出: (batch, 30, 6) # 重构序列
异动判断: reconstruction_error > threshold
```
## 6维特征
1. `alpha` - 超额收益（概念涨幅 - 大盘涨幅）
2. `alpha_delta` - Alpha 5分钟变化
3. `amt_ratio` - 成交额 / 20分钟均值
4. `amt_delta` - 成交额变化率
5. `rank_pct` - Alpha 排名百分位
6. `limit_up_ratio` - 涨停股占比
## 训练产出
训练完成后，`ml/checkpoints/` 包含:
- `best_model.pt` - 最佳模型权重
- `thresholds.json` - 异动阈值 (P90/P95/P99)
- `normalization_stats.json` - 数据标准化参数
- `config.json` - 训练配置
## 使用示例
```python
from ml.inference import ConceptAnomalyDetector
detector = ConceptAnomalyDetector('ml/checkpoints')
# 实时检测
is_anomaly, score = detector.detect(
concept_name="人工智能",
features={
'alpha': 2.5,
'alpha_delta': 0.8,
'amt_ratio': 1.5,
'amt_delta': 0.3,
'rank_pct': 0.95,
'limit_up_ratio': 0.15,
}
)
if is_anomaly:
print(f"检测到异动！分数: {score}")
```

10
ml/__init__.py
View File

@@ -1,10 +0,0 @@
# -*- coding: utf-8 -*-
"""
概念异动检测 ML 模块
提供基于 Transformer Autoencoder 的异动检测功能
"""
from .inference import ConceptAnomalyDetector, MLAnomalyService
__all__ = ['ConceptAnomalyDetector', 'MLAnomalyService']

BIN
ml/__pycache__/realtime_detector.cpython-310.pyc
View File

Binary file not shown.

481
ml/backtest.py
View File

@@ -1,481 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
历史异动回测脚本
使用训练好的模型，对历史数据进行异动检测，生成异动记录
使用方法：
# 回测指定日期范围
python backtest.py --start 2024-01-01 --end 2024-12-01
# 回测单天
python backtest.py --start 2024-11-01 --end 2024-11-01
# 只生成结果，不写入数据库
python backtest.py --start 2024-01-01 --dry-run
"""
import os
import sys
import argparse
import json
from datetime import datetime, timedelta
from pathlib import Path
from typing import Dict, List, Tuple, Optional
from collections import defaultdict
import numpy as np
import pandas as pd
import torch
from tqdm import tqdm
from sqlalchemy import create_engine, text
# 添加父目录到路径
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from model import TransformerAutoencoder
# ==================== 配置 ====================
MYSQL_ENGINE = create_engine(
"mysql+pymysql://root:Zzl5588161!@192.168.1.5:3306/stock",
echo=False
)
# 特征列表（与训练一致）
FEATURES = [
'alpha',
'alpha_delta',
'amt_ratio',
'amt_delta',
'rank_pct',
'limit_up_ratio',
]
# 回测配置
BACKTEST_CONFIG = {
'seq_len': 30, # 序列长度
'threshold_key': 'p95', # 使用的阈值
'min_alpha_abs': 0.5, # 最小 Alpha 绝对值（过滤微小波动）
'cooldown_minutes': 8, # 同一概念冷却时间
'max_alerts_per_minute': 15, # 每分钟最多异动数
'clip_value': 10.0, # 极端值截断
}
# ==================== 模型加载 ====================
class AnomalyDetector:
"""异动检测器"""
def __init__(self, checkpoint_dir: str = 'ml/checkpoints', device: str = 'auto'):
self.checkpoint_dir = Path(checkpoint_dir)
# 设备
if device == 'auto':
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
else:
self.device = torch.device(device)
# 加载配置
self._load_config()
# 加载模型
self._load_model()
# 加载阈值
self._load_thresholds()
print(f"AnomalyDetector 初始化完成")
print(f" 设备: {self.device}")
print(f" 阈值 ({BACKTEST_CONFIG['threshold_key']}): {self.threshold:.6f}")
def _load_config(self):
config_path = self.checkpoint_dir / 'config.json'
with open(config_path, 'r') as f:
self.config = json.load(f)
def _load_model(self):
model_path = self.checkpoint_dir / 'best_model.pt'
checkpoint = torch.load(model_path, map_location=self.device)
model_config = self.config['model'].copy()
model_config['use_instance_norm'] = self.config.get('use_instance_norm', True)
self.model = TransformerAutoencoder(**model_config)
self.model.load_state_dict(checkpoint['model_state_dict'])
self.model.to(self.device)
self.model.eval()
def _load_thresholds(self):
thresholds_path = self.checkpoint_dir / 'thresholds.json'
with open(thresholds_path, 'r') as f:
thresholds = json.load(f)
self.threshold = thresholds[BACKTEST_CONFIG['threshold_key']]
@torch.no_grad()
def compute_anomaly_scores(self, sequences: np.ndarray) -> np.ndarray:
"""
计算异动分数
Args:
sequences: (n_sequences, seq_len, n_features)
Returns:
scores: (n_sequences,) 每个序列最后时刻的异动分数
"""
# 截断极端值
sequences = np.clip(sequences, -BACKTEST_CONFIG['clip_value'], BACKTEST_CONFIG['clip_value'])
# 转为 tensor
x = torch.FloatTensor(sequences).to(self.device)
# 计算重构误差
errors = self.model.compute_reconstruction_error(x, reduction='none')
# 取最后一个时刻的误差
scores = errors[:, -1].cpu().numpy()
return scores
def is_anomaly(self, score: float) -> bool:
"""判断是否异动"""
return score > self.threshold
# ==================== 数据加载 ====================
def load_daily_features(data_dir: str, date: str) -> Optional[pd.DataFrame]:
"""加载单天的特征数据"""
file_path = Path(data_dir) / f"features_{date}.parquet"
if not file_path.exists():
return None
df = pd.read_parquet(file_path)
return df
def get_available_dates(data_dir: str, start_date: str, end_date: str) -> List[str]:
"""获取可用的日期列表"""
data_path = Path(data_dir)
all_files = sorted(data_path.glob("features_*.parquet"))
dates = []
for f in all_files:
date = f.stem.replace('features_', '')
if start_date <= date <= end_date:
dates.append(date)
return dates
# ==================== 回测逻辑 ====================
def backtest_single_day(
detector: AnomalyDetector,
df: pd.DataFrame,
date: str,
seq_len: int = 30
) -> List[Dict]:
"""
回测单天数据
Args:
detector: 异动检测器
df: 当天的特征数据
date: 日期
seq_len: 序列长度
Returns:
alerts: 异动列表
"""
alerts = []
# 按概念分组
grouped = df.groupby('concept_id', sort=False)
# 冷却记录 {concept_id: last_alert_timestamp}
cooldown = {}
# 获取所有时间点
all_timestamps = sorted(df['timestamp'].unique())
if len(all_timestamps) < seq_len:
return alerts
# 对每个时间点进行检测（从第 seq_len 个开始）
for t_idx in range(seq_len - 1, len(all_timestamps)):
current_time = all_timestamps[t_idx]
window_start_time = all_timestamps[t_idx - seq_len + 1]
minute_alerts = []
# 收集该时刻所有概念的序列
concept_sequences = []
concept_infos = []
for concept_id, concept_df in grouped:
# 获取该概念在时间窗口内的数据
mask = (concept_df['timestamp'] >= window_start_time) & (concept_df['timestamp'] <= current_time)
window_df = concept_df[mask].sort_values('timestamp')
if len(window_df) < seq_len:
continue
# 取最后 seq_len 个点
window_df = window_df.tail(seq_len)
# 提取特征
features = window_df[FEATURES].values
# 处理缺失值
features = np.nan_to_num(features, nan=0.0, posinf=0.0, neginf=0.0)
# 获取当前时刻的信息
current_row = window_df.iloc[-1]
concept_sequences.append(features)
concept_infos.append({
'concept_id': concept_id,
'timestamp': current_time,
'alpha': current_row.get('alpha', 0),
'alpha_delta': current_row.get('alpha_delta', 0),
'amt_ratio': current_row.get('amt_ratio', 1),
'limit_up_ratio': current_row.get('limit_up_ratio', 0),
'limit_down_ratio': current_row.get('limit_down_ratio', 0),
'rank_pct': current_row.get('rank_pct', 0.5),
'stock_count': current_row.get('stock_count', 0),
'total_amt': current_row.get('total_amt', 0),
})
if not concept_sequences:
continue
# 批量计算异动分数
sequences_array = np.array(concept_sequences)
scores = detector.compute_anomaly_scores(sequences_array)
# 检测异动
for i, (info, score) in enumerate(zip(concept_infos, scores)):
concept_id = info['concept_id']
alpha = info['alpha']
# 过滤小波动
if abs(alpha) < BACKTEST_CONFIG['min_alpha_abs']:
continue
# 检查冷却
if concept_id in cooldown:
last_alert = cooldown[concept_id]
if isinstance(current_time, datetime):
time_diff = (current_time - last_alert).total_seconds() / 60
else:
# timestamp 是字符串或其他格式
time_diff = BACKTEST_CONFIG['cooldown_minutes'] + 1 # 跳过冷却检查
if time_diff < BACKTEST_CONFIG['cooldown_minutes']:
continue
# 判断是否异动
if not detector.is_anomaly(score):
continue
# 记录异动
alert_type = 'surge_up' if alpha > 0 else 'surge_down'
alert = {
'concept_id': concept_id,
'alert_time': current_time,
'trade_date': date,
'alert_type': alert_type,
'anomaly_score': float(score),
'threshold': detector.threshold,
**info
}
minute_alerts.append(alert)
cooldown[concept_id] = current_time
# 按分数排序，限制数量
minute_alerts.sort(key=lambda x: x['anomaly_score'], reverse=True)
alerts.extend(minute_alerts[:BACKTEST_CONFIG['max_alerts_per_minute']])
return alerts
# ==================== 数据库写入 ====================
def save_alerts_to_mysql(alerts: List[Dict], dry_run: bool = False) -> int:
"""保存异动到 MySQL"""
if not alerts:
return 0
if dry_run:
print(f" [Dry Run] 将写入 {len(alerts)} 条异动")
return len(alerts)
saved = 0
with MYSQL_ENGINE.begin() as conn:
for alert in alerts:
try:
# 检查是否已存在
check_sql = text("""
SELECT id FROM concept_minute_alert
WHERE concept_id = :concept_id
AND alert_time = :alert_time
AND trade_date = :trade_date
""")
exists = conn.execute(check_sql, {
'concept_id': alert['concept_id'],
'alert_time': alert['alert_time'],
'trade_date': alert['trade_date'],
}).fetchone()
if exists:
continue
# 插入新记录
insert_sql = text("""
INSERT INTO concept_minute_alert
(concept_id, concept_name, alert_time, alert_type, trade_date,
change_pct, zscore, importance_score, stock_count, extra_info)
VALUES
(:concept_id, :concept_name, :alert_time, :alert_type, :trade_date,
:change_pct, :zscore, :importance_score, :stock_count, :extra_info)
""")
conn.execute(insert_sql, {
'concept_id': alert['concept_id'],
'concept_name': alert.get('concept_name', ''),
'alert_time': alert['alert_time'],
'alert_type': alert['alert_type'],
'trade_date': alert['trade_date'],
'change_pct': alert.get('alpha', 0),
'zscore': alert['anomaly_score'],
'importance_score': alert['anomaly_score'],
'stock_count': alert.get('stock_count', 0),
'extra_info': json.dumps({
'detection_method': 'ml_autoencoder',
'threshold': alert['threshold'],
'alpha': alert.get('alpha', 0),
'amt_ratio': alert.get('amt_ratio', 1),
}, ensure_ascii=False)
})
saved += 1
except Exception as e:
print(f" 保存失败: {alert['concept_id']} - {e}")
return saved
def export_alerts_to_csv(alerts: List[Dict], output_path: str):
"""导出异动到 CSV"""
if not alerts:
return
df = pd.DataFrame(alerts)
df.to_csv(output_path, index=False, encoding='utf-8-sig')
print(f"已导出到: {output_path}")
# ==================== 主函数 ====================
def main():
parser = argparse.ArgumentParser(description='历史异动回测')
parser.add_argument('--data_dir', type=str, default='ml/data',
help='特征数据目录')
parser.add_argument('--checkpoint_dir', type=str, default='ml/checkpoints',
help='模型检查点目录')
parser.add_argument('--start', type=str, required=True,
help='开始日期 (YYYY-MM-DD)')
parser.add_argument('--end', type=str, required=True,
help='结束日期 (YYYY-MM-DD)')
parser.add_argument('--dry-run', action='store_true',
help='只计算，不写入数据库')
parser.add_argument('--export-csv', type=str, default=None,
help='导出 CSV 文件路径')
parser.add_argument('--device', type=str, default='auto',
help='设备 (auto/cuda/cpu)')
args = parser.parse_args()
print("=" * 60)
print("历史异动回测")
print("=" * 60)
print(f"日期范围: {args.start} ~ {args.end}")
print(f"数据目录: {args.data_dir}")
print(f"模型目录: {args.checkpoint_dir}")
print(f"Dry Run: {args.dry_run}")
print("=" * 60)
# 初始化检测器
detector = AnomalyDetector(args.checkpoint_dir, args.device)
# 获取可用日期
dates = get_available_dates(args.data_dir, args.start, args.end)
if not dates:
print(f"未找到 {args.start} ~ {args.end} 范围内的数据")
return
print(f"\n找到 {len(dates)} 天的数据")
# 回测
all_alerts = []
total_saved = 0
for date in tqdm(dates, desc="回测进度"):
# 加载数据
df = load_daily_features(args.data_dir, date)
if df is None or df.empty:
continue
# 回测单天
alerts = backtest_single_day(
detector, df, date,
seq_len=BACKTEST_CONFIG['seq_len']
)
if alerts:
all_alerts.extend(alerts)
# 写入数据库
saved = save_alerts_to_mysql(alerts, dry_run=args.dry_run)
total_saved += saved
if not args.dry_run:
tqdm.write(f" {date}: 检测到 {len(alerts)} 个异动，保存 {saved} 条")
# 导出 CSV
if args.export_csv and all_alerts:
export_alerts_to_csv(all_alerts, args.export_csv)
# 汇总
print("\n" + "=" * 60)
print("回测完成！")
print("=" * 60)
print(f"总计检测到: {len(all_alerts)} 个异动")
print(f"保存到数据库: {total_saved} 条")
# 统计
if all_alerts:
df_alerts = pd.DataFrame(all_alerts)
print(f"\n异动类型分布:")
print(df_alerts['alert_type'].value_counts())
print(f"\n异动分数统计:")
print(f" Mean: {df_alerts['anomaly_score'].mean():.4f}")
print(f" Max: {df_alerts['anomaly_score'].max():.4f}")
print(f" Min: {df_alerts['anomaly_score'].min():.4f}")
print("=" * 60)
if __name__ == "__main__":
main()

859
ml/backtest_fast.py
View File

@@ -1,859 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
快速融合异动回测脚本
优化策略：
1. 预先构建所有序列（向量化），避免循环内重复切片
2. 批量 ML 推理（一次推理所有候选）
3. 使用 NumPy 向量化操作替代 Python 循环
性能对比：
- 原版：5分钟/天
- 优化版：预计 10-30秒/天
"""
import os
import sys
import argparse
import json
from datetime import datetime
from pathlib import Path
from typing import Dict, List, Optional, Tuple
from collections import defaultdict
import numpy as np
import pandas as pd
import torch
from tqdm import tqdm
from sqlalchemy import create_engine, text
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
# ==================== 配置 ====================
MYSQL_ENGINE = create_engine(
"mysql+pymysql://root:Zzl5588161!@192.168.1.5:3306/stock",
echo=False
)
FEATURES = ['alpha', 'alpha_delta', 'amt_ratio', 'amt_delta', 'rank_pct', 'limit_up_ratio']
CONFIG = {
'seq_len': 15, # 序列长度（支持跨日后可从 9:30 检测）
'min_alpha_abs': 0.3, # 最小 alpha 过滤
'cooldown_minutes': 8,
'max_alerts_per_minute': 20,
'clip_value': 10.0,
# === 融合权重：均衡 ===
'rule_weight': 0.5,
'ml_weight': 0.5,
# === 触发阈值 ===
'rule_trigger': 65, # 60 -> 65，略提高规则门槛
'ml_trigger': 70, # 75 -> 70，略降低 ML 门槛
'fusion_trigger': 45,
}
# ==================== 规则评分（向量化版）====================
def get_size_adjusted_thresholds(stock_count: np.ndarray) -> dict:
"""
根据概念股票数量计算动态阈值
设计思路：
- 小概念（<10 只）：波动大是正常的，需要更高阈值
- 中概念（10-50 只）：标准阈值
- 大概念（>50 只）：能有明显波动说明是真异动，降低阈值
返回各指标的调整系数（乘以基准阈值）
"""
n = len(stock_count)
# 基于股票数量的调整系数
# 小概念：系数 > 1（提高阈值，更难触发）
# 大概念：系数 < 1（降低阈值，更容易触发）
size_factor = np.ones(n)
# 微型概念（<5 只）：阈值 × 1.8
tiny = stock_count < 5
size_factor[tiny] = 1.8
# 小概念（5-10 只）：阈值 × 1.4
small = (stock_count >= 5) & (stock_count < 10)
size_factor[small] = 1.4
# 中小概念（10-20 只）：阈值 × 1.2
medium_small = (stock_count >= 10) & (stock_count < 20)
size_factor[medium_small] = 1.2
# 中概念（20-50 只）：标准阈值 × 1.0
medium = (stock_count >= 20) & (stock_count < 50)
size_factor[medium] = 1.0
# 大概念（50-100 只）：阈值 × 0.85
large = (stock_count >= 50) & (stock_count < 100)
size_factor[large] = 0.85
# 超大概念（>100 只）：阈值 × 0.7
xlarge = stock_count >= 100
size_factor[xlarge] = 0.7
return size_factor
def score_rules_batch(df: pd.DataFrame) -> Tuple[np.ndarray, List[List[str]]]:
"""
批量计算规则得分（向量化）- 考虑概念规模版
设计原则：
- 规则作为辅助信号，不应单独主导决策
- 根据概念股票数量动态调整阈值
- 大概念异动更有价值，小概念需要更大波动才算异动
Args:
df: DataFrame，包含所有特征列（必须包含 stock_count）
Returns:
scores: (n,) 规则得分数组
triggered_rules: 每行触发的规则列表
"""
n = len(df)
scores = np.zeros(n)
triggered = [[] for _ in range(n)]
alpha = df['alpha'].values
alpha_delta = df['alpha_delta'].values
amt_ratio = df['amt_ratio'].values
amt_delta = df['amt_delta'].values
rank_pct = df['rank_pct'].values
limit_up_ratio = df['limit_up_ratio'].values
stock_count = df['stock_count'].values if 'stock_count' in df.columns else np.full(n, 20)
alpha_abs = np.abs(alpha)
alpha_delta_abs = np.abs(alpha_delta)
# 获取基于规模的调整系数
size_factor = get_size_adjusted_thresholds(stock_count)
# ========== Alpha 规则（动态阈值）==========
# 基准阈值：极强 5%，强 4%，中等 3%
# 实际阈值 = 基准 × size_factor
# 极强信号
alpha_extreme_thresh = 5.0 * size_factor
mask = alpha_abs >= alpha_extreme_thresh
scores[mask] += 20
for i in np.where(mask)[0]: triggered[i].append('alpha_extreme')
# 强信号
alpha_strong_thresh = 4.0 * size_factor
mask = (alpha_abs >= alpha_strong_thresh) & (alpha_abs < alpha_extreme_thresh)
scores[mask] += 15
for i in np.where(mask)[0]: triggered[i].append('alpha_strong')
# 中等信号
alpha_medium_thresh = 3.0 * size_factor
mask = (alpha_abs >= alpha_medium_thresh) & (alpha_abs < alpha_strong_thresh)
scores[mask] += 10
for i in np.where(mask)[0]: triggered[i].append('alpha_medium')
# ========== Alpha 加速度规则（动态阈值）==========
delta_strong_thresh = 2.0 * size_factor
mask = alpha_delta_abs >= delta_strong_thresh
scores[mask] += 15
for i in np.where(mask)[0]: triggered[i].append('alpha_delta_strong')
delta_medium_thresh = 1.5 * size_factor
mask = (alpha_delta_abs >= delta_medium_thresh) & (alpha_delta_abs < delta_strong_thresh)
scores[mask] += 10
for i in np.where(mask)[0]: triggered[i].append('alpha_delta_medium')
# ========== 成交额规则（不受规模影响，放量就是放量）==========
mask = amt_ratio >= 10.0
scores[mask] += 20
for i in np.where(mask)[0]: triggered[i].append('volume_extreme')
mask = (amt_ratio >= 6.0) & (amt_ratio < 10.0)
scores[mask] += 12
for i in np.where(mask)[0]: triggered[i].append('volume_strong')
# ========== 排名规则 ==========
mask = rank_pct >= 0.98
scores[mask] += 15
for i in np.where(mask)[0]: triggered[i].append('rank_top')
mask = rank_pct <= 0.02
scores[mask] += 15
for i in np.where(mask)[0]: triggered[i].append('rank_bottom')
# ========== 涨停规则（动态阈值）==========
# 大概念有涨停更有意义
limit_high_thresh = 0.30 * size_factor
mask = limit_up_ratio >= limit_high_thresh
scores[mask] += 20
for i in np.where(mask)[0]: triggered[i].append('limit_up_high')
limit_medium_thresh = 0.20 * size_factor
mask = (limit_up_ratio >= limit_medium_thresh) & (limit_up_ratio < limit_high_thresh)
scores[mask] += 12
for i in np.where(mask)[0]: triggered[i].append('limit_up_medium')
# ========== 概念规模加分（大概念异动更有价值）==========
# 大概念（50+）额外加分
large_concept = stock_count >= 50
has_signal = scores > 0 # 至少触发了某个规则
mask = large_concept & has_signal
scores[mask] += 10
for i in np.where(mask)[0]: triggered[i].append('large_concept_bonus')
# 超大概念（100+）再加分
xlarge_concept = stock_count >= 100
mask = xlarge_concept & has_signal
scores[mask] += 10
for i in np.where(mask)[0]: triggered[i].append('xlarge_concept_bonus')
# ========== 组合规则（动态阈值）==========
combo_alpha_thresh = 3.0 * size_factor
# Alpha + 放量 + 排名（三重验证）
mask = (alpha_abs >= combo_alpha_thresh) & (amt_ratio >= 5.0) & ((rank_pct >= 0.95) | (rank_pct <= 0.05))
scores[mask] += 20
for i in np.where(mask)[0]: triggered[i].append('triple_signal')
# Alpha + 涨停（强组合）
mask = (alpha_abs >= combo_alpha_thresh) & (limit_up_ratio >= 0.15 * size_factor)
scores[mask] += 15
for i in np.where(mask)[0]: triggered[i].append('alpha_with_limit')
# ========== 小概念惩罚（过滤噪音）==========
# 微型概念（<5 只）如果只有单一信号，减分
tiny_concept = stock_count < 5
single_rule = np.array([len(t) <= 1 for t in triggered])
mask = tiny_concept & single_rule & (scores > 0)
scores[mask] *= 0.5 # 减半
for i in np.where(mask)[0]: triggered[i].append('tiny_concept_penalty')
scores = np.clip(scores, 0, 100)
return scores, triggered
# ==================== ML 评分器 ====================
class FastMLScorer:
"""快速 ML 评分器"""
def __init__(self, checkpoint_dir: str = 'ml/checkpoints', device: str = 'auto'):
self.checkpoint_dir = Path(checkpoint_dir)
if device == 'auto':
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
elif device == 'cuda' and not torch.cuda.is_available():
print("警告: CUDA 不可用，使用 CPU")
self.device = torch.device('cpu')
else:
self.device = torch.device(device)
self.model = None
self.thresholds = None
self._load_model()
def _load_model(self):
model_path = self.checkpoint_dir / 'best_model.pt'
thresholds_path = self.checkpoint_dir / 'thresholds.json'
config_path = self.checkpoint_dir / 'config.json'
if not model_path.exists():
print(f"警告: 模型不存在 {model_path}")
return
try:
from model import LSTMAutoencoder
config = {}
if config_path.exists():
with open(config_path) as f:
config = json.load(f).get('model', {})
# 处理旧配置键名
if 'd_model' in config:
config['hidden_dim'] = config.pop('d_model') // 2
for key in ['num_encoder_layers', 'num_decoder_layers', 'nhead', 'dim_feedforward', 'max_seq_len', 'use_instance_norm']:
config.pop(key, None)
if 'num_layers' not in config:
config['num_layers'] = 1
checkpoint = torch.load(model_path, map_location='cpu')
self.model = LSTMAutoencoder(**config)
self.model.load_state_dict(checkpoint['model_state_dict'])
self.model.to(self.device)
self.model.eval()
if thresholds_path.exists():
with open(thresholds_path) as f:
self.thresholds = json.load(f)
print(f"ML模型加载成功 (设备: {self.device})")
except Exception as e:
print(f"ML模型加载失败: {e}")
self.model = None
def is_ready(self):
return self.model is not None
@torch.no_grad()
def score_batch(self, sequences: np.ndarray) -> np.ndarray:
"""
批量计算 ML 得分
Args:
sequences: (batch, seq_len, n_features)
Returns:
scores: (batch,) 0-100 分数
"""
if not self.is_ready() or len(sequences) == 0:
return np.zeros(len(sequences))
x = torch.FloatTensor(sequences).to(self.device)
output, _ = self.model(x)
mse = ((output - x) ** 2).mean(dim=-1)
errors = mse[:, -1].cpu().numpy()
p95 = self.thresholds.get('p95', 0.1) if self.thresholds else 0.1
scores = np.clip(errors / p95 * 50, 0, 100)
return scores
# ==================== 快速回测 ====================
def build_sequences_fast(
df: pd.DataFrame,
seq_len: int = 30,
prev_df: pd.DataFrame = None
) -> Tuple[np.ndarray, pd.DataFrame]:
"""
快速构建所有有效序列
支持跨日序列：用前一天收盘数据 + 当天开盘数据拼接，实现 9:30 就能检测
Args:
df: 当天数据
seq_len: 序列长度
prev_df: 前一天数据（可选，用于构建开盘时的序列）
返回：
sequences: (n_valid, seq_len, n_features) 所有有效序列
info_df: 对应的元信息 DataFrame
"""
# 确保按概念和时间排序
df = df.sort_values(['concept_id', 'timestamp']).reset_index(drop=True)
# 如果有前一天数据，按概念构建尾部缓存（取每个概念最后 seq_len-1 条）
prev_cache = {}
if prev_df is not None and len(prev_df) > 0:
prev_df = prev_df.sort_values(['concept_id', 'timestamp'])
for concept_id, gdf in prev_df.groupby('concept_id'):
tail_data = gdf.tail(seq_len - 1)
if len(tail_data) > 0:
feat_matrix = tail_data[FEATURES].values
feat_matrix = np.nan_to_num(feat_matrix, nan=0.0, posinf=0.0, neginf=0.0)
feat_matrix = np.clip(feat_matrix, -CONFIG['clip_value'], CONFIG['clip_value'])
prev_cache[concept_id] = feat_matrix
# 按概念分组
groups = df.groupby('concept_id')
sequences = []
infos = []
for concept_id, gdf in groups:
gdf = gdf.reset_index(drop=True)
# 获取特征矩阵
feat_matrix = gdf[FEATURES].values
feat_matrix = np.nan_to_num(feat_matrix, nan=0.0, posinf=0.0, neginf=0.0)
feat_matrix = np.clip(feat_matrix, -CONFIG['clip_value'], CONFIG['clip_value'])
# 如果有前一天缓存，拼接到当天数据前面
if concept_id in prev_cache:
prev_data = prev_cache[concept_id]
combined_matrix = np.vstack([prev_data, feat_matrix])
# 计算偏移量：前一天数据的长度
offset = len(prev_data)
else:
combined_matrix = feat_matrix
offset = 0
# 滑动窗口构建序列
n_total = len(combined_matrix)
if n_total < seq_len:
continue
for i in range(n_total - seq_len + 1):
seq = combined_matrix[i:i + seq_len]
# 计算对应当天数据的索引
# 序列最后一个点的位置 = i + seq_len - 1
# 对应当天数据的索引 = (i + seq_len - 1) - offset
today_idx = i + seq_len - 1 - offset
# 只要序列的最后一个点是当天的数据，就记录
if today_idx < 0 or today_idx >= len(gdf):
continue
sequences.append(seq)
# 记录最后一个时间步的信息（当天的）
row = gdf.iloc[today_idx]
infos.append({
'concept_id': concept_id,
'timestamp': row['timestamp'],
'alpha': row['alpha'],
'alpha_delta': row.get('alpha_delta', 0),
'amt_ratio': row.get('amt_ratio', 1),
'amt_delta': row.get('amt_delta', 0),
'rank_pct': row.get('rank_pct', 0.5),
'limit_up_ratio': row.get('limit_up_ratio', 0),
'stock_count': row.get('stock_count', 0),
'total_amt': row.get('total_amt', 0),
})
if not sequences:
return np.array([]), pd.DataFrame()
return np.array(sequences), pd.DataFrame(infos)
def backtest_single_day_fast(
ml_scorer: FastMLScorer,
df: pd.DataFrame,
date: str,
config: Dict,
prev_df: pd.DataFrame = None
) -> List[Dict]:
"""
快速回测单天（向量化版本）
Args:
ml_scorer: ML 评分器
df: 当天数据
date: 日期
config: 配置
prev_df: 前一天数据（用于 9:30 开始检测）
"""
seq_len = config.get('seq_len', 30)
# 1. 构建所有序列（支持跨日）
sequences, info_df = build_sequences_fast(df, seq_len, prev_df)
if len(sequences) == 0:
return []
# 2. 过滤小波动
alpha_abs = np.abs(info_df['alpha'].values)
valid_mask = alpha_abs >= config['min_alpha_abs']
sequences = sequences[valid_mask]
info_df = info_df[valid_mask].reset_index(drop=True)
if len(sequences) == 0:
return []
# 3. 批量规则评分
rule_scores, triggered_rules = score_rules_batch(info_df)
# 4. 批量 ML 评分（分批处理避免显存溢出）
batch_size = 2048
ml_scores = []
for i in range(0, len(sequences), batch_size):
batch_seq = sequences[i:i+batch_size]
batch_scores = ml_scorer.score_batch(batch_seq)
ml_scores.append(batch_scores)
ml_scores = np.concatenate(ml_scores) if ml_scores else np.zeros(len(sequences))
# 5. 融合得分
w1, w2 = config['rule_weight'], config['ml_weight']
final_scores = w1 * rule_scores + w2 * ml_scores
# 6. 判断异动
is_anomaly = (
(rule_scores >= config['rule_trigger']) |
(ml_scores >= config['ml_trigger']) |
(final_scores >= config['fusion_trigger'])
)
# 7. 应用冷却期（按概念+时间排序后处理）
info_df['rule_score'] = rule_scores
info_df['ml_score'] = ml_scores
info_df['final_score'] = final_scores
info_df['is_anomaly'] = is_anomaly
info_df['triggered_rules'] = triggered_rules
# 只保留异动
anomaly_df = info_df[info_df['is_anomaly']].copy()
if len(anomaly_df) == 0:
return []
# 应用冷却期
anomaly_df = anomaly_df.sort_values(['concept_id', 'timestamp'])
cooldown = {}
keep_mask = []
for _, row in anomaly_df.iterrows():
cid = row['concept_id']
ts = row['timestamp']
if cid in cooldown:
try:
diff = (ts - cooldown[cid]).total_seconds() / 60
except:
diff = config['cooldown_minutes'] + 1
if diff < config['cooldown_minutes']:
keep_mask.append(False)
continue
cooldown[cid] = ts
keep_mask.append(True)
anomaly_df = anomaly_df[keep_mask]
# 8. 按时间分组，每分钟最多 max_alerts_per_minute 个
alerts = []
for ts, group in anomaly_df.groupby('timestamp'):
group = group.nlargest(config['max_alerts_per_minute'], 'final_score')
for _, row in group.iterrows():
alpha = row['alpha']
if alpha >= 1.5:
atype = 'surge_up'
elif alpha <= -1.5:
atype = 'surge_down'
elif row['amt_ratio'] >= 3.0:
atype = 'volume_spike'
else:
atype = 'unknown'
rule_score = row['rule_score']
ml_score = row['ml_score']
final_score = row['final_score']
if rule_score >= config['rule_trigger']:
trigger = f'规则强信号({rule_score:.0f}分)'
elif ml_score >= config['ml_trigger']:
trigger = f'ML强信号({ml_score:.0f}分)'
else:
trigger = f'融合触发({final_score:.0f}分)'
alerts.append({
'concept_id': row['concept_id'],
'alert_time': row['timestamp'],
'trade_date': date,
'alert_type': atype,
'final_score': final_score,
'rule_score': rule_score,
'ml_score': ml_score,
'trigger_reason': trigger,
'triggered_rules': row['triggered_rules'],
'alpha': alpha,
'alpha_delta': row['alpha_delta'],
'amt_ratio': row['amt_ratio'],
'amt_delta': row['amt_delta'],
'rank_pct': row['rank_pct'],
'limit_up_ratio': row['limit_up_ratio'],
'stock_count': row['stock_count'],
'total_amt': row['total_amt'],
})
return alerts
# ==================== 数据加载 ====================
def load_daily_features(data_dir: str, date: str) -> Optional[pd.DataFrame]:
file_path = Path(data_dir) / f"features_{date}.parquet"
if not file_path.exists():
return None
return pd.read_parquet(file_path)
def get_available_dates(data_dir: str, start: str, end: str) -> List[str]:
data_path = Path(data_dir)
dates = []
for f in sorted(data_path.glob("features_*.parquet")):
d = f.stem.replace('features_', '')
if start <= d <= end:
dates.append(d)
return dates
def get_prev_trading_day(data_dir: str, date: str) -> Optional[str]:
"""获取给定日期之前最近的有数据的交易日"""
data_path = Path(data_dir)
all_dates = sorted([f.stem.replace('features_', '') for f in data_path.glob("features_*.parquet")])
for i, d in enumerate(all_dates):
if d == date and i > 0:
return all_dates[i - 1]
return None
def export_to_csv(alerts: List[Dict], path: str):
if alerts:
pd.DataFrame(alerts).to_csv(path, index=False, encoding='utf-8-sig')
print(f"已导出: {path}")
# ==================== 数据库写入 ====================
def init_db_table():
"""
初始化数据库表（如果不存在则创建）
表结构说明：
- concept_id: 概念ID
- alert_time: 异动时间（精确到分钟）
- trade_date: 交易日期
- alert_type: 异动类型（surge_up/surge_down/volume_spike/unknown）
- final_score: 最终得分（0-100）
- rule_score: 规则得分（0-100）
- ml_score: ML得分（0-100）
- trigger_reason: 触发原因
- alpha: 超额收益率
- alpha_delta: alpha变化速度
- amt_ratio: 成交额放大倍数
- rank_pct: 排名百分位
- stock_count: 概念股票数量
- triggered_rules: 触发的规则列表（JSON）
"""
create_sql = text("""
CREATE TABLE IF NOT EXISTS concept_anomaly_hybrid (
id INT AUTO_INCREMENT PRIMARY KEY,
concept_id VARCHAR(64) NOT NULL,
alert_time DATETIME NOT NULL,
trade_date DATE NOT NULL,
alert_type VARCHAR(32) NOT NULL,
final_score FLOAT NOT NULL,
rule_score FLOAT NOT NULL,
ml_score FLOAT NOT NULL,
trigger_reason VARCHAR(64),
alpha FLOAT,
alpha_delta FLOAT,
amt_ratio FLOAT,
amt_delta FLOAT,
rank_pct FLOAT,
limit_up_ratio FLOAT,
stock_count INT,
total_amt FLOAT,
triggered_rules JSON,
created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
UNIQUE KEY uk_concept_time (concept_id, alert_time, trade_date),
INDEX idx_trade_date (trade_date),
INDEX idx_concept_id (concept_id),
INDEX idx_final_score (final_score),
INDEX idx_alert_type (alert_type)
) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COMMENT='概念异动检测结果（融合版）'
""")
with MYSQL_ENGINE.begin() as conn:
conn.execute(create_sql)
print("数据库表已就绪: concept_anomaly_hybrid")
def save_alerts_to_mysql(alerts: List[Dict], dry_run: bool = False) -> int:
"""
保存异动到 MySQL
Args:
alerts: 异动列表
dry_run: 是否只模拟，不实际写入
Returns:
实际保存的记录数
"""
if not alerts:
return 0
if dry_run:
print(f" [Dry Run] 将写入 {len(alerts)} 条异动")
return len(alerts)
saved = 0
skipped = 0
with MYSQL_ENGINE.begin() as conn:
for alert in alerts:
try:
# 检查是否已存在（使用 INSERT IGNORE 更高效）
insert_sql = text("""
INSERT IGNORE INTO concept_anomaly_hybrid
(concept_id, alert_time, trade_date, alert_type,
final_score, rule_score, ml_score, trigger_reason,
alpha, alpha_delta, amt_ratio, amt_delta,
rank_pct, limit_up_ratio, stock_count, total_amt,
triggered_rules)
VALUES
(:concept_id, :alert_time, :trade_date, :alert_type,
:final_score, :rule_score, :ml_score, :trigger_reason,
:alpha, :alpha_delta, :amt_ratio, :amt_delta,
:rank_pct, :limit_up_ratio, :stock_count, :total_amt,
:triggered_rules)
""")
result = conn.execute(insert_sql, {
'concept_id': alert['concept_id'],
'alert_time': alert['alert_time'],
'trade_date': alert['trade_date'],
'alert_type': alert['alert_type'],
'final_score': alert['final_score'],
'rule_score': alert['rule_score'],
'ml_score': alert['ml_score'],
'trigger_reason': alert['trigger_reason'],
'alpha': alert.get('alpha', 0),
'alpha_delta': alert.get('alpha_delta', 0),
'amt_ratio': alert.get('amt_ratio', 1),
'amt_delta': alert.get('amt_delta', 0),
'rank_pct': alert.get('rank_pct', 0.5),
'limit_up_ratio': alert.get('limit_up_ratio', 0),
'stock_count': alert.get('stock_count', 0),
'total_amt': alert.get('total_amt', 0),
'triggered_rules': json.dumps(alert.get('triggered_rules', []), ensure_ascii=False),
})
if result.rowcount > 0:
saved += 1
else:
skipped += 1
except Exception as e:
print(f" 保存失败: {alert['concept_id']} @ {alert['alert_time']} - {e}")
if skipped > 0:
print(f" 跳过 {skipped} 条重复记录")
return saved
def clear_alerts_by_date(trade_date: str) -> int:
"""清除指定日期的异动记录（用于重新回测）"""
with MYSQL_ENGINE.begin() as conn:
result = conn.execute(
text("DELETE FROM concept_anomaly_hybrid WHERE trade_date = :trade_date"),
{'trade_date': trade_date}
)
return result.rowcount
def analyze_alerts(alerts: List[Dict]):
if not alerts:
print("无异动")
return
df = pd.DataFrame(alerts)
print(f"\n总异动: {len(alerts)}")
print(f"\n类型分布:\n{df['alert_type'].value_counts()}")
print(f"\n得分统计:")
print(f" 最终: {df['final_score'].mean():.1f} (max: {df['final_score'].max():.1f})")
print(f" 规则: {df['rule_score'].mean():.1f} (max: {df['rule_score'].max():.1f})")
print(f" ML: {df['ml_score'].mean():.1f} (max: {df['ml_score'].max():.1f})")
trigger_type = df['trigger_reason'].apply(
lambda x: '规则' if '规则' in x else ('ML' if 'ML' in x else '融合')
)
print(f"\n触发来源:\n{trigger_type.value_counts()}")
# ==================== 主函数 ====================
def main():
parser = argparse.ArgumentParser(description='快速融合异动回测')
parser.add_argument('--data_dir', default='ml/data')
parser.add_argument('--checkpoint_dir', default='ml/checkpoints')
parser.add_argument('--start', required=True)
parser.add_argument('--end', default=None)
parser.add_argument('--dry-run', action='store_true', help='模拟运行，不写入数据库')
parser.add_argument('--export-csv', default=None, help='导出 CSV 文件路径')
parser.add_argument('--save-db', action='store_true', help='保存结果到数据库')
parser.add_argument('--clear-first', action='store_true', help='写入前先清除该日期的旧数据')
parser.add_argument('--device', default='auto')
args = parser.parse_args()
if args.end is None:
args.end = args.start
print("=" * 60)
print("快速融合异动回测")
print("=" * 60)
print(f"日期: {args.start} ~ {args.end}")
print(f"设备: {args.device}")
print(f"保存数据库: {args.save_db}")
print("=" * 60)
# 初始化数据库表（如果需要保存）
if args.save_db and not args.dry_run:
init_db_table()
# 初始化 ML 评分器
ml_scorer = FastMLScorer(args.checkpoint_dir, args.device)
# 获取日期
dates = get_available_dates(args.data_dir, args.start, args.end)
if not dates:
print("无数据")
return
print(f"找到 {len(dates)} 天数据\n")
# 回测（支持跨日序列）
all_alerts = []
total_saved = 0
prev_df = None # 缓存前一天数据
for i, date in enumerate(tqdm(dates, desc="回测")):
df = load_daily_features(args.data_dir, date)
if df is None or df.empty:
prev_df = None # 当天无数据，清空缓存
continue
# 第一天需要加载前一天数据（如果存在）
if i == 0 and prev_df is None:
prev_date = get_prev_trading_day(args.data_dir, date)
if prev_date:
prev_df = load_daily_features(args.data_dir, prev_date)
if prev_df is not None:
tqdm.write(f" 加载前一天数据: {prev_date}")
alerts = backtest_single_day_fast(ml_scorer, df, date, CONFIG, prev_df)
all_alerts.extend(alerts)
# 保存到数据库
if args.save_db and alerts:
if args.clear_first and not args.dry_run:
cleared = clear_alerts_by_date(date)
if cleared > 0:
tqdm.write(f" 清除 {date} 旧数据: {cleared} 条")
saved = save_alerts_to_mysql(alerts, dry_run=args.dry_run)
total_saved += saved
tqdm.write(f" {date}: {len(alerts)} 个异动, 保存 {saved} 条")
elif alerts:
tqdm.write(f" {date}: {len(alerts)} 个异动")
# 当天数据成为下一天的 prev_df
prev_df = df
# 导出 CSV
if args.export_csv:
export_to_csv(all_alerts, args.export_csv)
# 分析
analyze_alerts(all_alerts)
print(f"\n总计: {len(all_alerts)} 个异动")
if args.save_db:
print(f"已保存到数据库: {total_saved} 条")
if __name__ == "__main__":
main()

481
ml/backtest_hybrid.py
View File

@@ -1,481 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
融合异动回测脚本
使用 HybridAnomalyDetector 进行回测：
- 规则评分 + LSTM Autoencoder 融合判断
- 输出更丰富的异动信息
使用方法：
python backtest_hybrid.py --start 2024-01-01 --end 2024-12-01
python backtest_hybrid.py --start 2024-11-01 --dry-run
"""
import os
import sys
import argparse
import json
from datetime import datetime
from pathlib import Path
from typing import Dict, List, Optional
from collections import defaultdict
import numpy as np
import pandas as pd
from tqdm import tqdm
from sqlalchemy import create_engine, text
# 添加父目录到路径
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from detector import HybridAnomalyDetector, create_detector
# ==================== 配置 ====================
MYSQL_ENGINE = create_engine(
"mysql+pymysql://root:Zzl5588161!@192.168.1.5:3306/stock",
echo=False
)
FEATURES = [
'alpha',
'alpha_delta',
'amt_ratio',
'amt_delta',
'rank_pct',
'limit_up_ratio',
]
BACKTEST_CONFIG = {
'seq_len': 30,
'min_alpha_abs': 0.3, # 降低阈值，让规则也能发挥作用
'cooldown_minutes': 8,
'max_alerts_per_minute': 20,
'clip_value': 10.0,
}
# ==================== 数据加载 ====================
def load_daily_features(data_dir: str, date: str) -> Optional[pd.DataFrame]:
"""加载单天的特征数据"""
file_path = Path(data_dir) / f"features_{date}.parquet"
if not file_path.exists():
return None
df = pd.read_parquet(file_path)
return df
def get_available_dates(data_dir: str, start_date: str, end_date: str) -> List[str]:
"""获取可用的日期列表"""
data_path = Path(data_dir)
all_files = sorted(data_path.glob("features_*.parquet"))
dates = []
for f in all_files:
date = f.stem.replace('features_', '')
if start_date <= date <= end_date:
dates.append(date)
return dates
# ==================== 融合回测 ====================
def backtest_single_day_hybrid(
detector: HybridAnomalyDetector,
df: pd.DataFrame,
date: str,
seq_len: int = 30
) -> List[Dict]:
"""
使用融合检测器回测单天数据（批量优化版）
"""
alerts = []
# 按概念分组，预先构建字典
grouped_dict = {cid: cdf for cid, cdf in df.groupby('concept_id', sort=False)}
# 冷却记录
cooldown = {}
# 获取所有时间点
all_timestamps = sorted(df['timestamp'].unique())
if len(all_timestamps) < seq_len:
return alerts
# 对每个时间点进行检测
for t_idx in range(seq_len - 1, len(all_timestamps)):
current_time = all_timestamps[t_idx]
window_start_time = all_timestamps[t_idx - seq_len + 1]
# 批量收集该时刻所有候选概念
batch_sequences = []
batch_features = []
batch_infos = []
for concept_id, concept_df in grouped_dict.items():
# 检查冷却（提前过滤）
if concept_id in cooldown:
last_alert = cooldown[concept_id]
if isinstance(current_time, datetime):
time_diff = (current_time - last_alert).total_seconds() / 60
else:
time_diff = BACKTEST_CONFIG['cooldown_minutes'] + 1
if time_diff < BACKTEST_CONFIG['cooldown_minutes']:
continue
# 获取时间窗口内的数据
mask = (concept_df['timestamp'] >= window_start_time) & (concept_df['timestamp'] <= current_time)
window_df = concept_df.loc[mask]
if len(window_df) < seq_len:
continue
window_df = window_df.sort_values('timestamp').tail(seq_len)
# 当前时刻特征
current_row = window_df.iloc[-1]
alpha = current_row.get('alpha', 0)
# 过滤微小波动（提前过滤）
if abs(alpha) < BACKTEST_CONFIG['min_alpha_abs']:
continue
# 提取特征序列
sequence = window_df[FEATURES].values
sequence = np.nan_to_num(sequence, nan=0.0, posinf=0.0, neginf=0.0)
sequence = np.clip(sequence, -BACKTEST_CONFIG['clip_value'], BACKTEST_CONFIG['clip_value'])
current_features = {
'alpha': alpha,
'alpha_delta': current_row.get('alpha_delta', 0),
'amt_ratio': current_row.get('amt_ratio', 1),
'amt_delta': current_row.get('amt_delta', 0),
'rank_pct': current_row.get('rank_pct', 0.5),
'limit_up_ratio': current_row.get('limit_up_ratio', 0),
}
batch_sequences.append(sequence)
batch_features.append(current_features)
batch_infos.append({
'concept_id': concept_id,
'stock_count': current_row.get('stock_count', 0),
'total_amt': current_row.get('total_amt', 0),
})
if not batch_sequences:
continue
# 批量 ML 推理
sequences_array = np.array(batch_sequences)
ml_scores = detector.ml_scorer.score(sequences_array) if detector.ml_scorer.is_ready() else [0.0] * len(batch_sequences)
if isinstance(ml_scores, float):
ml_scores = [ml_scores]
# 批量规则评分 + 融合
minute_alerts = []
for i, (features, info) in enumerate(zip(batch_features, batch_infos)):
concept_id = info['concept_id']
# 规则评分
rule_score, rule_details = detector.rule_scorer.score(features)
# ML 评分
ml_score = ml_scores[i] if i < len(ml_scores) else 0.0
# 融合
w1 = detector.config['rule_weight']
w2 = detector.config['ml_weight']
final_score = w1 * rule_score + w2 * ml_score
# 判断是否异动
is_anomaly = False
trigger_reason = ''
if rule_score >= detector.config['rule_trigger']:
is_anomaly = True
trigger_reason = f'规则强信号({rule_score:.0f}分)'
elif ml_score >= detector.config['ml_trigger']:
is_anomaly = True
trigger_reason = f'ML强信号({ml_score:.0f}分)'
elif final_score >= detector.config['fusion_trigger']:
is_anomaly = True
trigger_reason = f'融合触发({final_score:.0f}分)'
if not is_anomaly:
continue
# 异动类型
alpha = features.get('alpha', 0)
if alpha >= 1.5:
anomaly_type = 'surge_up'
elif alpha <= -1.5:
anomaly_type = 'surge_down'
elif features.get('amt_ratio', 1) >= 3.0:
anomaly_type = 'volume_spike'
else:
anomaly_type = 'unknown'
alert = {
'concept_id': concept_id,
'alert_time': current_time,
'trade_date': date,
'alert_type': anomaly_type,
'final_score': final_score,
'rule_score': rule_score,
'ml_score': ml_score,
'trigger_reason': trigger_reason,
'triggered_rules': list(rule_details.keys()),
**features,
**info,
}
minute_alerts.append(alert)
cooldown[concept_id] = current_time
# 按最终得分排序
minute_alerts.sort(key=lambda x: x['final_score'], reverse=True)
alerts.extend(minute_alerts[:BACKTEST_CONFIG['max_alerts_per_minute']])
return alerts
# ==================== 数据库写入 ====================
def save_alerts_to_mysql(alerts: List[Dict], dry_run: bool = False) -> int:
"""保存异动到 MySQL（增强版字段）"""
if not alerts:
return 0
if dry_run:
print(f" [Dry Run] 将写入 {len(alerts)} 条异动")
return len(alerts)
saved = 0
with MYSQL_ENGINE.begin() as conn:
for alert in alerts:
try:
# 检查是否已存在
check_sql = text("""
SELECT id FROM concept_minute_alert
WHERE concept_id = :concept_id
AND alert_time = :alert_time
AND trade_date = :trade_date
""")
exists = conn.execute(check_sql, {
'concept_id': alert['concept_id'],
'alert_time': alert['alert_time'],
'trade_date': alert['trade_date'],
}).fetchone()
if exists:
continue
# 插入新记录
insert_sql = text("""
INSERT INTO concept_minute_alert
(concept_id, concept_name, alert_time, alert_type, trade_date,
change_pct, zscore, importance_score, stock_count, extra_info)
VALUES
(:concept_id, :concept_name, :alert_time, :alert_type, :trade_date,
:change_pct, :zscore, :importance_score, :stock_count, :extra_info)
""")
extra_info = {
'detection_method': 'hybrid',
'final_score': alert['final_score'],
'rule_score': alert['rule_score'],
'ml_score': alert['ml_score'],
'trigger_reason': alert['trigger_reason'],
'triggered_rules': alert['triggered_rules'],
'alpha': alert.get('alpha', 0),
'alpha_delta': alert.get('alpha_delta', 0),
'amt_ratio': alert.get('amt_ratio', 1),
}
conn.execute(insert_sql, {
'concept_id': alert['concept_id'],
'concept_name': alert.get('concept_name', ''),
'alert_time': alert['alert_time'],
'alert_type': alert['alert_type'],
'trade_date': alert['trade_date'],
'change_pct': alert.get('alpha', 0),
'zscore': alert['final_score'], # 用最终得分作为 zscore
'importance_score': alert['final_score'],
'stock_count': alert.get('stock_count', 0),
'extra_info': json.dumps(extra_info, ensure_ascii=False)
})
saved += 1
except Exception as e:
print(f" 保存失败: {alert['concept_id']} - {e}")
return saved
def export_alerts_to_csv(alerts: List[Dict], output_path: str):
"""导出异动到 CSV"""
if not alerts:
return
df = pd.DataFrame(alerts)
df.to_csv(output_path, index=False, encoding='utf-8-sig')
print(f"已导出到: {output_path}")
# ==================== 统计分析 ====================
def analyze_alerts(alerts: List[Dict]):
"""分析异动结果"""
if not alerts:
print("无异动数据")
return
df = pd.DataFrame(alerts)
print("\n" + "=" * 60)
print("异动统计分析")
print("=" * 60)
# 1. 基本统计
print(f"\n总异动数: {len(alerts)}")
# 2. 按类型统计
print(f"\n异动类型分布:")
print(df['alert_type'].value_counts())
# 3. 得分统计
print(f"\n得分统计:")
print(f" 最终得分 - Mean: {df['final_score'].mean():.1f}, Max: {df['final_score'].max():.1f}")
print(f" 规则得分 - Mean: {df['rule_score'].mean():.1f}, Max: {df['rule_score'].max():.1f}")
print(f" ML得分 - Mean: {df['ml_score'].mean():.1f}, Max: {df['ml_score'].max():.1f}")
# 4. 触发来源分析
print(f"\n触发来源分析:")
trigger_counts = df['trigger_reason'].apply(
lambda x: '规则' if '规则' in x else ('ML' if 'ML' in x else '融合')
).value_counts()
print(trigger_counts)
# 5. 规则触发频率
all_rules = []
for rules in df['triggered_rules']:
if isinstance(rules, list):
all_rules.extend(rules)
if all_rules:
print(f"\n最常触发的规则 (Top 10):")
from collections import Counter
rule_counts = Counter(all_rules)
for rule, count in rule_counts.most_common(10):
print(f" {rule}: {count}")
# ==================== 主函数 ====================
def main():
parser = argparse.ArgumentParser(description='融合异动回测')
parser.add_argument('--data_dir', type=str, default='ml/data',
help='特征数据目录')
parser.add_argument('--checkpoint_dir', type=str, default='ml/checkpoints',
help='模型检查点目录')
parser.add_argument('--start', type=str, required=True,
help='开始日期 (YYYY-MM-DD)')
parser.add_argument('--end', type=str, default=None,
help='结束日期 (YYYY-MM-DD)，默认=start')
parser.add_argument('--dry-run', action='store_true',
help='只计算，不写入数据库')
parser.add_argument('--export-csv', type=str, default=None,
help='导出 CSV 文件路径')
parser.add_argument('--rule-weight', type=float, default=0.6,
help='规则权重 (0-1)')
parser.add_argument('--ml-weight', type=float, default=0.4,
help='ML权重 (0-1)')
parser.add_argument('--device', type=str, default='cuda',
help='设备 (cuda/cpu)，默认 cuda')
args = parser.parse_args()
if args.end is None:
args.end = args.start
print("=" * 60)
print("融合异动回测 (规则 + LSTM)")
print("=" * 60)
print(f"日期范围: {args.start} ~ {args.end}")
print(f"数据目录: {args.data_dir}")
print(f"模型目录: {args.checkpoint_dir}")
print(f"规则权重: {args.rule_weight}")
print(f"ML权重: {args.ml_weight}")
print(f"设备: {args.device}")
print(f"Dry Run: {args.dry_run}")
print("=" * 60)
# 初始化融合检测器（使用 GPU）
config = {
'rule_weight': args.rule_weight,
'ml_weight': args.ml_weight,
}
# 修改 detector.py 中 MLScorer 的设备
from detector import HybridAnomalyDetector
detector = HybridAnomalyDetector(config, args.checkpoint_dir, device=args.device)
# 获取可用日期
dates = get_available_dates(args.data_dir, args.start, args.end)
if not dates:
print(f"未找到 {args.start} ~ {args.end} 范围内的数据")
return
print(f"\n找到 {len(dates)} 天的数据")
# 回测
all_alerts = []
total_saved = 0
for date in tqdm(dates, desc="回测进度"):
df = load_daily_features(args.data_dir, date)
if df is None or df.empty:
continue
alerts = backtest_single_day_hybrid(
detector, df, date,
seq_len=BACKTEST_CONFIG['seq_len']
)
if alerts:
all_alerts.extend(alerts)
saved = save_alerts_to_mysql(alerts, dry_run=args.dry_run)
total_saved += saved
if not args.dry_run:
tqdm.write(f" {date}: 检测到 {len(alerts)} 个异动，保存 {saved} 条")
# 导出 CSV
if args.export_csv and all_alerts:
export_alerts_to_csv(all_alerts, args.export_csv)
# 统计分析
analyze_alerts(all_alerts)
# 汇总
print("\n" + "=" * 60)
print("回测完成！")
print("=" * 60)
print(f"总计检测到: {len(all_alerts)} 个异动")
print(f"保存到数据库: {total_saved} 条")
print("=" * 60)
if __name__ == "__main__":
main()

294
ml/backtest_v2.py
View File

@@ -1,294 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
V2 回测脚本 - 验证时间片对齐 + 持续性确认的效果
回测指标：
1. 准确率：异动后 N 分钟内 alpha 是否继续上涨/下跌
2. 虚警率：多少异动是噪音
3. 持续性：平均异动持续时长
"""
import os
import sys
import json
import argparse
from datetime import datetime
from pathlib import Path
from typing import Dict, List, Tuple
from collections import defaultdict
import numpy as np
import pandas as pd
from tqdm import tqdm
from sqlalchemy import create_engine, text
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from ml.detector_v2 import AnomalyDetectorV2, CONFIG
# ==================== 配置 ====================
MYSQL_ENGINE = create_engine(
"mysql+pymysql://root:Zzl5588161!@192.168.1.5:3306/stock",
echo=False
)
# ==================== 回测评估 ====================
def evaluate_alerts(
alerts: List[Dict],
raw_data: pd.DataFrame,
lookahead_minutes: int = 10
) -> Dict:
"""
评估异动质量
指标：
1. 方向正确率：异动后 N 分钟 alpha 方向是否一致
2. 持续率：异动后 N 分钟内有多少时刻 alpha 保持同向
3. 峰值收益：异动后 N 分钟内的最大 alpha
"""
if not alerts:
return {'accuracy': 0, 'sustained_rate': 0, 'avg_peak': 0, 'total_alerts': 0}
results = []
for alert in alerts:
concept_id = alert['concept_id']
alert_time = alert['alert_time']
alert_alpha = alert['alpha']
is_up = alert_alpha > 0
# 获取该概念在异动后的数据
concept_data = raw_data[
(raw_data['concept_id'] == concept_id) &
(raw_data['timestamp'] > alert_time)
].head(lookahead_minutes)
if len(concept_data) < 3:
continue
future_alphas = concept_data['alpha'].values
# 方向正确：未来 alpha 平均值与当前同向
avg_future_alpha = np.mean(future_alphas)
direction_correct = (is_up and avg_future_alpha > 0) or (not is_up and avg_future_alpha < 0)
# 持续率：有多少时刻保持同向
if is_up:
sustained_count = sum(1 for a in future_alphas if a > 0)
else:
sustained_count = sum(1 for a in future_alphas if a < 0)
sustained_rate = sustained_count / len(future_alphas)
# 峰值收益
if is_up:
peak = max(future_alphas)
else:
peak = min(future_alphas)
results.append({
'direction_correct': direction_correct,
'sustained_rate': sustained_rate,
'peak': peak,
'alert_alpha': alert_alpha,
})
if not results:
return {'accuracy': 0, 'sustained_rate': 0, 'avg_peak': 0, 'total_alerts': 0}
return {
'accuracy': np.mean([r['direction_correct'] for r in results]),
'sustained_rate': np.mean([r['sustained_rate'] for r in results]),
'avg_peak': np.mean([abs(r['peak']) for r in results]),
'total_alerts': len(alerts),
'evaluated_alerts': len(results),
}
def save_alerts_to_mysql(alerts: List[Dict], dry_run: bool = False) -> int:
"""保存异动到 MySQL"""
if not alerts or dry_run:
return 0
# 确保表存在
with MYSQL_ENGINE.begin() as conn:
conn.execute(text("""
CREATE TABLE IF NOT EXISTS concept_anomaly_v2 (
id INT AUTO_INCREMENT PRIMARY KEY,
concept_id VARCHAR(64) NOT NULL,
alert_time DATETIME NOT NULL,
trade_date DATE NOT NULL,
alert_type VARCHAR(32) NOT NULL,
final_score FLOAT NOT NULL,
rule_score FLOAT NOT NULL,
ml_score FLOAT NOT NULL,
trigger_reason VARCHAR(128),
confirm_ratio FLOAT,
alpha FLOAT,
alpha_zscore FLOAT,
amt_zscore FLOAT,
rank_zscore FLOAT,
momentum_3m FLOAT,
momentum_5m FLOAT,
limit_up_ratio FLOAT,
triggered_rules JSON,
created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
UNIQUE KEY uk_concept_time (concept_id, alert_time, trade_date),
INDEX idx_trade_date (trade_date),
INDEX idx_final_score (final_score)
) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COMMENT='概念异动 V2（时间片对齐+持续确认）'
"""))
# 插入数据
saved = 0
with MYSQL_ENGINE.begin() as conn:
for alert in alerts:
try:
conn.execute(text("""
INSERT IGNORE INTO concept_anomaly_v2
(concept_id, alert_time, trade_date, alert_type,
final_score, rule_score, ml_score, trigger_reason, confirm_ratio,
alpha, alpha_zscore, amt_zscore, rank_zscore,
momentum_3m, momentum_5m, limit_up_ratio, triggered_rules)
VALUES
(:concept_id, :alert_time, :trade_date, :alert_type,
:final_score, :rule_score, :ml_score, :trigger_reason, :confirm_ratio,
:alpha, :alpha_zscore, :amt_zscore, :rank_zscore,
:momentum_3m, :momentum_5m, :limit_up_ratio, :triggered_rules)
"""), {
'concept_id': alert['concept_id'],
'alert_time': alert['alert_time'],
'trade_date': alert['trade_date'],
'alert_type': alert['alert_type'],
'final_score': alert['final_score'],
'rule_score': alert['rule_score'],
'ml_score': alert['ml_score'],
'trigger_reason': alert['trigger_reason'],
'confirm_ratio': alert.get('confirm_ratio', 0),
'alpha': alert['alpha'],
'alpha_zscore': alert.get('alpha_zscore', 0),
'amt_zscore': alert.get('amt_zscore', 0),
'rank_zscore': alert.get('rank_zscore', 0),
'momentum_3m': alert.get('momentum_3m', 0),
'momentum_5m': alert.get('momentum_5m', 0),
'limit_up_ratio': alert.get('limit_up_ratio', 0),
'triggered_rules': json.dumps(alert.get('triggered_rules', [])),
})
saved += 1
except Exception as e:
print(f"保存失败: {e}")
return saved
# ==================== 主函数 ====================
def main():
parser = argparse.ArgumentParser(description='V2 回测')
parser.add_argument('--start', type=str, required=True, help='开始日期')
parser.add_argument('--end', type=str, default=None, help='结束日期')
parser.add_argument('--model_dir', type=str, default='ml/checkpoints_v2')
parser.add_argument('--baseline_dir', type=str, default='ml/data_v2/baselines')
parser.add_argument('--save', action='store_true', help='保存到数据库')
parser.add_argument('--lookahead', type=int, default=10, help='评估前瞻时间（分钟）')
args = parser.parse_args()
end_date = args.end or args.start
print("=" * 60)
print("V2 回测 - 时间片对齐 + 持续性确认")
print("=" * 60)
print(f"日期范围: {args.start} ~ {end_date}")
print(f"模型目录: {args.model_dir}")
print(f"评估前瞻: {args.lookahead} 分钟")
# 初始化检测器
detector = AnomalyDetectorV2(
model_dir=args.model_dir,
baseline_dir=args.baseline_dir
)
# 获取交易日
from prepare_data_v2 import get_trading_days
trading_days = get_trading_days(args.start, end_date)
if not trading_days:
print("无交易日")
return
print(f"交易日数: {len(trading_days)}")
# 回测统计
total_stats = {
'total_alerts': 0,
'accuracy_sum': 0,
'sustained_sum': 0,
'peak_sum': 0,
'day_count': 0,
}
all_alerts = []
for trade_date in tqdm(trading_days, desc="回测进度"):
# 检测异动
alerts = detector.detect(trade_date)
if not alerts:
continue
all_alerts.extend(alerts)
# 评估
raw_data = detector._compute_raw_features(trade_date)
if raw_data.empty:
continue
stats = evaluate_alerts(alerts, raw_data, args.lookahead)
if stats['evaluated_alerts'] > 0:
total_stats['total_alerts'] += stats['total_alerts']
total_stats['accuracy_sum'] += stats['accuracy'] * stats['evaluated_alerts']
total_stats['sustained_sum'] += stats['sustained_rate'] * stats['evaluated_alerts']
total_stats['peak_sum'] += stats['avg_peak'] * stats['evaluated_alerts']
total_stats['day_count'] += 1
print(f"\n[{trade_date}] 异动: {stats['total_alerts']}, "
f"准确率: {stats['accuracy']:.1%}, "
f"持续率: {stats['sustained_rate']:.1%}, "
f"峰值: {stats['avg_peak']:.2f}%")
# 汇总
print("\n" + "=" * 60)
print("回测汇总")
print("=" * 60)
if total_stats['total_alerts'] > 0:
avg_accuracy = total_stats['accuracy_sum'] / total_stats['total_alerts']
avg_sustained = total_stats['sustained_sum'] / total_stats['total_alerts']
avg_peak = total_stats['peak_sum'] / total_stats['total_alerts']
print(f"总异动数: {total_stats['total_alerts']}")
print(f"回测天数: {total_stats['day_count']}")
print(f"平均每天: {total_stats['total_alerts'] / max(1, total_stats['day_count']):.1f} 个")
print(f"方向准确率: {avg_accuracy:.1%}")
print(f"持续率: {avg_sustained:.1%}")
print(f"平均峰值: {avg_peak:.2f}%")
else:
print("无异动检测结果")
# 保存
if args.save and all_alerts:
print(f"\n保存 {len(all_alerts)} 条异动到数据库...")
saved = save_alerts_to_mysql(all_alerts)
print(f"保存完成: {saved} 条")
print("=" * 60)
if __name__ == "__main__":
main()

31
ml/checkpoints_v2/config.json
View File

@@ -1,31 +0,0 @@
{
"seq_len": 10,
"stride": 2,
"train_end_date": "2025-06-30",
"val_end_date": "2025-09-30",
"features": [
"alpha_zscore",
"amt_zscore",
"rank_zscore",
"momentum_3m",
"momentum_5m",
"limit_up_ratio"
],
"batch_size": 32768,
"epochs": 150,
"learning_rate": 0.0006,
"weight_decay": 1e-05,
"gradient_clip": 1.0,
"patience": 15,
"min_delta": 1e-06,
"model": {
"n_features": 6,
"hidden_dim": 32,
"latent_dim": 4,
"num_layers": 1,
"dropout": 0.2,
"bidirectional": true
},
"clip_value": 5.0,
"threshold_percentiles": [90, 95, 99]
}

8
ml/checkpoints_v2/thresholds.json
View File

@@ -1,8 +0,0 @@
{
"p90": 0.15,
"p95": 0.25,
"p99": 0.50,
"mean": 0.08,
"std": 0.12,
"median": 0.06
}

635
ml/detector.py
View File

@@ -1,635 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
概念异动检测器 - 融合版
结合两种方法的优势：
1. 规则评分系统：可解释、稳定、覆盖已知模式
2. LSTM Autoencoder：发现未知的异常模式
融合策略：
┌─────────────────────────────────────────────────────────┐
│ 输入特征 │
│ (alpha, alpha_delta, amt_ratio, amt_delta, rank_pct, │
│ limit_up_ratio) │
├─────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────┐ ┌──────────────┐ │
│ │ 规则评分系统 │ │ LSTM Autoencoder │ │
│ │ (0-100分) │ │ (重构误差) │ │
│ └──────┬───────┘ └──────┬───────┘ │
│ │ │ │
│ ▼ ▼ │
│ rule_score (0-100) ml_score (标准化后 0-100) │
│ │
├─────────────────────────────────────────────────────────┤
│ 融合策略 │
│ │
│ final_score = w1 * rule_score + w2 * ml_score │
│ │
│ 异动判定： │
│ - rule_score >= 60 → 直接触发（规则强信号） │
│ - ml_score >= 80 → 直接触发（ML强信号） │
│ - final_score >= 50 → 融合触发 │
│ │
└─────────────────────────────────────────────────────────┘
优势：
- 规则系统保证已知模式的检出率
- ML模型捕捉规则未覆盖的异常
- 两者互相验证，减少误报
"""
import json
from pathlib import Path
from typing import Dict, List, Optional, Tuple
from dataclasses import dataclass
import numpy as np
import torch
# 尝试导入模型（可能不存在）
try:
from model import LSTMAutoencoder, create_model
HAS_MODEL = True
except ImportError:
HAS_MODEL = False
@dataclass
class AnomalyResult:
"""异动检测结果"""
is_anomaly: bool
final_score: float # 最终得分 (0-100)
rule_score: float # 规则得分 (0-100)
ml_score: float # ML得分 (0-100)
trigger_reason: str # 触发原因
rule_details: Dict # 规则明细
anomaly_type: str # 异动类型: surge_up / surge_down / volume_spike / unknown
class RuleBasedScorer:
"""
基于规则的评分系统
设计原则：
- 每个规则独立打分
- 分数可叠加
- 阈值可配置
"""
# 默认规则配置
DEFAULT_RULES = {
# Alpha 相关（超额收益）
'alpha_strong': {
'condition': lambda r: abs(r.get('alpha', 0)) >= 3.0,
'score': 35,
'description': 'Alpha强信号(|α|≥3%)'
},
'alpha_medium': {
'condition': lambda r: 2.0 <= abs(r.get('alpha', 0)) < 3.0,
'score': 25,
'description': 'Alpha中等(2%≤|α|<3%)'
},
'alpha_weak': {
'condition': lambda r: 1.5 <= abs(r.get('alpha', 0)) < 2.0,
'score': 15,
'description': 'Alpha轻微(1.5%≤|α|<2%)'
},
# Alpha 变化率（加速度）
'alpha_delta_strong': {
'condition': lambda r: abs(r.get('alpha_delta', 0)) >= 1.0,
'score': 30,
'description': 'Alpha加速强(|Δα|≥1%)'
},
'alpha_delta_medium': {
'condition': lambda r: 0.5 <= abs(r.get('alpha_delta', 0)) < 1.0,
'score': 20,
'description': 'Alpha加速中(0.5%≤|Δα|<1%)'
},
# 成交额比率（放量）
'volume_spike_strong': {
'condition': lambda r: r.get('amt_ratio', 1) >= 5.0,
'score': 30,
'description': '极度放量(≥5倍)'
},
'volume_spike_medium': {
'condition': lambda r: 3.0 <= r.get('amt_ratio', 1) < 5.0,
'score': 20,
'description': '显著放量(3-5倍)'
},
'volume_spike_weak': {
'condition': lambda r: 2.0 <= r.get('amt_ratio', 1) < 3.0,
'score': 10,
'description': '轻微放量(2-3倍)'
},
# 成交额变化率
'amt_delta_strong': {
'condition': lambda r: abs(r.get('amt_delta', 0)) >= 1.0,
'score': 15,
'description': '成交额急变(|Δamt|≥100%)'
},
# 排名跳变
'rank_top': {
'condition': lambda r: r.get('rank_pct', 0.5) >= 0.95,
'score': 25,
'description': '排名前5%'
},
'rank_bottom': {
'condition': lambda r: r.get('rank_pct', 0.5) <= 0.05,
'score': 25,
'description': '排名后5%'
},
'rank_high': {
'condition': lambda r: 0.9 <= r.get('rank_pct', 0.5) < 0.95,
'score': 15,
'description': '排名前10%'
},
# 涨停比例
'limit_up_high': {
'condition': lambda r: r.get('limit_up_ratio', 0) >= 0.2,
'score': 25,
'description': '涨停比例≥20%'
},
'limit_up_medium': {
'condition': lambda r: 0.1 <= r.get('limit_up_ratio', 0) < 0.2,
'score': 15,
'description': '涨停比例10-20%'
},
# 组合条件（更可靠的信号）
'alpha_with_volume': {
'condition': lambda r: abs(r.get('alpha', 0)) >= 1.5 and r.get('amt_ratio', 1) >= 2.0,
'score': 20, # 额外加分
'description': 'Alpha+放量组合'
},
'acceleration_with_rank': {
'condition': lambda r: abs(r.get('alpha_delta', 0)) >= 0.5 and (r.get('rank_pct', 0.5) >= 0.9 or r.get('rank_pct', 0.5) <= 0.1),
'score': 15, # 额外加分
'description': '加速+排名异常组合'
},
}
def __init__(self, rules: Dict = None):
"""
初始化规则评分器
Args:
rules: 自定义规则，格式同 DEFAULT_RULES
"""
self.rules = rules or self.DEFAULT_RULES
def score(self, features: Dict) -> Tuple[float, Dict]:
"""
计算规则得分
Args:
features: 特征字典，包含 alpha, alpha_delta, amt_ratio 等
Returns:
score: 总分 (0-100)
details: 触发的规则明细
"""
total_score = 0
triggered_rules = {}
for rule_name, rule_config in self.rules.items():
try:
if rule_config['condition'](features):
total_score += rule_config['score']
triggered_rules[rule_name] = {
'score': rule_config['score'],
'description': rule_config['description']
}
except Exception:
# 忽略规则计算错误
pass
# 限制在 0-100
total_score = min(100, max(0, total_score))
return total_score, triggered_rules
def get_anomaly_type(self, features: Dict) -> str:
"""判断异动类型"""
alpha = features.get('alpha', 0)
amt_ratio = features.get('amt_ratio', 1)
if alpha >= 1.5:
return 'surge_up'
elif alpha <= -1.5:
return 'surge_down'
elif amt_ratio >= 3.0:
return 'volume_spike'
else:
return 'unknown'
class MLScorer:
"""
基于 LSTM Autoencoder 的评分器
将重构误差转换为 0-100 的分数
"""
def __init__(
self,
checkpoint_dir: str = 'ml/checkpoints',
device: str = 'auto'
):
self.checkpoint_dir = Path(checkpoint_dir)
# 设备检测
if device == 'auto':
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
elif device == 'cuda' and not torch.cuda.is_available():
print("警告: CUDA 不可用，使用 CPU")
self.device = torch.device('cpu')
else:
self.device = torch.device(device)
self.model = None
self.thresholds = None
self.config = None
# 尝试加载模型
self._load_model()
def _load_model(self):
"""加载模型和阈值"""
if not HAS_MODEL:
print("警告: 无法导入模型模块")
return
model_path = self.checkpoint_dir / 'best_model.pt'
thresholds_path = self.checkpoint_dir / 'thresholds.json'
config_path = self.checkpoint_dir / 'config.json'
if not model_path.exists():
print(f"警告: 模型文件不存在 {model_path}")
return
try:
# 加载配置
if config_path.exists():
with open(config_path, 'r') as f:
self.config = json.load(f)
# 先用 CPU 加载模型（避免 CUDA 不可用问题），再移动到目标设备
checkpoint = torch.load(model_path, map_location='cpu')
model_config = self.config.get('model', {}) if self.config else {}
self.model = create_model(model_config)
self.model.load_state_dict(checkpoint['model_state_dict'])
self.model.to(self.device)
self.model.eval()
# 加载阈值
if thresholds_path.exists():
with open(thresholds_path, 'r') as f:
self.thresholds = json.load(f)
print(f"MLScorer 加载成功 (设备: {self.device})")
except Exception as e:
print(f"警告: 模型加载失败 - {e}")
import traceback
traceback.print_exc()
self.model = None
def is_ready(self) -> bool:
"""检查模型是否就绪"""
return self.model is not None
@torch.no_grad()
def score(self, sequence: np.ndarray) -> float:
"""
计算 ML 得分
Args:
sequence: (seq_len, n_features) 或 (batch, seq_len, n_features)
Returns:
score: 0-100 的分数，越高越异常
"""
if not self.is_ready():
return 0.0
# 确保是 3D
if sequence.ndim == 2:
sequence = sequence[np.newaxis, ...]
# 转为 tensor
x = torch.FloatTensor(sequence).to(self.device)
# 计算重构误差
output, _ = self.model(x)
mse = ((output - x) ** 2).mean(dim=-1) # (batch, seq_len)
# 取最后时刻的误差
error = mse[:, -1].cpu().numpy()
# 转换为 0-100 分数
# 使用 p95 阈值作为参考
if self.thresholds:
p95 = self.thresholds.get('p95', 0.1)
p99 = self.thresholds.get('p99', 0.2)
else:
p95, p99 = 0.1, 0.2
# 线性映射：p95 -> 50分, p99 -> 80分
# error=0 -> 0分, error>=p99*1.5 -> 100分
score = np.clip(error / p95 * 50, 0, 100)
return float(score[0]) if len(score) == 1 else score.tolist()
class HybridAnomalyDetector:
"""
融合异动检测器
结合规则系统和 ML 模型
"""
# 默认配置
DEFAULT_CONFIG = {
# 权重配置
'rule_weight': 0.6, # 规则权重
'ml_weight': 0.4, # ML权重
# 触发阈值
'rule_trigger': 60, # 规则直接触发阈值
'ml_trigger': 80, # ML直接触发阈值
'fusion_trigger': 50, # 融合触发阈值
# 特征列表
'features': [
'alpha', 'alpha_delta', 'amt_ratio',
'amt_delta', 'rank_pct', 'limit_up_ratio'
],
# 序列长度（ML模型需要）
'seq_len': 30,
}
def __init__(
self,
config: Dict = None,
checkpoint_dir: str = 'ml/checkpoints',
device: str = 'auto'
):
self.config = {**self.DEFAULT_CONFIG, **(config or {})}
# 初始化评分器
self.rule_scorer = RuleBasedScorer()
self.ml_scorer = MLScorer(checkpoint_dir, device)
print(f"HybridAnomalyDetector 初始化完成")
print(f" 规则权重: {self.config['rule_weight']}")
print(f" ML权重: {self.config['ml_weight']}")
print(f" ML模型: {'就绪' if self.ml_scorer.is_ready() else '未加载'}")
def detect(
self,
features: Dict,
sequence: np.ndarray = None
) -> AnomalyResult:
"""
检测异动
Args:
features: 当前时刻的特征字典
sequence: 历史序列 (seq_len, n_features)，ML模型需要
Returns:
AnomalyResult: 检测结果
"""
# 1. 规则评分
rule_score, rule_details = self.rule_scorer.score(features)
# 2. ML评分
ml_score = 0.0
if sequence is not None and self.ml_scorer.is_ready():
ml_score = self.ml_scorer.score(sequence)
# 3. 融合得分
w1 = self.config['rule_weight']
w2 = self.config['ml_weight']
# 如果ML不可用，全部权重给规则
if not self.ml_scorer.is_ready():
w1, w2 = 1.0, 0.0
final_score = w1 * rule_score + w2 * ml_score
# 4. 判断是否异动
is_anomaly = False
trigger_reason = ''
if rule_score >= self.config['rule_trigger']:
is_anomaly = True
trigger_reason = f'规则强信号({rule_score:.0f}分)'
elif ml_score >= self.config['ml_trigger']:
is_anomaly = True
trigger_reason = f'ML强信号({ml_score:.0f}分)'
elif final_score >= self.config['fusion_trigger']:
is_anomaly = True
trigger_reason = f'融合触发({final_score:.0f}分)'
# 5. 判断异动类型
anomaly_type = self.rule_scorer.get_anomaly_type(features) if is_anomaly else ''
return AnomalyResult(
is_anomaly=is_anomaly,
final_score=final_score,
rule_score=rule_score,
ml_score=ml_score,
trigger_reason=trigger_reason,
rule_details=rule_details,
anomaly_type=anomaly_type
)
def detect_batch(
self,
features_list: List[Dict],
sequences: np.ndarray = None
) -> List[AnomalyResult]:
"""
批量检测
Args:
features_list: 特征字典列表
sequences: (batch, seq_len, n_features)
Returns:
List[AnomalyResult]
"""
results = []
for i, features in enumerate(features_list):
seq = sequences[i] if sequences is not None else None
result = self.detect(features, seq)
results.append(result)
return results
# ==================== 便捷函数 ====================
def create_detector(
checkpoint_dir: str = 'ml/checkpoints',
config: Dict = None
) -> HybridAnomalyDetector:
"""创建融合检测器"""
return HybridAnomalyDetector(config, checkpoint_dir)
def quick_detect(features: Dict) -> bool:
"""
快速检测（只用规则，不需要ML模型）
适用于：
- 实时检测
- ML模型未训练完成时
"""
scorer = RuleBasedScorer()
score, _ = scorer.score(features)
return score >= 50
# ==================== 测试 ====================
if __name__ == "__main__":
print("=" * 60)
print("融合异动检测器测试")
print("=" * 60)
# 创建检测器
detector = create_detector()
# 测试用例
test_cases = [
{
'name': '正常情况',
'features': {
'alpha': 0.5,
'alpha_delta': 0.1,
'amt_ratio': 1.2,
'amt_delta': 0.1,
'rank_pct': 0.5,
'limit_up_ratio': 0.02
}
},
{
'name': 'Alpha异动',
'features': {
'alpha': 3.5,
'alpha_delta': 0.8,
'amt_ratio': 2.5,
'amt_delta': 0.5,
'rank_pct': 0.92,
'limit_up_ratio': 0.05
}
},
{
'name': '放量异动',
'features': {
'alpha': 1.2,
'alpha_delta': 0.3,
'amt_ratio': 6.0,
'amt_delta': 1.5,
'rank_pct': 0.85,
'limit_up_ratio': 0.08
}
},
{
'name': '涨停潮',
'features': {
'alpha': 2.5,
'alpha_delta': 0.6,
'amt_ratio': 3.5,
'amt_delta': 0.8,
'rank_pct': 0.98,
'limit_up_ratio': 0.25
}
},
]
print("\n" + "-" * 60)
print("测试1: 只用规则（无序列数据）")
print("-" * 60)
for case in test_cases:
result = detector.detect(case['features'])
print(f"\n{case['name']}:")
print(f" 异动: {'是' if result.is_anomaly else '否'}")
print(f" 最终得分: {result.final_score:.1f}")
print(f" 规则得分: {result.rule_score:.1f}")
print(f" ML得分: {result.ml_score:.1f}")
if result.is_anomaly:
print(f" 触发原因: {result.trigger_reason}")
print(f" 异动类型: {result.anomaly_type}")
print(f" 触发规则: {list(result.rule_details.keys())}")
# 测试2: 带序列数据的融合检测
print("\n" + "-" * 60)
print("测试2: 融合检测（规则 + ML）")
print("-" * 60)
# 生成模拟序列数据
seq_len = 30
n_features = 6
# 正常序列：小幅波动
normal_sequence = np.random.randn(seq_len, n_features) * 0.3
normal_sequence[:, 0] = np.linspace(0, 0.5, seq_len) # alpha 缓慢上升
normal_sequence[:, 2] = np.abs(normal_sequence[:, 2]) + 1 # amt_ratio > 0
# 异常序列：最后几个时间步突然变化
anomaly_sequence = np.random.randn(seq_len, n_features) * 0.3
anomaly_sequence[-5:, 0] = np.linspace(1, 4, 5) # alpha 突然飙升
anomaly_sequence[-5:, 1] = np.linspace(0.2, 1.5, 5) # alpha_delta 加速
anomaly_sequence[-5:, 2] = np.linspace(2, 6, 5) # amt_ratio 放量
anomaly_sequence[:, 2] = np.abs(anomaly_sequence[:, 2]) + 1
# 测试正常序列
normal_features = {
'alpha': float(normal_sequence[-1, 0]),
'alpha_delta': float(normal_sequence[-1, 1]),
'amt_ratio': float(normal_sequence[-1, 2]),
'amt_delta': float(normal_sequence[-1, 3]),
'rank_pct': 0.5,
'limit_up_ratio': 0.02
}
result = detector.detect(normal_features, normal_sequence)
print(f"\n正常序列:")
print(f" 异动: {'是' if result.is_anomaly else '否'}")
print(f" 最终得分: {result.final_score:.1f}")
print(f" 规则得分: {result.rule_score:.1f}")
print(f" ML得分: {result.ml_score:.1f}")
# 测试异常序列
anomaly_features = {
'alpha': float(anomaly_sequence[-1, 0]),
'alpha_delta': float(anomaly_sequence[-1, 1]),
'amt_ratio': float(anomaly_sequence[-1, 2]),
'amt_delta': float(anomaly_sequence[-1, 3]),
'rank_pct': 0.95,
'limit_up_ratio': 0.15
}
result = detector.detect(anomaly_features, anomaly_sequence)
print(f"\n异常序列:")
print(f" 异动: {'是' if result.is_anomaly else '否'}")
print(f" 最终得分: {result.final_score:.1f}")
print(f" 规则得分: {result.rule_score:.1f}")
print(f" ML得分: {result.ml_score:.1f}")
if result.is_anomaly:
print(f" 触发原因: {result.trigger_reason}")
print(f" 异动类型: {result.anomaly_type}")
print("\n" + "=" * 60)
print("测试完成!")

716
ml/detector_v2.py
View File

@@ -1,716 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
异动检测器 V2 - 基于时间片对齐 + 持续性确认
核心改进：
1. Z-Score 特征：相对于同时间片历史的偏离
2. 短序列 LSTM：10分钟序列，开盘即可用
3. 持续性确认：5分钟窗口内60%时刻超标才确认为异动
检测流程：
1. 计算当前时刻的 Z-Score（对比同时间片历史基线）
2. 构建最近10分钟的 Z-Score 序列
3. LSTM 计算重构误差（ML分数）
4. 规则评分（基于 Z-Score 的规则）
5. 滑动窗口确认：最近5分钟内是否有足够多的时刻超标
6. 只有通过持续性确认的才输出为异动
"""
import os
import sys
import json
import pickle
from datetime import datetime, timedelta
from typing import Dict, List, Optional, Tuple
from collections import defaultdict, deque
import numpy as np
import pandas as pd
import torch
from sqlalchemy import create_engine, text
from elasticsearch import Elasticsearch
from clickhouse_driver import Client
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from ml.model import TransformerAutoencoder
# ==================== 配置 ====================
MYSQL_ENGINE = create_engine(
"mysql+pymysql://root:Zzl5588161!@192.168.1.5:3306/stock",
echo=False
)
ES_CLIENT = Elasticsearch(['http://127.0.0.1:9200'])
ES_INDEX = 'concept_library_v3'
CLICKHOUSE_CONFIG = {
'host': '127.0.0.1',
'port': 9000,
'user': 'default',
'password': 'Zzl33818!',
'database': 'stock'
}
REFERENCE_INDEX = '000001.SH'
# 检测配置
CONFIG = {
# 序列配置
'seq_len': 10, # LSTM 序列长度（分钟）
# 持续性确认配置（核心！）
'confirm_window': 5, # 确认窗口（分钟）
'confirm_ratio': 0.6, # 确认比例（60%时刻需要超标）
# Z-Score 阈值
'alpha_zscore_threshold': 2.0, # Alpha Z-Score 阈值
'amt_zscore_threshold': 2.5, # 成交额 Z-Score 阈值
# 融合权重
'rule_weight': 0.5,
'ml_weight': 0.5,
# 触发阈值
'rule_trigger': 60,
'ml_trigger': 70,
'fusion_trigger': 50,
# 冷却期
'cooldown_minutes': 10,
'max_alerts_per_minute': 15,
# Z-Score 截断
'zscore_clip': 5.0,
}
# V2 特征列表
FEATURES_V2 = [
'alpha_zscore', 'amt_zscore', 'rank_zscore',
'momentum_3m', 'momentum_5m', 'limit_up_ratio'
]
# ==================== 工具函数 ====================
def get_ch_client():
return Client(**CLICKHOUSE_CONFIG)
def code_to_ch_format(code: str) -> str:
if not code or len(code) != 6 or not code.isdigit():
return None
if code.startswith('6'):
return f"{code}.SH"
elif code.startswith('0') or code.startswith('3'):
return f"{code}.SZ"
else:
return f"{code}.BJ"
def time_to_slot(ts) -> str:
"""时间戳转时间片（HH:MM）"""
if isinstance(ts, str):
return ts
return ts.strftime('%H:%M')
# ==================== 基线加载 ====================
def load_baselines(baseline_dir: str = 'ml/data_v2/baselines') -> Dict[str, pd.DataFrame]:
"""加载时间片基线"""
baseline_file = os.path.join(baseline_dir, 'baselines.pkl')
if os.path.exists(baseline_file):
with open(baseline_file, 'rb') as f:
return pickle.load(f)
return {}
# ==================== 规则评分（基于 Z-Score）====================
def score_rules_zscore(row: Dict) -> Tuple[float, List[str]]:
"""
基于 Z-Score 的规则评分
设计思路：Z-Score 已经标准化，直接用阈值判断
"""
score = 0.0
triggered = []
alpha_zscore = row.get('alpha_zscore', 0)
amt_zscore = row.get('amt_zscore', 0)
rank_zscore = row.get('rank_zscore', 0)
momentum_3m = row.get('momentum_3m', 0)
momentum_5m = row.get('momentum_5m', 0)
limit_up_ratio = row.get('limit_up_ratio', 0)
alpha_zscore_abs = abs(alpha_zscore)
amt_zscore_abs = abs(amt_zscore)
# ========== Alpha Z-Score 规则 ==========
if alpha_zscore_abs >= 4.0:
score += 25
triggered.append('alpha_zscore_extreme')
elif alpha_zscore_abs >= 3.0:
score += 18
triggered.append('alpha_zscore_strong')
elif alpha_zscore_abs >= 2.0:
score += 10
triggered.append('alpha_zscore_moderate')
# ========== 成交额 Z-Score 规则 ==========
if amt_zscore >= 4.0:
score += 20
triggered.append('amt_zscore_extreme')
elif amt_zscore >= 3.0:
score += 12
triggered.append('amt_zscore_strong')
elif amt_zscore >= 2.0:
score += 6
triggered.append('amt_zscore_moderate')
# ========== 排名 Z-Score 规则 ==========
if abs(rank_zscore) >= 3.0:
score += 15
triggered.append('rank_zscore_extreme')
elif abs(rank_zscore) >= 2.0:
score += 8
triggered.append('rank_zscore_strong')
# ========== 动量规则 ==========
if momentum_3m >= 1.0:
score += 12
triggered.append('momentum_3m_strong')
elif momentum_3m >= 0.5:
score += 6
triggered.append('momentum_3m_moderate')
if momentum_5m >= 1.5:
score += 10
triggered.append('momentum_5m_strong')
# ========== 涨停比例规则 ==========
if limit_up_ratio >= 0.3:
score += 20
triggered.append('limit_up_extreme')
elif limit_up_ratio >= 0.15:
score += 12
triggered.append('limit_up_strong')
elif limit_up_ratio >= 0.08:
score += 5
triggered.append('limit_up_moderate')
# ========== 组合规则 ==========
# Alpha Z-Score + 成交额放大
if alpha_zscore_abs >= 2.0 and amt_zscore >= 2.0:
score += 15
triggered.append('combo_alpha_amt')
# Alpha Z-Score + 涨停
if alpha_zscore_abs >= 2.0 and limit_up_ratio >= 0.1:
score += 12
triggered.append('combo_alpha_limitup')
return min(score, 100), triggered
# ==================== ML 评分器 ====================
class MLScorerV2:
"""V2 ML 评分器"""
def __init__(self, model_dir: str = 'ml/checkpoints_v2'):
self.model_dir = model_dir
self.model = None
self.thresholds = None
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self._load_model()
def _load_model(self):
"""加载模型和阈值"""
model_path = os.path.join(self.model_dir, 'best_model.pt')
threshold_path = os.path.join(self.model_dir, 'thresholds.json')
config_path = os.path.join(self.model_dir, 'config.json')
if not os.path.exists(model_path):
print(f"警告: 模型文件不存在: {model_path}")
return
# 加载配置
with open(config_path, 'r') as f:
config = json.load(f)
# 创建模型
model_config = config.get('model', {})
self.model = TransformerAutoencoder(**model_config)
# 加载权重
checkpoint = torch.load(model_path, map_location=self.device)
self.model.load_state_dict(checkpoint['model_state_dict'])
self.model.to(self.device)
self.model.eval()
# 加载阈值
if os.path.exists(threshold_path):
with open(threshold_path, 'r') as f:
self.thresholds = json.load(f)
print(f"V2 模型加载完成: {model_path}")
@torch.no_grad()
def score_batch(self, sequences: np.ndarray) -> np.ndarray:
"""
批量计算 ML 分数
返回 0-100 的分数，越高越异常
"""
if self.model is None:
return np.zeros(len(sequences))
# 转换为 tensor
x = torch.FloatTensor(sequences).to(self.device)
# 计算重构误差
errors = self.model.compute_reconstruction_error(x, reduction='none')
# 取最后一个时刻的误差
last_errors = errors[:, -1].cpu().numpy()
# 转换为 0-100 分数
if self.thresholds:
p50 = self.thresholds.get('median', 0.1)
p99 = self.thresholds.get('p99', 1.0)
# 线性映射：p50 -> 50分，p99 -> 99分
scores = 50 + (last_errors - p50) / (p99 - p50) * 49
scores = np.clip(scores, 0, 100)
else:
# 没有阈值时，简单归一化
scores = last_errors * 100
scores = np.clip(scores, 0, 100)
return scores
# ==================== 实时数据管理器 ====================
class RealtimeDataManagerV2:
"""
V2 实时数据管理器
维护：
1. 每个概念的历史 Z-Score 序列（用于 LSTM 输入）
2. 每个概念的异动候选队列（用于持续性确认）
"""
def __init__(self, concepts: List[dict], baselines: Dict[str, pd.DataFrame]):
self.concepts = {c['concept_id']: c for c in concepts}
self.baselines = baselines
# 概念到股票的映射
self.concept_stocks = {c['concept_id']: set(c['stocks']) for c in concepts}
# 历史 Z-Score 序列（每个概念）
# {concept_id: deque([(timestamp, features_dict), ...], maxlen=seq_len)}
self.zscore_history = defaultdict(lambda: deque(maxlen=CONFIG['seq_len']))
# 异动候选队列（用于持续性确认）
# {concept_id: deque([(timestamp, score), ...], maxlen=confirm_window)}
self.anomaly_candidates = defaultdict(lambda: deque(maxlen=CONFIG['confirm_window']))
# 冷却期记录
self.cooldown = {}
# 上一次更新的时间戳
self.last_timestamp = None
def compute_zscore_features(
self,
concept_id: str,
timestamp,
alpha: float,
total_amt: float,
rank_pct: float,
limit_up_ratio: float
) -> Optional[Dict]:
"""计算单个概念单个时刻的 Z-Score 特征"""
if concept_id not in self.baselines:
return None
baseline = self.baselines[concept_id]
time_slot = time_to_slot(timestamp)
# 查找对应时间片的基线
bl_row = baseline[baseline['time_slot'] == time_slot]
if bl_row.empty:
return None
bl = bl_row.iloc[0]
# 检查样本量
if bl.get('sample_count', 0) < 10:
return None
# 计算 Z-Score
alpha_zscore = (alpha - bl['alpha_mean']) / bl['alpha_std']
amt_zscore = (total_amt - bl['amt_mean']) / bl['amt_std']
rank_zscore = (rank_pct - bl['rank_mean']) / bl['rank_std']
# 截断
clip = CONFIG['zscore_clip']
alpha_zscore = np.clip(alpha_zscore, -clip, clip)
amt_zscore = np.clip(amt_zscore, -clip, clip)
rank_zscore = np.clip(rank_zscore, -clip, clip)
# 计算动量（需要历史）
history = self.zscore_history[concept_id]
momentum_3m = 0
momentum_5m = 0
if len(history) >= 3:
recent_alphas = [h[1]['alpha'] for h in list(history)[-3:]]
older_alphas = [h[1]['alpha'] for h in list(history)[-6:-3]] if len(history) >= 6 else [alpha]
momentum_3m = np.mean(recent_alphas) - np.mean(older_alphas)
if len(history) >= 5:
recent_alphas = [h[1]['alpha'] for h in list(history)[-5:]]
older_alphas = [h[1]['alpha'] for h in list(history)[-10:-5]] if len(history) >= 10 else [alpha]
momentum_5m = np.mean(recent_alphas) - np.mean(older_alphas)
return {
'alpha': alpha,
'alpha_zscore': alpha_zscore,
'amt_zscore': amt_zscore,
'rank_zscore': rank_zscore,
'momentum_3m': momentum_3m,
'momentum_5m': momentum_5m,
'limit_up_ratio': limit_up_ratio,
'total_amt': total_amt,
'rank_pct': rank_pct,
}
def update(self, concept_id: str, timestamp, features: Dict):
"""更新概念的历史数据"""
self.zscore_history[concept_id].append((timestamp, features))
def get_sequence(self, concept_id: str) -> Optional[np.ndarray]:
"""获取用于 LSTM 的序列"""
history = self.zscore_history[concept_id]
if len(history) < CONFIG['seq_len']:
return None
# 提取特征
feature_list = []
for _, features in history:
feature_list.append([
features['alpha_zscore'],
features['amt_zscore'],
features['rank_zscore'],
features['momentum_3m'],
features['momentum_5m'],
features['limit_up_ratio'],
])
return np.array(feature_list)
def add_anomaly_candidate(self, concept_id: str, timestamp, score: float):
"""添加异动候选"""
self.anomaly_candidates[concept_id].append((timestamp, score))
def check_sustained_anomaly(self, concept_id: str, threshold: float) -> Tuple[bool, float]:
"""
检查是否为持续性异动
返回：(是否确认, 确认比例)
"""
candidates = self.anomaly_candidates[concept_id]
if len(candidates) < CONFIG['confirm_window']:
return False, 0.0
# 统计超过阈值的时刻数量
exceed_count = sum(1 for _, score in candidates if score >= threshold)
ratio = exceed_count / len(candidates)
return ratio >= CONFIG['confirm_ratio'], ratio
def check_cooldown(self, concept_id: str, timestamp) -> bool:
"""检查是否在冷却期"""
if concept_id not in self.cooldown:
return False
last_alert = self.cooldown[concept_id]
try:
diff = (timestamp - last_alert).total_seconds() / 60
return diff < CONFIG['cooldown_minutes']
except:
return False
def set_cooldown(self, concept_id: str, timestamp):
"""设置冷却期"""
self.cooldown[concept_id] = timestamp
# ==================== 异动检测器 V2 ====================
class AnomalyDetectorV2:
"""
V2 异动检测器
核心流程：
1. 获取实时数据
2. 计算 Z-Score 特征
3. 规则评分 + ML 评分
4. 持续性确认
5. 输出异动
"""
def __init__(
self,
model_dir: str = 'ml/checkpoints_v2',
baseline_dir: str = 'ml/data_v2/baselines'
):
# 加载概念
self.concepts = self._load_concepts()
# 加载基线
self.baselines = load_baselines(baseline_dir)
print(f"加载了 {len(self.baselines)} 个概念的基线")
# 初始化 ML 评分器
self.ml_scorer = MLScorerV2(model_dir)
# 初始化数据管理器
self.data_manager = RealtimeDataManagerV2(self.concepts, self.baselines)
# 收集所有股票
self.all_stocks = list(set(s for c in self.concepts for s in c['stocks']))
def _load_concepts(self) -> List[dict]:
"""从 ES 加载概念"""
concepts = []
query = {"query": {"match_all": {}}, "size": 100, "_source": ["concept_id", "concept", "stocks"]}
resp = ES_CLIENT.search(index=ES_INDEX, body=query, scroll='2m')
scroll_id = resp['_scroll_id']
hits = resp['hits']['hits']
while len(hits) > 0:
for hit in hits:
source = hit['_source']
stocks = []
if 'stocks' in source and isinstance(source['stocks'], list):
for stock in source['stocks']:
if isinstance(stock, dict) and 'code' in stock and stock['code']:
stocks.append(stock['code'])
if stocks:
concepts.append({
'concept_id': source.get('concept_id'),
'concept_name': source.get('concept'),
'stocks': stocks
})
resp = ES_CLIENT.scroll(scroll_id=scroll_id, scroll='2m')
scroll_id = resp['_scroll_id']
hits = resp['hits']['hits']
ES_CLIENT.clear_scroll(scroll_id=scroll_id)
print(f"加载了 {len(concepts)} 个概念")
return concepts
def detect(self, trade_date: str) -> List[Dict]:
"""
检测指定日期的异动
返回异动列表
"""
print(f"\n检测 {trade_date} 的异动...")
# 获取原始数据
raw_features = self._compute_raw_features(trade_date)
if raw_features.empty:
print("无数据")
return []
# 按时间排序
timestamps = sorted(raw_features['timestamp'].unique())
print(f"时间点数: {len(timestamps)}")
all_alerts = []
for ts in timestamps:
ts_data = raw_features[raw_features['timestamp'] == ts]
ts_alerts = self._process_timestamp(ts, ts_data, trade_date)
all_alerts.extend(ts_alerts)
print(f"共检测到 {len(all_alerts)} 个异动")
return all_alerts
def _compute_raw_features(self, trade_date: str) -> pd.DataFrame:
"""计算原始特征（同 prepare_data_v2）"""
# 这里简化处理，直接调用数据准备逻辑
from prepare_data_v2 import compute_raw_concept_features
return compute_raw_concept_features(trade_date, self.concepts, self.all_stocks)
def _process_timestamp(self, timestamp, ts_data: pd.DataFrame, trade_date: str) -> List[Dict]:
"""处理单个时间戳"""
alerts = []
candidates = [] # (concept_id, features, rule_score, triggered_rules)
for _, row in ts_data.iterrows():
concept_id = row['concept_id']
# 计算 Z-Score 特征
features = self.data_manager.compute_zscore_features(
concept_id, timestamp,
row['alpha'], row['total_amt'], row['rank_pct'], row['limit_up_ratio']
)
if features is None:
continue
# 更新历史
self.data_manager.update(concept_id, timestamp, features)
# 规则评分
rule_score, triggered_rules = score_rules_zscore(features)
# 收集候选
candidates.append((concept_id, features, rule_score, triggered_rules))
if not candidates:
return []
# 批量 ML 评分
sequences = []
valid_candidates = []
for concept_id, features, rule_score, triggered_rules in candidates:
seq = self.data_manager.get_sequence(concept_id)
if seq is not None:
sequences.append(seq)
valid_candidates.append((concept_id, features, rule_score, triggered_rules))
if not sequences:
return []
sequences = np.array(sequences)
ml_scores = self.ml_scorer.score_batch(sequences)
# 融合评分 + 持续性确认
for i, (concept_id, features, rule_score, triggered_rules) in enumerate(valid_candidates):
ml_score = ml_scores[i]
final_score = CONFIG['rule_weight'] * rule_score + CONFIG['ml_weight'] * ml_score
# 判断是否触发
is_triggered = (
rule_score >= CONFIG['rule_trigger'] or
ml_score >= CONFIG['ml_trigger'] or
final_score >= CONFIG['fusion_trigger']
)
# 添加到候选队列
self.data_manager.add_anomaly_candidate(concept_id, timestamp, final_score)
if not is_triggered:
continue
# 检查冷却期
if self.data_manager.check_cooldown(concept_id, timestamp):
continue
# 持续性确认
is_sustained, confirm_ratio = self.data_manager.check_sustained_anomaly(
concept_id, CONFIG['fusion_trigger']
)
if not is_sustained:
continue
# 确认为异动！
self.data_manager.set_cooldown(concept_id, timestamp)
# 确定异动类型
alpha = features['alpha']
if alpha >= 1.5:
alert_type = 'surge_up'
elif alpha <= -1.5:
alert_type = 'surge_down'
elif features['amt_zscore'] >= 3.0:
alert_type = 'volume_spike'
else:
alert_type = 'surge'
# 确定触发原因
if rule_score >= CONFIG['rule_trigger']:
trigger_reason = f'规则({rule_score:.0f})+持续确认({confirm_ratio:.0%})'
elif ml_score >= CONFIG['ml_trigger']:
trigger_reason = f'ML({ml_score:.0f})+持续确认({confirm_ratio:.0%})'
else:
trigger_reason = f'融合({final_score:.0f})+持续确认({confirm_ratio:.0%})'
alerts.append({
'concept_id': concept_id,
'concept_name': self.data_manager.concepts.get(concept_id, {}).get('concept_name', concept_id),
'alert_time': timestamp,
'trade_date': trade_date,
'alert_type': alert_type,
'final_score': final_score,
'rule_score': rule_score,
'ml_score': ml_score,
'trigger_reason': trigger_reason,
'confirm_ratio': confirm_ratio,
'alpha': alpha,
'alpha_zscore': features['alpha_zscore'],
'amt_zscore': features['amt_zscore'],
'rank_zscore': features['rank_zscore'],
'momentum_3m': features['momentum_3m'],
'momentum_5m': features['momentum_5m'],
'limit_up_ratio': features['limit_up_ratio'],
'triggered_rules': triggered_rules,
})
# 每分钟最多 N 个
if len(alerts) > CONFIG['max_alerts_per_minute']:
alerts = sorted(alerts, key=lambda x: x['final_score'], reverse=True)
alerts = alerts[:CONFIG['max_alerts_per_minute']]
return alerts
# ==================== 主函数 ====================
def main():
import argparse
parser = argparse.ArgumentParser(description='V2 异动检测器')
parser.add_argument('--date', type=str, default=None, help='检测日期（默认今天）')
parser.add_argument('--model_dir', type=str, default='ml/checkpoints_v2')
parser.add_argument('--baseline_dir', type=str, default='ml/data_v2/baselines')
args = parser.parse_args()
trade_date = args.date or datetime.now().strftime('%Y-%m-%d')
detector = AnomalyDetectorV2(
model_dir=args.model_dir,
baseline_dir=args.baseline_dir
)
alerts = detector.detect(trade_date)
print(f"\n检测结果:")
for alert in alerts[:20]:
print(f" [{alert['alert_time'].strftime('%H:%M') if hasattr(alert['alert_time'], 'strftime') else alert['alert_time']}] "
f"{alert['concept_name']} ({alert['alert_type']}) "
f"分数={alert['final_score']:.0f} "
f"确认率={alert['confirm_ratio']:.0%}")
if len(alerts) > 20:
print(f" ... 共 {len(alerts)} 个异动")
if __name__ == "__main__":
main()

526
ml/enhanced_detector.py
View File

@@ -1,526 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
增强版概念异动检测器
融合两种检测方法：
1. Alpha-based Z-Score（规则方法，实时性好）
2. Transformer Autoencoder（ML方法，更准确）
使用策略：
- 当 ML 模型可用且历史数据足够时，优先使用 ML 方法
- 否则回退到 Alpha-based 方法
- 可以配置两种方法的融合权重
"""
import os
import sys
import logging
from datetime import datetime
from pathlib import Path
from typing import Dict, List, Tuple, Optional
from dataclasses import dataclass, field
from collections import deque
import numpy as np
# 添加父目录到路径
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
logger = logging.getLogger(__name__)
# ==================== 配置 ====================
ENHANCED_CONFIG = {
# 融合策略
'fusion_mode': 'adaptive', # 'ml_only', 'alpha_only', 'adaptive', 'ensemble'
# ML 权重（在 ensemble 模式下）
'ml_weight': 0.6,
'alpha_weight': 0.4,
# ML 模型配置
'ml_checkpoint_dir': 'ml/checkpoints',
'ml_threshold_key': 'p95', # p90, p95, p99
# Alpha 配置（与 concept_alert_alpha.py 一致）
'alpha_zscore_threshold': 2.0,
'alpha_absolute_threshold': 1.5,
'alpha_history_window': 60,
'alpha_min_history': 5,
# 共享配置
'cooldown_minutes': 8,
'max_alerts_per_minute': 15,
'min_alpha_abs': 0.5,
}
# 特征配置（与训练一致）
FEATURE_NAMES = [
'alpha',
'alpha_delta',
'amt_ratio',
'amt_delta',
'rank_pct',
'limit_up_ratio',
]
# ==================== 数据结构 ====================
@dataclass
class AlphaStats:
"""概念的Alpha统计信息"""
history: deque = field(default_factory=lambda: deque(maxlen=ENHANCED_CONFIG['alpha_history_window']))
mean: float = 0.0
std: float = 1.0
def update(self, alpha: float):
self.history.append(alpha)
if len(self.history) >= 2:
self.mean = np.mean(self.history)
self.std = max(np.std(self.history), 0.1)
def get_zscore(self, alpha: float) -> float:
if len(self.history) < ENHANCED_CONFIG['alpha_min_history']:
return 0.0
return (alpha - self.mean) / self.std
def is_ready(self) -> bool:
return len(self.history) >= ENHANCED_CONFIG['alpha_min_history']
@dataclass
class ConceptFeatures:
"""概念的实时特征"""
alpha: float = 0.0
alpha_delta: float = 0.0
amt_ratio: float = 1.0
amt_delta: float = 0.0
rank_pct: float = 0.5
limit_up_ratio: float = 0.0
def to_dict(self) -> Dict[str, float]:
return {
'alpha': self.alpha,
'alpha_delta': self.alpha_delta,
'amt_ratio': self.amt_ratio,
'amt_delta': self.amt_delta,
'rank_pct': self.rank_pct,
'limit_up_ratio': self.limit_up_ratio,
}
# ==================== 增强检测器 ====================
class EnhancedAnomalyDetector:
"""
增强版异动检测器
融合 Alpha-based 和 ML 两种方法
"""
def __init__(
self,
config: Dict = None,
ml_enabled: bool = True
):
self.config = config or ENHANCED_CONFIG
self.ml_enabled = ml_enabled
self.ml_detector = None
# Alpha 统计
self.alpha_stats: Dict[str, AlphaStats] = {}
# 特征历史（用于计算 delta）
self.feature_history: Dict[str, deque] = {}
# 冷却记录
self.cooldown_cache: Dict[str, datetime] = {}
# 尝试加载 ML 模型
if ml_enabled:
self._load_ml_model()
logger.info(f"EnhancedAnomalyDetector 初始化完成")
logger.info(f" 融合模式: {self.config['fusion_mode']}")
logger.info(f" ML 可用: {self.ml_detector is not None}")
def _load_ml_model(self):
"""加载 ML 模型"""
try:
from inference import ConceptAnomalyDetector
checkpoint_dir = Path(__file__).parent / 'checkpoints'
if (checkpoint_dir / 'best_model.pt').exists():
self.ml_detector = ConceptAnomalyDetector(
checkpoint_dir=str(checkpoint_dir),
threshold_key=self.config['ml_threshold_key']
)
logger.info("ML 模型加载成功")
else:
logger.warning(f"ML 模型不存在: {checkpoint_dir / 'best_model.pt'}")
except Exception as e:
logger.warning(f"ML 模型加载失败: {e}")
self.ml_detector = None
def _get_alpha_stats(self, concept_id: str) -> AlphaStats:
"""获取或创建 Alpha 统计"""
if concept_id not in self.alpha_stats:
self.alpha_stats[concept_id] = AlphaStats()
return self.alpha_stats[concept_id]
def _get_feature_history(self, concept_id: str) -> deque:
"""获取特征历史"""
if concept_id not in self.feature_history:
self.feature_history[concept_id] = deque(maxlen=10)
return self.feature_history[concept_id]
def _check_cooldown(self, concept_id: str, current_time: datetime) -> bool:
"""检查冷却"""
if concept_id not in self.cooldown_cache:
return False
last_alert = self.cooldown_cache[concept_id]
cooldown_td = (current_time - last_alert).total_seconds() / 60
return cooldown_td < self.config['cooldown_minutes']
def _set_cooldown(self, concept_id: str, current_time: datetime):
"""设置冷却"""
self.cooldown_cache[concept_id] = current_time
def compute_features(
self,
concept_id: str,
alpha: float,
amt_ratio: float,
rank_pct: float,
limit_up_ratio: float
) -> ConceptFeatures:
"""
计算概念的完整特征
Args:
concept_id: 概念ID
alpha: 当前超额收益
amt_ratio: 成交额比率
rank_pct: 排名百分位
limit_up_ratio: 涨停股占比
Returns:
完整特征
"""
history = self._get_feature_history(concept_id)
# 计算变化率
alpha_delta = 0.0
amt_delta = 0.0
if len(history) > 0:
last_features = history[-1]
alpha_delta = alpha - last_features.alpha
if last_features.amt_ratio > 0:
amt_delta = (amt_ratio - last_features.amt_ratio) / last_features.amt_ratio
features = ConceptFeatures(
alpha=alpha,
alpha_delta=alpha_delta,
amt_ratio=amt_ratio,
amt_delta=amt_delta,
rank_pct=rank_pct,
limit_up_ratio=limit_up_ratio,
)
# 更新历史
history.append(features)
return features
def detect_alpha_anomaly(
self,
concept_id: str,
alpha: float
) -> Tuple[bool, float, str]:
"""
Alpha-based 异动检测
Returns:
is_anomaly: 是否异动
score: 异动分数（Z-Score 绝对值）
reason: 触发原因
"""
stats = self._get_alpha_stats(concept_id)
# 计算 Z-Score（在更新前）
zscore = stats.get_zscore(alpha)
# 更新统计
stats.update(alpha)
# 判断
if stats.is_ready():
if abs(zscore) >= self.config['alpha_zscore_threshold']:
return True, abs(zscore), f"Z={zscore:.2f}"
else:
if abs(alpha) >= self.config['alpha_absolute_threshold']:
fake_zscore = alpha / 0.5
return True, abs(fake_zscore), f"Alpha={alpha:+.2f}%"
return False, abs(zscore) if zscore else 0.0, ""
def detect_ml_anomaly(
self,
concept_id: str,
features: ConceptFeatures
) -> Tuple[bool, float]:
"""
ML-based 异动检测
Returns:
is_anomaly: 是否异动
score: 异动分数（重构误差）
"""
if self.ml_detector is None:
return False, 0.0
try:
is_anomaly, score = self.ml_detector.detect(
concept_id,
features.to_dict()
)
return is_anomaly, score or 0.0
except Exception as e:
logger.warning(f"ML 检测失败: {e}")
return False, 0.0
def detect(
self,
concept_id: str,
concept_name: str,
alpha: float,
amt_ratio: float,
rank_pct: float,
limit_up_ratio: float,
change_pct: float,
index_change: float,
current_time: datetime,
**extra_data
) -> Optional[Dict]:
"""
融合检测
Args:
concept_id: 概念ID
concept_name: 概念名称
alpha: 超额收益
amt_ratio: 成交额比率
rank_pct: 排名百分位
limit_up_ratio: 涨停股占比
change_pct: 概念涨跌幅
index_change: 大盘涨跌幅
current_time: 当前时间
**extra_data: 其他数据（limit_up_count, stock_count 等）
Returns:
异动信息（如果触发），否则 None
"""
# Alpha 太小，不关注
if abs(alpha) < self.config['min_alpha_abs']:
return None
# 检查冷却
if self._check_cooldown(concept_id, current_time):
return None
# 计算特征
features = self.compute_features(
concept_id, alpha, amt_ratio, rank_pct, limit_up_ratio
)
# 执行检测
fusion_mode = self.config['fusion_mode']
alpha_anomaly, alpha_score, alpha_reason = self.detect_alpha_anomaly(concept_id, alpha)
ml_anomaly, ml_score = False, 0.0
if fusion_mode in ('ml_only', 'adaptive', 'ensemble'):
ml_anomaly, ml_score = self.detect_ml_anomaly(concept_id, features)
# 根据融合模式判断
is_anomaly = False
final_score = 0.0
detection_method = ''
if fusion_mode == 'alpha_only':
is_anomaly = alpha_anomaly
final_score = alpha_score
detection_method = 'alpha'
elif fusion_mode == 'ml_only':
is_anomaly = ml_anomaly
final_score = ml_score
detection_method = 'ml'
elif fusion_mode == 'adaptive':
# 优先 ML，回退 Alpha
if self.ml_detector and ml_score > 0:
is_anomaly = ml_anomaly
final_score = ml_score
detection_method = 'ml'
else:
is_anomaly = alpha_anomaly
final_score = alpha_score
detection_method = 'alpha'
elif fusion_mode == 'ensemble':
# 加权融合
# 归一化分数
norm_alpha = min(alpha_score / 5.0, 1.0) # Z > 5 视为 1.0
norm_ml = min(ml_score / (self.ml_detector.threshold if self.ml_detector else 1.0), 1.0)
final_score = (
self.config['alpha_weight'] * norm_alpha +
self.config['ml_weight'] * norm_ml
)
is_anomaly = final_score > 0.5 or alpha_anomaly or ml_anomaly
detection_method = 'ensemble'
if not is_anomaly:
return None
# 构建异动记录
self._set_cooldown(concept_id, current_time)
alert_type = 'surge_up' if alpha > 0 else 'surge_down'
alert = {
'concept_id': concept_id,
'concept_name': concept_name,
'alert_type': alert_type,
'alert_time': current_time,
'change_pct': change_pct,
'alpha': alpha,
'alpha_zscore': alpha_score,
'index_change_pct': index_change,
'detection_method': detection_method,
'alpha_score': alpha_score,
'ml_score': ml_score,
'final_score': final_score,
**extra_data
}
return alert
def batch_detect(
self,
concepts_data: List[Dict],
current_time: datetime
) -> List[Dict]:
"""
批量检测
Args:
concepts_data: 概念数据列表
current_time: 当前时间
Returns:
异动列表（按分数排序，限制数量）
"""
alerts = []
for data in concepts_data:
alert = self.detect(
concept_id=data['concept_id'],
concept_name=data['concept_name'],
alpha=data.get('alpha', 0),
amt_ratio=data.get('amt_ratio', 1.0),
rank_pct=data.get('rank_pct', 0.5),
limit_up_ratio=data.get('limit_up_ratio', 0),
change_pct=data.get('change_pct', 0),
index_change=data.get('index_change', 0),
current_time=current_time,
limit_up_count=data.get('limit_up_count', 0),
limit_down_count=data.get('limit_down_count', 0),
stock_count=data.get('stock_count', 0),
concept_type=data.get('concept_type', 'leaf'),
)
if alert:
alerts.append(alert)
# 排序并限制数量
alerts.sort(key=lambda x: x['final_score'], reverse=True)
return alerts[:self.config['max_alerts_per_minute']]
def reset(self):
"""重置所有状态（新交易日）"""
self.alpha_stats.clear()
self.feature_history.clear()
self.cooldown_cache.clear()
if self.ml_detector:
self.ml_detector.clear_history()
logger.info("检测器状态已重置")
# ==================== 测试 ====================
if __name__ == "__main__":
import random
print("测试 EnhancedAnomalyDetector...")
# 初始化
detector = EnhancedAnomalyDetector(ml_enabled=False) # 不加载 ML（可能不存在）
# 模拟数据
concepts = [
{'concept_id': 'ai_001', 'concept_name': '人工智能'},
{'concept_id': 'chip_002', 'concept_name': '芯片半导体'},
{'concept_id': 'car_003', 'concept_name': '新能源汽车'},
]
print("\n模拟实时检测...")
current_time = datetime.now()
for minute in range(50):
concepts_data = []
for c in concepts:
# 生成随机数据
alpha = random.gauss(0, 0.8)
amt_ratio = max(0.3, random.gauss(1, 0.3))
rank_pct = random.random()
limit_up_ratio = random.random() * 0.1
# 模拟异动（第30分钟人工智能暴涨）
if minute == 30 and c['concept_id'] == 'ai_001':
alpha = 4.5
amt_ratio = 2.5
limit_up_ratio = 0.3
concepts_data.append({
**c,
'alpha': alpha,
'amt_ratio': amt_ratio,
'rank_pct': rank_pct,
'limit_up_ratio': limit_up_ratio,
'change_pct': alpha + 0.5,
'index_change': 0.5,
})
# 检测
alerts = detector.batch_detect(concepts_data, current_time)
if alerts:
for alert in alerts:
print(f" t={minute:02d} 🔥 {alert['concept_name']} "
f"Alpha={alert['alpha']:+.2f}% "
f"Score={alert['final_score']:.2f} "
f"Method={alert['detection_method']}")
current_time = current_time.replace(minute=current_time.minute + 1 if current_time.minute < 59 else 0)
print("\n测试完成！")

455
ml/inference.py
View File

@@ -1,455 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
概念异动检测推理服务
在实时场景中使用训练好的 Transformer Autoencoder 进行异动检测
使用方法：
from ml.inference import ConceptAnomalyDetector
detector = ConceptAnomalyDetector('ml/checkpoints')
# 检测异动
features = {...} # 实时特征数据
is_anomaly, score = detector.detect(features)
"""
import os
import json
from pathlib import Path
from typing import Dict, List, Tuple, Optional
from collections import deque
import numpy as np
import torch
from model import TransformerAutoencoder
class ConceptAnomalyDetector:
"""
概念异动检测器
使用训练好的 Transformer Autoencoder 进行实时异动检测
"""
def __init__(
self,
checkpoint_dir: str = 'ml/checkpoints',
device: str = 'auto',
threshold_key: str = 'p95'
):
"""
初始化检测器
Args:
checkpoint_dir: 模型检查点目录
device: 设备 (auto/cuda/cpu)
threshold_key: 使用的阈值键 (p90/p95/p99)
"""
self.checkpoint_dir = Path(checkpoint_dir)
self.threshold_key = threshold_key
# 设备选择
if device == 'auto':
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
else:
self.device = torch.device(device)
# 加载配置
self._load_config()
# 加载模型
self._load_model()
# 加载阈值
self._load_thresholds()
# 加载标准化统计量
self._load_normalization_stats()
# 概念历史数据缓存
# {concept_name: deque(maxlen=seq_len)}
self.history_cache: Dict[str, deque] = {}
print(f"ConceptAnomalyDetector 初始化完成")
print(f" 设备: {self.device}")
print(f" 阈值: {self.threshold_key} = {self.threshold:.6f}")
print(f" 序列长度: {self.seq_len}")
def _load_config(self):
"""加载配置"""
config_path = self.checkpoint_dir / 'config.json'
if not config_path.exists():
raise FileNotFoundError(f"配置文件不存在: {config_path}")
with open(config_path, 'r') as f:
self.config = json.load(f)
self.features = self.config['features']
self.seq_len = self.config['seq_len']
self.model_config = self.config['model']
def _load_model(self):
"""加载模型"""
model_path = self.checkpoint_dir / 'best_model.pt'
if not model_path.exists():
raise FileNotFoundError(f"模型文件不存在: {model_path}")
# 创建模型
self.model = TransformerAutoencoder(**self.model_config)
# 加载权重
checkpoint = torch.load(model_path, map_location=self.device)
self.model.load_state_dict(checkpoint['model_state_dict'])
self.model.to(self.device)
self.model.eval()
print(f"模型已加载: {model_path}")
def _load_thresholds(self):
"""加载阈值"""
thresholds_path = self.checkpoint_dir / 'thresholds.json'
if not thresholds_path.exists():
raise FileNotFoundError(f"阈值文件不存在: {thresholds_path}")
with open(thresholds_path, 'r') as f:
self.thresholds = json.load(f)
if self.threshold_key not in self.thresholds:
available_keys = list(self.thresholds.keys())
raise KeyError(f"阈值键 '{self.threshold_key}' 不存在，可用: {available_keys}")
self.threshold = self.thresholds[self.threshold_key]
def _load_normalization_stats(self):
"""加载标准化统计量"""
stats_path = self.checkpoint_dir / 'normalization_stats.json'
if not stats_path.exists():
raise FileNotFoundError(f"标准化统计量文件不存在: {stats_path}")
with open(stats_path, 'r') as f:
stats = json.load(f)
self.norm_mean = np.array(stats['mean'])
self.norm_std = np.array(stats['std'])
def normalize(self, features: np.ndarray) -> np.ndarray:
"""标准化特征"""
return (features - self.norm_mean) / self.norm_std
def update_history(
self,
concept_name: str,
features: Dict[str, float]
):
"""
更新概念历史数据
Args:
concept_name: 概念名称
features: 当前时刻的特征字典
"""
# 初始化历史缓存
if concept_name not in self.history_cache:
self.history_cache[concept_name] = deque(maxlen=self.seq_len)
# 提取特征向量
feature_vector = np.array([
features.get(f, 0.0) for f in self.features
])
# 处理异常值
feature_vector = np.nan_to_num(feature_vector, nan=0.0, posinf=0.0, neginf=0.0)
# 添加到历史
self.history_cache[concept_name].append(feature_vector)
def get_history_length(self, concept_name: str) -> int:
"""获取概念的历史数据长度"""
if concept_name not in self.history_cache:
return 0
return len(self.history_cache[concept_name])
@torch.no_grad()
def detect(
self,
concept_name: str,
features: Dict[str, float] = None,
return_score: bool = True
) -> Tuple[bool, Optional[float]]:
"""
检测概念是否异动
Args:
concept_name: 概念名称
features: 当前时刻的特征（如果提供，会先更新历史）
return_score: 是否返回异动分数
Returns:
is_anomaly: 是否异动
score: 异动分数（如果 return_score=True）
"""
# 更新历史
if features is not None:
self.update_history(concept_name, features)
# 检查历史数据是否足够
if concept_name not in self.history_cache:
return False, None
history = self.history_cache[concept_name]
if len(history) < self.seq_len:
return False, None
# 构建输入序列
sequence = np.array(list(history)) # (seq_len, n_features)
# 标准化
sequence = self.normalize(sequence)
# 转为 tensor
x = torch.FloatTensor(sequence).unsqueeze(0) # (1, seq_len, n_features)
x = x.to(self.device)
# 计算重构误差
error = self.model.compute_reconstruction_error(x, reduction='none')
# 取最后一个时刻的误差作为当前分数
score = error[0, -1].item()
# 判断是否异动
is_anomaly = score > self.threshold
if return_score:
return is_anomaly, score
else:
return is_anomaly, None
@torch.no_grad()
def batch_detect(
self,
concept_features: Dict[str, Dict[str, float]]
) -> Dict[str, Tuple[bool, float]]:
"""
批量检测多个概念
Args:
concept_features: {concept_name: {feature_name: value}}
Returns:
results: {concept_name: (is_anomaly, score)}
"""
results = {}
for concept_name, features in concept_features.items():
is_anomaly, score = self.detect(concept_name, features)
results[concept_name] = (is_anomaly, score)
return results
def get_anomaly_type(
self,
concept_name: str,
features: Dict[str, float]
) -> str:
"""
判断异动类型
Args:
concept_name: 概念名称
features: 当前特征
Returns:
anomaly_type: 'surge_up' / 'surge_down' / 'normal'
"""
is_anomaly, score = self.detect(concept_name, features)
if not is_anomaly:
return 'normal'
# 根据 alpha 判断涨跌
alpha = features.get('alpha', 0.0)
if alpha > 0:
return 'surge_up'
else:
return 'surge_down'
def get_top_anomalies(
self,
concept_features: Dict[str, Dict[str, float]],
top_k: int = 10
) -> List[Tuple[str, float, str]]:
"""
获取异动分数最高的 top_k 个概念
Args:
concept_features: {concept_name: {feature_name: value}}
top_k: 返回数量
Returns:
anomalies: [(concept_name, score, anomaly_type), ...]
"""
results = self.batch_detect(concept_features)
# 按分数排序
sorted_results = sorted(
[(name, is_anomaly, score) for name, (is_anomaly, score) in results.items() if score is not None],
key=lambda x: x[2],
reverse=True
)
# 取 top_k
top_anomalies = []
for name, is_anomaly, score in sorted_results[:top_k]:
if is_anomaly:
alpha = concept_features[name].get('alpha', 0.0)
anomaly_type = 'surge_up' if alpha > 0 else 'surge_down'
top_anomalies.append((name, score, anomaly_type))
return top_anomalies
def clear_history(self, concept_name: str = None):
"""
清除历史缓存
Args:
concept_name: 概念名称（如果为 None，清除所有）
"""
if concept_name is None:
self.history_cache.clear()
elif concept_name in self.history_cache:
del self.history_cache[concept_name]
# ==================== 集成到现有系统 ====================
class MLAnomalyService:
"""
ML 异动检测服务
用于替换或增强现有的 Alpha-based 检测
"""
def __init__(
self,
checkpoint_dir: str = 'ml/checkpoints',
fallback_to_alpha: bool = True
):
"""
Args:
checkpoint_dir: 模型检查点目录
fallback_to_alpha: 当 ML 模型不可用时是否回退到 Alpha 方法
"""
self.fallback_to_alpha = fallback_to_alpha
self.ml_detector = None
try:
self.ml_detector = ConceptAnomalyDetector(checkpoint_dir)
print("ML 异动检测服务初始化成功")
except Exception as e:
print(f"ML 模型加载失败: {e}")
if not fallback_to_alpha:
raise
print("将回退到 Alpha-based 检测")
def is_ml_available(self) -> bool:
"""检查 ML 模型是否可用"""
return self.ml_detector is not None
def detect_anomaly(
self,
concept_name: str,
features: Dict[str, float],
alpha_threshold: float = 2.0
) -> Tuple[bool, float, str]:
"""
检测异动
Args:
concept_name: 概念名称
features: 特征字典（需包含 alpha, amt_ratio 等）
alpha_threshold: Alpha Z-Score 阈值（用于回退）
Returns:
is_anomaly: 是否异动
score: 异动分数
method: 检测方法 ('ml' / 'alpha')
"""
# 优先使用 ML 检测
if self.ml_detector is not None:
history_len = self.ml_detector.get_history_length(concept_name)
# 历史数据足够时使用 ML
if history_len >= self.ml_detector.seq_len - 1:
is_anomaly, score = self.ml_detector.detect(concept_name, features)
if score is not None:
return is_anomaly, score, 'ml'
else:
# 更新历史但使用 Alpha 方法
self.ml_detector.update_history(concept_name, features)
# 回退到 Alpha 方法
if self.fallback_to_alpha:
alpha = features.get('alpha', 0.0)
alpha_zscore = features.get('alpha_zscore', 0.0)
is_anomaly = abs(alpha_zscore) > alpha_threshold
score = abs(alpha_zscore)
return is_anomaly, score, 'alpha'
return False, 0.0, 'none'
# ==================== 测试 ====================
if __name__ == "__main__":
import random
print("测试 ConceptAnomalyDetector...")
# 检查模型是否存在
checkpoint_dir = Path('ml/checkpoints')
if not (checkpoint_dir / 'best_model.pt').exists():
print("模型文件不存在，跳过测试")
print("请先运行 train.py 训练模型")
exit(0)
# 初始化检测器
detector = ConceptAnomalyDetector('ml/checkpoints')
# 模拟数据
print("\n模拟实时检测...")
concept_name = "人工智能"
for i in range(40):
# 生成随机特征
features = {
'alpha': random.gauss(0, 1),
'alpha_delta': random.gauss(0, 0.5),
'amt_ratio': random.gauss(1, 0.3),
'amt_delta': random.gauss(0, 0.2),
'rank_pct': random.random(),
'limit_up_ratio': random.random() * 0.1,
}
# 在第 35 分钟模拟异动
if i == 35:
features['alpha'] = 5.0
features['alpha_delta'] = 2.0
features['amt_ratio'] = 3.0
is_anomaly, score = detector.detect(concept_name, features)
history_len = detector.get_history_length(concept_name)
if score is not None:
status = "🔥 异动！" if is_anomaly else "正常"
print(f" t={i:02d} | 历史={history_len} | 分数={score:.4f} | {status}")
else:
print(f" t={i:02d} | 历史={history_len} | 数据不足")
print("\n测试完成！")

393
ml/model.py
View File

@@ -1,393 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
LSTM Autoencoder 模型定义
用于概念异动检测：
- 学习"正常"市场模式
- 重构误差大的时刻 = 异动
模型结构（简洁有效）:
┌─────────────────────────────────────┐
│ 输入: (batch, seq_len, n_features) │
│ 过去30分钟的特征序列 │
├─────────────────────────────────────┤
│ LSTM Encoder │
│ - 双向 LSTM │
│ - 输出最后隐藏状态 │
├─────────────────────────────────────┤
│ Bottleneck (压缩层) │
│ 降维到 latent_dim（关键！） │
├─────────────────────────────────────┤
│ LSTM Decoder │
│ - 单向 LSTM │
│ - 重构序列 │
├─────────────────────────────────────┤
│ 输出: (batch, seq_len, n_features) │
│ 重构的特征序列 │
└─────────────────────────────────────┘
为什么用 LSTM 而不是 Transformer:
1. 参数更少，不容易过拟合
2. 对于 6 维特征足够用
3. 训练更稳定
4. 瓶颈约束更容易控制
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple
class LSTMAutoencoder(nn.Module):
"""
LSTM Autoencoder for Anomaly Detection
设计原则：
- 足够简单，避免过拟合
- 瓶颈层严格限制，迫使模型只学习主要模式
- 异常难以通过狭窄瓶颈，重构误差大
"""
def __init__(
self,
n_features: int = 6,
hidden_dim: int = 32, # LSTM 隐藏维度（小！）
latent_dim: int = 4, # 瓶颈维度（非常小！关键参数）
num_layers: int = 1, # LSTM 层数
dropout: float = 0.2,
bidirectional: bool = True, # 双向编码器
):
super().__init__()
self.n_features = n_features
self.hidden_dim = hidden_dim
self.latent_dim = latent_dim
self.num_layers = num_layers
self.bidirectional = bidirectional
self.num_directions = 2 if bidirectional else 1
# Encoder: 双向 LSTM
self.encoder = nn.LSTM(
input_size=n_features,
hidden_size=hidden_dim,
num_layers=num_layers,
batch_first=True,
dropout=dropout if num_layers > 1 else 0,
bidirectional=bidirectional
)
# Bottleneck: 压缩到极小的 latent space
encoder_output_dim = hidden_dim * self.num_directions
self.bottleneck_down = nn.Sequential(
nn.Linear(encoder_output_dim, latent_dim),
nn.Tanh(), # 限制范围，增加约束
)
# 使用 LeakyReLU 替代 ReLU
# 原因：Z-Score 数据范围是 [-5, +5]，ReLU 会截断负值，丢失跌幅信息
# LeakyReLU 保留负值信号（乘以 0.1）
self.bottleneck_up = nn.Sequential(
nn.Linear(latent_dim, hidden_dim),
nn.LeakyReLU(negative_slope=0.1),
)
# Decoder: 单向 LSTM
self.decoder = nn.LSTM(
input_size=hidden_dim,
hidden_size=hidden_dim,
num_layers=num_layers,
batch_first=True,
dropout=dropout if num_layers > 1 else 0,
bidirectional=False # 解码器用单向
)
# 输出层
self.output_layer = nn.Linear(hidden_dim, n_features)
# Dropout
self.dropout = nn.Dropout(dropout)
# 初始化
self._init_weights()
def _init_weights(self):
"""初始化权重"""
for name, param in self.named_parameters():
if 'weight_ih' in name:
nn.init.xavier_uniform_(param)
elif 'weight_hh' in name:
nn.init.orthogonal_(param)
elif 'bias' in name:
nn.init.zeros_(param)
def encode(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
编码器
Args:
x: (batch, seq_len, n_features)
Returns:
latent: (batch, seq_len, latent_dim) 每个时间步的压缩表示
encoder_outputs: (batch, seq_len, hidden_dim * num_directions)
"""
# LSTM 编码
encoder_outputs, (h_n, c_n) = self.encoder(x)
# encoder_outputs: (batch, seq_len, hidden_dim * num_directions)
encoder_outputs = self.dropout(encoder_outputs)
# 压缩到 latent space（对每个时间步）
latent = self.bottleneck_down(encoder_outputs)
# latent: (batch, seq_len, latent_dim)
return latent, encoder_outputs
def decode(self, latent: torch.Tensor, seq_len: int) -> torch.Tensor:
"""
解码器
Args:
latent: (batch, seq_len, latent_dim)
seq_len: 序列长度
Returns:
output: (batch, seq_len, n_features)
"""
# 从 latent space 恢复
decoder_input = self.bottleneck_up(latent)
# decoder_input: (batch, seq_len, hidden_dim)
# LSTM 解码
decoder_outputs, _ = self.decoder(decoder_input)
# decoder_outputs: (batch, seq_len, hidden_dim)
decoder_outputs = self.dropout(decoder_outputs)
# 投影到原始特征空间
output = self.output_layer(decoder_outputs)
# output: (batch, seq_len, n_features)
return output
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
前向传播
Args:
x: (batch, seq_len, n_features)
Returns:
output: (batch, seq_len, n_features) 重构结果
latent: (batch, seq_len, latent_dim) 隐向量
"""
batch_size, seq_len, _ = x.shape
# 编码
latent, _ = self.encode(x)
# 解码
output = self.decode(latent, seq_len)
return output, latent
def compute_reconstruction_error(
self,
x: torch.Tensor,
reduction: str = 'none'
) -> torch.Tensor:
"""
计算重构误差
Args:
x: (batch, seq_len, n_features)
reduction: 'none' | 'mean' | 'sum'
Returns:
error: 重构误差
"""
output, _ = self.forward(x)
# MSE per feature per timestep
error = F.mse_loss(output, x, reduction='none')
if reduction == 'none':
# (batch, seq_len, n_features) -> (batch, seq_len)
return error.mean(dim=-1)
elif reduction == 'mean':
return error.mean()
elif reduction == 'sum':
return error.sum()
else:
raise ValueError(f"Unknown reduction: {reduction}")
def detect_anomaly(
self,
x: torch.Tensor,
threshold: float = None,
return_scores: bool = True
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
"""
检测异动
Args:
x: (batch, seq_len, n_features)
threshold: 异动阈值（如果为 None，只返回分数）
return_scores: 是否返回异动分数
Returns:
is_anomaly: (batch, seq_len) bool tensor (if threshold is not None)
scores: (batch, seq_len) 异动分数 (if return_scores)
"""
scores = self.compute_reconstruction_error(x, reduction='none')
is_anomaly = None
if threshold is not None:
is_anomaly = scores > threshold
if return_scores:
return is_anomaly, scores
else:
return is_anomaly, None
# 为了兼容性，创建别名
TransformerAutoencoder = LSTMAutoencoder
# ==================== 损失函数 ====================
class AnomalyDetectionLoss(nn.Module):
"""
异动检测损失函数
简单的 MSE 重构损失
"""
def __init__(
self,
feature_weights: torch.Tensor = None,
):
super().__init__()
self.feature_weights = feature_weights
def forward(
self,
output: torch.Tensor,
target: torch.Tensor,
latent: torch.Tensor = None
) -> Tuple[torch.Tensor, dict]:
"""
Args:
output: (batch, seq_len, n_features) 重构结果
target: (batch, seq_len, n_features) 原始输入
latent: (batch, seq_len, latent_dim) 隐向量（未使用）
Returns:
loss: 总损失
loss_dict: 各项损失详情
"""
# 重构损失 (MSE)
mse = F.mse_loss(output, target, reduction='none')
# 特征加权（可选）
if self.feature_weights is not None:
weights = self.feature_weights.to(mse.device)
mse = mse * weights
reconstruction_loss = mse.mean()
loss_dict = {
'total': reconstruction_loss.item(),
'reconstruction': reconstruction_loss.item(),
}
return reconstruction_loss, loss_dict
# ==================== 工具函数 ====================
def count_parameters(model: nn.Module) -> int:
"""统计模型参数量"""
return sum(p.numel() for p in model.parameters() if p.requires_grad)
def create_model(config: dict = None) -> LSTMAutoencoder:
"""
创建模型
默认使用小型 LSTM 配置，适合异动检测
"""
default_config = {
'n_features': 6,
'hidden_dim': 32, # 小！
'latent_dim': 4, # 非常小！关键
'num_layers': 1,
'dropout': 0.2,
'bidirectional': True,
}
if config:
# 兼容旧的 Transformer 配置键名
if 'd_model' in config:
config['hidden_dim'] = config.pop('d_model') // 2
if 'num_encoder_layers' in config:
config['num_layers'] = config.pop('num_encoder_layers')
if 'num_decoder_layers' in config:
config.pop('num_decoder_layers')
if 'nhead' in config:
config.pop('nhead')
if 'dim_feedforward' in config:
config.pop('dim_feedforward')
if 'max_seq_len' in config:
config.pop('max_seq_len')
if 'use_instance_norm' in config:
config.pop('use_instance_norm')
default_config.update(config)
model = LSTMAutoencoder(**default_config)
param_count = count_parameters(model)
print(f"模型参数量: {param_count:,}")
if param_count > 100000:
print(f"⚠️ 警告: 参数量较大（{param_count:,}），可能过拟合")
else:
print(f"✓ 参数量适中（LSTM Autoencoder）")
return model
if __name__ == "__main__":
# 测试模型
print("测试 LSTM Autoencoder...")
# 创建模型
model = create_model()
# 测试输入
batch_size = 32
seq_len = 30
n_features = 6
x = torch.randn(batch_size, seq_len, n_features)
# 前向传播
output, latent = model(x)
print(f"输入形状: {x.shape}")
print(f"输出形状: {output.shape}")
print(f"隐向量形状: {latent.shape}")
# 计算重构误差
error = model.compute_reconstruction_error(x)
print(f"重构误差形状: {error.shape}")
print(f"平均重构误差: {error.mean().item():.4f}")
# 测试异动检测
is_anomaly, scores = model.detect_anomaly(x, threshold=0.5)
print(f"异动检测结果形状: {is_anomaly.shape if is_anomaly is not None else 'None'}")
print(f"异动分数形状: {scores.shape}")
# 测试损失函数
criterion = AnomalyDetectionLoss()
loss, loss_dict = criterion(output, x, latent)
print(f"损失: {loss.item():.4f}")
print("\n测试通过！")

537
ml/prepare_data.py
View File

@@ -1,537 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
数据准备脚本 - 为 Transformer Autoencoder 准备训练数据
从 ClickHouse 提取历史分钟数据，计算以下特征：
1. alpha - 超额收益（概念涨幅 - 大盘涨幅）
2. alpha_delta - Alpha 变化率（5分钟）
3. amt_ratio - 成交额相对均值（当前/过去20分钟均值）
4. amt_delta - 成交额变化率
5. rank_pct - Alpha 排名百分位
6. limit_up_ratio - 涨停股占比
输出：按交易日存储的特征文件（parquet格式）
"""
import os
import sys
import numpy as np
import pandas as pd
from datetime import datetime, timedelta, date
from sqlalchemy import create_engine, text
from elasticsearch import Elasticsearch
from clickhouse_driver import Client
import hashlib
import json
import logging
from typing import Dict, List, Set, Tuple
from concurrent.futures import ProcessPoolExecutor, as_completed
from multiprocessing import Manager
import multiprocessing
import warnings
warnings.filterwarnings('ignore')
# ==================== 配置 ====================
MYSQL_ENGINE = create_engine(
"mysql+pymysql://root:Zzl5588161!@192.168.1.5:3306/stock",
echo=False
)
ES_CLIENT = Elasticsearch(['http://127.0.0.1:9200'])
ES_INDEX = 'concept_library_v3'
CLICKHOUSE_CONFIG = {
'host': '127.0.0.1',
'port': 9000,
'user': 'default',
'password': 'Zzl33818!',
'database': 'stock'
}
# 输出目录
OUTPUT_DIR = os.path.join(os.path.dirname(__file__), 'data')
os.makedirs(OUTPUT_DIR, exist_ok=True)
# 特征计算参数
FEATURE_CONFIG = {
'alpha_delta_window': 5, # Alpha变化窗口（分钟）
'amt_ma_window': 20, # 成交额均值窗口（分钟）
'limit_up_threshold': 9.8, # 涨停阈值（%）
'limit_down_threshold': -9.8, # 跌停阈值（%）
}
REFERENCE_INDEX = '000001.SH'
# ==================== 日志 ====================
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)
# ==================== 工具函数 ====================
def get_ch_client():
return Client(**CLICKHOUSE_CONFIG)
def generate_id(name: str) -> str:
return hashlib.md5(name.encode('utf-8')).hexdigest()[:16]
def code_to_ch_format(code: str) -> str:
if not code or len(code) != 6 or not code.isdigit():
return None
if code.startswith('6'):
return f"{code}.SH"
elif code.startswith('0') or code.startswith('3'):
return f"{code}.SZ"
else:
return f"{code}.BJ"
# ==================== 获取概念列表 ====================
def get_all_concepts() -> List[dict]:
"""从ES获取所有叶子概念"""
concepts = []
query = {
"query": {"match_all": {}},
"size": 100,
"_source": ["concept_id", "concept", "stocks"]
}
resp = ES_CLIENT.search(index=ES_INDEX, body=query, scroll='2m')
scroll_id = resp['_scroll_id']
hits = resp['hits']['hits']
while len(hits) > 0:
for hit in hits:
source = hit['_source']
stocks = []
if 'stocks' in source and isinstance(source['stocks'], list):
for stock in source['stocks']:
if isinstance(stock, dict) and 'code' in stock and stock['code']:
stocks.append(stock['code'])
if stocks:
concepts.append({
'concept_id': source.get('concept_id'),
'concept_name': source.get('concept'),
'stocks': stocks
})
resp = ES_CLIENT.scroll(scroll_id=scroll_id, scroll='2m')
scroll_id = resp['_scroll_id']
hits = resp['hits']['hits']
ES_CLIENT.clear_scroll(scroll_id=scroll_id)
print(f"获取到 {len(concepts)} 个概念")
return concepts
# ==================== 获取交易日列表 ====================
def get_trading_days(start_date: str, end_date: str) -> List[str]:
"""获取交易日列表"""
client = get_ch_client()
query = f"""
SELECT DISTINCT toDate(timestamp) as trade_date
FROM stock_minute
WHERE toDate(timestamp) >= '{start_date}'
AND toDate(timestamp) <= '{end_date}'
ORDER BY trade_date
"""
result = client.execute(query)
days = [row[0].strftime('%Y-%m-%d') for row in result]
print(f"找到 {len(days)} 个交易日: {days[0]} ~ {days[-1]}")
return days
# ==================== 获取单日数据 ====================
def get_daily_stock_data(trade_date: str, stock_codes: List[str]) -> pd.DataFrame:
"""获取单日所有股票的分钟数据"""
client = get_ch_client()
# 转换代码格式
ch_codes = []
code_map = {}
for code in stock_codes:
ch_code = code_to_ch_format(code)
if ch_code:
ch_codes.append(ch_code)
code_map[ch_code] = code
if not ch_codes:
return pd.DataFrame()
ch_codes_str = "','".join(ch_codes)
query = f"""
SELECT
code,
timestamp,
close,
volume,
amt
FROM stock_minute
WHERE toDate(timestamp) = '{trade_date}'
AND code IN ('{ch_codes_str}')
ORDER BY code, timestamp
"""
result = client.execute(query)
if not result:
return pd.DataFrame()
df = pd.DataFrame(result, columns=['ch_code', 'timestamp', 'close', 'volume', 'amt'])
df['code'] = df['ch_code'].map(code_map)
df = df.dropna(subset=['code'])
return df[['code', 'timestamp', 'close', 'volume', 'amt']]
def get_daily_index_data(trade_date: str, index_code: str = REFERENCE_INDEX) -> pd.DataFrame:
"""获取单日指数分钟数据"""
client = get_ch_client()
query = f"""
SELECT
timestamp,
close,
volume,
amt
FROM index_minute
WHERE toDate(timestamp) = '{trade_date}'
AND code = '{index_code}'
ORDER BY timestamp
"""
result = client.execute(query)
if not result:
return pd.DataFrame()
df = pd.DataFrame(result, columns=['timestamp', 'close', 'volume', 'amt'])
return df
def get_prev_close(stock_codes: List[str], trade_date: str) -> Dict[str, float]:
"""获取昨收价（上一交易日的收盘价 F007N）"""
valid_codes = [c for c in stock_codes if c and len(c) == 6 and c.isdigit()]
if not valid_codes:
return {}
codes_str = "','".join(valid_codes)
# 注意：F007N 是"最近成交价"即当日收盘价，F002N 是"昨日收盘价"
# 我们需要查上一交易日的 F007N（那天的收盘价）作为今天的昨收
query = f"""
SELECT SECCODE, F007N
FROM ea_trade
WHERE SECCODE IN ('{codes_str}')
AND TRADEDATE = (
SELECT MAX(TRADEDATE) FROM ea_trade WHERE TRADEDATE < '{trade_date}'
)
AND F007N IS NOT NULL AND F007N > 0
"""
try:
with MYSQL_ENGINE.connect() as conn:
result = conn.execute(text(query))
return {row[0]: float(row[1]) for row in result if row[1]}
except Exception as e:
print(f"获取昨收价失败: {e}")
return {}
def get_index_prev_close(trade_date: str, index_code: str = REFERENCE_INDEX) -> float:
"""获取指数昨收价"""
code_no_suffix = index_code.split('.')[0]
try:
with MYSQL_ENGINE.connect() as conn:
result = conn.execute(text("""
SELECT F006N FROM ea_exchangetrade
WHERE INDEXCODE = :code AND TRADEDATE < :today
ORDER BY TRADEDATE DESC LIMIT 1
"""), {'code': code_no_suffix, 'today': trade_date}).fetchone()
if result and result[0]:
return float(result[0])
except Exception as e:
print(f"获取指数昨收失败: {e}")
return None
# ==================== 计算特征 ====================
def compute_daily_features(
trade_date: str,
concepts: List[dict],
all_stocks: List[str]
) -> pd.DataFrame:
"""
计算单日所有概念的特征
返回 DataFrame:
- index: (timestamp, concept_id)
- columns: alpha, alpha_delta, amt_ratio, amt_delta, rank_pct, limit_up_ratio
"""
# 1. 获取数据
stock_df = get_daily_stock_data(trade_date, all_stocks)
if stock_df.empty:
return pd.DataFrame()
index_df = get_daily_index_data(trade_date)
if index_df.empty:
return pd.DataFrame()
# 2. 获取昨收价
prev_close = get_prev_close(all_stocks, trade_date)
index_prev_close = get_index_prev_close(trade_date)
if not prev_close or not index_prev_close:
return pd.DataFrame()
# 3. 计算股票涨跌幅和成交额
stock_df['prev_close'] = stock_df['code'].map(prev_close)
stock_df = stock_df.dropna(subset=['prev_close'])
stock_df['change_pct'] = (stock_df['close'] - stock_df['prev_close']) / stock_df['prev_close'] * 100
# 4. 计算指数涨跌幅
index_df['change_pct'] = (index_df['close'] - index_prev_close) / index_prev_close * 100
index_change_map = dict(zip(index_df['timestamp'], index_df['change_pct']))
# 5. 获取所有时间点
timestamps = sorted(stock_df['timestamp'].unique())
# 6. 按时间点计算概念特征
results = []
# 概念到股票的映射
concept_stocks = {c['concept_id']: set(c['stocks']) for c in concepts}
concept_names = {c['concept_id']: c['concept_name'] for c in concepts}
# 历史数据缓存（用于计算变化率）
concept_history = {cid: {'alpha': [], 'amt': []} for cid in concept_stocks}
for ts in timestamps:
ts_stock_data = stock_df[stock_df['timestamp'] == ts]
index_change = index_change_map.get(ts, 0)
# 股票涨跌幅和成交额字典
stock_change = dict(zip(ts_stock_data['code'], ts_stock_data['change_pct']))
stock_amt = dict(zip(ts_stock_data['code'], ts_stock_data['amt']))
concept_features = []
for concept_id, stocks in concept_stocks.items():
# 该概念的股票数据
concept_changes = [stock_change[s] for s in stocks if s in stock_change]
concept_amts = [stock_amt.get(s, 0) for s in stocks if s in stock_change]
if not concept_changes:
continue
# 基础统计
avg_change = np.mean(concept_changes)
total_amt = sum(concept_amts)
# Alpha = 概念涨幅 - 指数涨幅
alpha = avg_change - index_change
# 涨停/跌停股占比
limit_up_count = sum(1 for c in concept_changes if c >= FEATURE_CONFIG['limit_up_threshold'])
limit_down_count = sum(1 for c in concept_changes if c <= FEATURE_CONFIG['limit_down_threshold'])
limit_up_ratio = limit_up_count / len(concept_changes)
limit_down_ratio = limit_down_count / len(concept_changes)
# 更新历史
history = concept_history[concept_id]
history['alpha'].append(alpha)
history['amt'].append(total_amt)
# 计算变化率
alpha_delta = 0
if len(history['alpha']) > FEATURE_CONFIG['alpha_delta_window']:
alpha_delta = alpha - history['alpha'][-FEATURE_CONFIG['alpha_delta_window']-1]
# 成交额相对均值
amt_ratio = 1.0
amt_delta = 0
if len(history['amt']) > FEATURE_CONFIG['amt_ma_window']:
amt_ma = np.mean(history['amt'][-FEATURE_CONFIG['amt_ma_window']-1:-1])
if amt_ma > 0:
amt_ratio = total_amt / amt_ma
amt_delta = total_amt - history['amt'][-2] if len(history['amt']) > 1 else 0
concept_features.append({
'concept_id': concept_id,
'alpha': alpha,
'alpha_delta': alpha_delta,
'amt_ratio': amt_ratio,
'amt_delta': amt_delta,
'limit_up_ratio': limit_up_ratio,
'limit_down_ratio': limit_down_ratio,
'total_amt': total_amt,
'stock_count': len(concept_changes),
})
if not concept_features:
continue
# 计算排名百分位
concept_df = pd.DataFrame(concept_features)
concept_df['rank_pct'] = concept_df['alpha'].rank(pct=True)
# 添加时间戳
concept_df['timestamp'] = ts
results.append(concept_df)
if not results:
return pd.DataFrame()
# 合并所有时间点
final_df = pd.concat(results, ignore_index=True)
# 标准化成交额变化率
if 'amt_delta' in final_df.columns:
amt_delta_std = final_df['amt_delta'].std()
if amt_delta_std > 0:
final_df['amt_delta'] = final_df['amt_delta'] / amt_delta_std
return final_df
# ==================== 主流程 ====================
def process_single_day(args) -> Tuple[str, bool]:
"""
处理单个交易日（多进程版本）
Args:
args: (trade_date, concepts, all_stocks) 元组
Returns:
(trade_date, success) 元组
"""
trade_date, concepts, all_stocks = args
output_file = os.path.join(OUTPUT_DIR, f'features_{trade_date}.parquet')
# 检查是否已处理
if os.path.exists(output_file):
print(f"[{trade_date}] 已存在，跳过")
return (trade_date, True)
print(f"[{trade_date}] 开始处理...")
try:
df = compute_daily_features(trade_date, concepts, all_stocks)
if df.empty:
print(f"[{trade_date}] 无数据")
return (trade_date, False)
# 保存
df.to_parquet(output_file, index=False)
print(f"[{trade_date}] 保存完成")
return (trade_date, True)
except Exception as e:
print(f"[{trade_date}] 处理失败: {e}")
import traceback
traceback.print_exc()
return (trade_date, False)
def main():
import argparse
from tqdm import tqdm
parser = argparse.ArgumentParser(description='准备训练数据')
parser.add_argument('--start', type=str, default='2022-01-01', help='开始日期')
parser.add_argument('--end', type=str, default=None, help='结束日期（默认今天）')
parser.add_argument('--workers', type=int, default=18, help='并行进程数（默认18）')
parser.add_argument('--force', action='store_true', help='强制重新处理已存在的文件')
args = parser.parse_args()
end_date = args.end or datetime.now().strftime('%Y-%m-%d')
print("=" * 60)
print("数据准备 - Transformer Autoencoder 训练数据")
print("=" * 60)
print(f"日期范围: {args.start} ~ {end_date}")
print(f"并行进程数: {args.workers}")
# 1. 获取概念列表
concepts = get_all_concepts()
# 收集所有股票
all_stocks = list(set(s for c in concepts for s in c['stocks']))
print(f"股票总数: {len(all_stocks)}")
# 2. 获取交易日列表
trading_days = get_trading_days(args.start, end_date)
if not trading_days:
print("无交易日数据")
return
# 如果强制模式，删除已有文件
if args.force:
for trade_date in trading_days:
output_file = os.path.join(OUTPUT_DIR, f'features_{trade_date}.parquet')
if os.path.exists(output_file):
os.remove(output_file)
print(f"删除已有文件: {output_file}")
# 3. 准备任务参数
tasks = [(trade_date, concepts, all_stocks) for trade_date in trading_days]
print(f"\n开始处理 {len(trading_days)} 个交易日（{args.workers} 进程并行）...")
# 4. 多进程处理
success_count = 0
failed_dates = []
with ProcessPoolExecutor(max_workers=args.workers) as executor:
# 提交所有任务
futures = {executor.submit(process_single_day, task): task[0] for task in tasks}
# 使用 tqdm 显示进度
with tqdm(total=len(futures), desc="处理进度", unit="天") as pbar:
for future in as_completed(futures):
trade_date = futures[future]
try:
result_date, success = future.result()
if success:
success_count += 1
else:
failed_dates.append(result_date)
except Exception as e:
print(f"\n[{trade_date}] 进程异常: {e}")
failed_dates.append(trade_date)
pbar.update(1)
print("\n" + "=" * 60)
print(f"处理完成: {success_count}/{len(trading_days)} 个交易日")
if failed_dates:
print(f"失败日期: {failed_dates[:10]}{'...' if len(failed_dates) > 10 else ''}")
print(f"数据保存在: {OUTPUT_DIR}")
print("=" * 60)
if __name__ == "__main__":
main()

715
ml/prepare_data_v2.py
View File

@@ -1,715 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
数据准备 V2 - 基于时间片对齐的特征计算（修复版）
核心改进：
1. 时间片对齐：9:35 和历史的 9:35 比，而不是和前30分钟比
2. Z-Score 特征：相对于同时间片历史分布的偏离程度
3. 滚动窗口基线：每个日期使用它之前 N 天的数据作为基线（不是固定的最后 N 天！）
4. 基于 Z-Score 的动量：消除一天内波动率异构性
修复：
- 滚动窗口基线：避免未来数据泄露
- Z-Score 动量：消除早盘/尾盘波动率差异
- 进程级数据库单例：避免连接池爆炸
"""
import os
import sys
import numpy as np
import pandas as pd
from datetime import datetime, timedelta
from sqlalchemy import create_engine, text
from elasticsearch import Elasticsearch
from clickhouse_driver import Client
from concurrent.futures import ProcessPoolExecutor, as_completed
from typing import Dict, List, Tuple, Optional
from tqdm import tqdm
from collections import defaultdict
import warnings
import pickle
warnings.filterwarnings('ignore')
# ==================== 配置 ====================
MYSQL_URL = "mysql+pymysql://root:Zzl5588161!@192.168.1.5:3306/stock"
ES_HOST = 'http://127.0.0.1:9200'
ES_INDEX = 'concept_library_v3'
CLICKHOUSE_CONFIG = {
'host': '127.0.0.1',
'port': 9000,
'user': 'default',
'password': 'Zzl33818!',
'database': 'stock'
}
REFERENCE_INDEX = '000001.SH'
# 输出目录
OUTPUT_DIR = os.path.join(os.path.dirname(__file__), 'data_v2')
BASELINE_DIR = os.path.join(OUTPUT_DIR, 'baselines')
RAW_CACHE_DIR = os.path.join(OUTPUT_DIR, 'raw_cache')
os.makedirs(OUTPUT_DIR, exist_ok=True)
os.makedirs(BASELINE_DIR, exist_ok=True)
os.makedirs(RAW_CACHE_DIR, exist_ok=True)
# 特征配置
CONFIG = {
'baseline_days': 20, # 滚动窗口大小
'min_baseline_samples': 10, # 最少需要10个样本才算有效基线
'limit_up_threshold': 9.8,
'limit_down_threshold': -9.8,
'zscore_clip': 5.0,
}
# 特征列表
FEATURES_V2 = [
'alpha', 'alpha_zscore', 'amt_zscore', 'rank_zscore',
'momentum_3m', 'momentum_5m', 'limit_up_ratio',
]
# ==================== 进程级单例（避免连接池爆炸）====================
# 进程级全局变量
_process_mysql_engine = None
_process_es_client = None
_process_ch_client = None
def init_process_connections():
"""进程初始化时调用，创建连接（单例）"""
global _process_mysql_engine, _process_es_client, _process_ch_client
_process_mysql_engine = create_engine(MYSQL_URL, echo=False, pool_pre_ping=True, pool_size=5)
_process_es_client = Elasticsearch([ES_HOST])
_process_ch_client = Client(**CLICKHOUSE_CONFIG)
def get_mysql_engine():
"""获取进程级 MySQL Engine（单例）"""
global _process_mysql_engine
if _process_mysql_engine is None:
_process_mysql_engine = create_engine(MYSQL_URL, echo=False, pool_pre_ping=True, pool_size=5)
return _process_mysql_engine
def get_es_client():
"""获取进程级 ES 客户端（单例）"""
global _process_es_client
if _process_es_client is None:
_process_es_client = Elasticsearch([ES_HOST])
return _process_es_client
def get_ch_client():
"""获取进程级 ClickHouse 客户端（单例）"""
global _process_ch_client
if _process_ch_client is None:
_process_ch_client = Client(**CLICKHOUSE_CONFIG)
return _process_ch_client
# ==================== 工具函数 ====================
def code_to_ch_format(code: str) -> str:
if not code or len(code) != 6 or not code.isdigit():
return None
if code.startswith('6'):
return f"{code}.SH"
elif code.startswith('0') or code.startswith('3'):
return f"{code}.SZ"
else:
return f"{code}.BJ"
def time_to_slot(ts) -> str:
"""将时间戳转换为时间片（HH:MM格式）"""
if isinstance(ts, str):
return ts
return ts.strftime('%H:%M')
# ==================== 获取概念列表 ====================
def get_all_concepts() -> List[dict]:
"""从ES获取所有叶子概念"""
es_client = get_es_client()
concepts = []
query = {
"query": {"match_all": {}},
"size": 100,
"_source": ["concept_id", "concept", "stocks"]
}
resp = es_client.search(index=ES_INDEX, body=query, scroll='2m')
scroll_id = resp['_scroll_id']
hits = resp['hits']['hits']
while len(hits) > 0:
for hit in hits:
source = hit['_source']
stocks = []
if 'stocks' in source and isinstance(source['stocks'], list):
for stock in source['stocks']:
if isinstance(stock, dict) and 'code' in stock and stock['code']:
stocks.append(stock['code'])
if stocks:
concepts.append({
'concept_id': source.get('concept_id'),
'concept_name': source.get('concept'),
'stocks': stocks
})
resp = es_client.scroll(scroll_id=scroll_id, scroll='2m')
scroll_id = resp['_scroll_id']
hits = resp['hits']['hits']
es_client.clear_scroll(scroll_id=scroll_id)
print(f"获取到 {len(concepts)} 个概念")
return concepts
# ==================== 获取交易日列表 ====================
def get_trading_days(start_date: str, end_date: str) -> List[str]:
"""获取交易日列表"""
client = get_ch_client()
query = f"""
SELECT DISTINCT toDate(timestamp) as trade_date
FROM stock_minute
WHERE toDate(timestamp) >= '{start_date}'
AND toDate(timestamp) <= '{end_date}'
ORDER BY trade_date
"""
result = client.execute(query)
days = [row[0].strftime('%Y-%m-%d') for row in result]
if days:
print(f"找到 {len(days)} 个交易日: {days[0]} ~ {days[-1]}")
return days
# ==================== 获取昨收价 ====================
def get_prev_close(stock_codes: List[str], trade_date: str) -> Dict[str, float]:
"""获取昨收价（上一交易日的收盘价 F007N）"""
valid_codes = [c for c in stock_codes if c and len(c) == 6 and c.isdigit()]
if not valid_codes:
return {}
codes_str = "','".join(valid_codes)
query = f"""
SELECT SECCODE, F007N
FROM ea_trade
WHERE SECCODE IN ('{codes_str}')
AND TRADEDATE = (
SELECT MAX(TRADEDATE) FROM ea_trade WHERE TRADEDATE < '{trade_date}'
)
AND F007N IS NOT NULL AND F007N > 0
"""
try:
engine = get_mysql_engine()
with engine.connect() as conn:
result = conn.execute(text(query))
return {row[0]: float(row[1]) for row in result if row[1]}
except Exception as e:
print(f"获取昨收价失败: {e}")
return {}
def get_index_prev_close(trade_date: str, index_code: str = REFERENCE_INDEX) -> float:
"""获取指数昨收价"""
code_no_suffix = index_code.split('.')[0]
try:
engine = get_mysql_engine()
with engine.connect() as conn:
result = conn.execute(text("""
SELECT F006N FROM ea_exchangetrade
WHERE INDEXCODE = :code AND TRADEDATE < :today
ORDER BY TRADEDATE DESC LIMIT 1
"""), {'code': code_no_suffix, 'today': trade_date}).fetchone()
if result and result[0]:
return float(result[0])
except Exception as e:
print(f"获取指数昨收失败: {e}")
return None
# ==================== 获取分钟数据 ====================
def get_daily_stock_data(trade_date: str, stock_codes: List[str]) -> pd.DataFrame:
"""获取单日所有股票的分钟数据"""
client = get_ch_client()
ch_codes = []
code_map = {}
for code in stock_codes:
ch_code = code_to_ch_format(code)
if ch_code:
ch_codes.append(ch_code)
code_map[ch_code] = code
if not ch_codes:
return pd.DataFrame()
ch_codes_str = "','".join(ch_codes)
query = f"""
SELECT code, timestamp, close, volume, amt
FROM stock_minute
WHERE toDate(timestamp) = '{trade_date}'
AND code IN ('{ch_codes_str}')
ORDER BY code, timestamp
"""
result = client.execute(query)
if not result:
return pd.DataFrame()
df = pd.DataFrame(result, columns=['ch_code', 'timestamp', 'close', 'volume', 'amt'])
df['code'] = df['ch_code'].map(code_map)
df = df.dropna(subset=['code'])
return df[['code', 'timestamp', 'close', 'volume', 'amt']]
def get_daily_index_data(trade_date: str, index_code: str = REFERENCE_INDEX) -> pd.DataFrame:
"""获取单日指数分钟数据"""
client = get_ch_client()
query = f"""
SELECT timestamp, close, volume, amt
FROM index_minute
WHERE toDate(timestamp) = '{trade_date}'
AND code = '{index_code}'
ORDER BY timestamp
"""
result = client.execute(query)
if not result:
return pd.DataFrame()
df = pd.DataFrame(result, columns=['timestamp', 'close', 'volume', 'amt'])
return df
# ==================== 计算原始概念特征（单日）====================
def compute_raw_concept_features(
trade_date: str,
concepts: List[dict],
all_stocks: List[str]
) -> pd.DataFrame:
"""计算单日概念的原始特征（alpha, amt, rank_pct, limit_up_ratio）"""
# 检查缓存
cache_file = os.path.join(RAW_CACHE_DIR, f'raw_{trade_date}.parquet')
if os.path.exists(cache_file):
return pd.read_parquet(cache_file)
# 获取数据
stock_df = get_daily_stock_data(trade_date, all_stocks)
if stock_df.empty:
return pd.DataFrame()
index_df = get_daily_index_data(trade_date)
if index_df.empty:
return pd.DataFrame()
# 获取昨收价
prev_close = get_prev_close(all_stocks, trade_date)
index_prev_close = get_index_prev_close(trade_date)
if not prev_close or not index_prev_close:
return pd.DataFrame()
# 计算涨跌幅
stock_df['prev_close'] = stock_df['code'].map(prev_close)
stock_df = stock_df.dropna(subset=['prev_close'])
stock_df['change_pct'] = (stock_df['close'] - stock_df['prev_close']) / stock_df['prev_close'] * 100
index_df['change_pct'] = (index_df['close'] - index_prev_close) / index_prev_close * 100
index_change_map = dict(zip(index_df['timestamp'], index_df['change_pct']))
# 获取所有时间点
timestamps = sorted(stock_df['timestamp'].unique())
# 概念到股票的映射
concept_stocks = {c['concept_id']: set(c['stocks']) for c in concepts}
results = []
for ts in timestamps:
ts_stock_data = stock_df[stock_df['timestamp'] == ts]
index_change = index_change_map.get(ts, 0)
stock_change = dict(zip(ts_stock_data['code'], ts_stock_data['change_pct']))
stock_amt = dict(zip(ts_stock_data['code'], ts_stock_data['amt']))
concept_features = []
for concept_id, stocks in concept_stocks.items():
concept_changes = [stock_change[s] for s in stocks if s in stock_change]
concept_amts = [stock_amt.get(s, 0) for s in stocks if s in stock_change]
if not concept_changes:
continue
avg_change = np.mean(concept_changes)
total_amt = sum(concept_amts)
alpha = avg_change - index_change
limit_up_count = sum(1 for c in concept_changes if c >= CONFIG['limit_up_threshold'])
limit_up_ratio = limit_up_count / len(concept_changes)
concept_features.append({
'concept_id': concept_id,
'alpha': alpha,
'total_amt': total_amt,
'limit_up_ratio': limit_up_ratio,
'stock_count': len(concept_changes),
})
if not concept_features:
continue
concept_df = pd.DataFrame(concept_features)
concept_df['rank_pct'] = concept_df['alpha'].rank(pct=True)
concept_df['timestamp'] = ts
concept_df['time_slot'] = time_to_slot(ts)
concept_df['trade_date'] = trade_date
results.append(concept_df)
if not results:
return pd.DataFrame()
result_df = pd.concat(results, ignore_index=True)
# 保存缓存
result_df.to_parquet(cache_file, index=False)
return result_df
# ==================== 滚动窗口基线计算 ====================
def compute_rolling_baseline(
historical_data: pd.DataFrame,
concept_id: str
) -> Dict[str, Dict]:
"""
计算单个概念的滚动基线
返回: {time_slot: {alpha_mean, alpha_std, amt_mean, amt_std, rank_mean, rank_std, sample_count}}
"""
if historical_data.empty:
return {}
concept_data = historical_data[historical_data['concept_id'] == concept_id]
if concept_data.empty:
return {}
baseline_dict = {}
for time_slot, group in concept_data.groupby('time_slot'):
if len(group) < CONFIG['min_baseline_samples']:
continue
alpha_std = group['alpha'].std()
amt_std = group['total_amt'].std()
rank_std = group['rank_pct'].std()
baseline_dict[time_slot] = {
'alpha_mean': group['alpha'].mean(),
'alpha_std': max(alpha_std if pd.notna(alpha_std) else 1.0, 0.1),
'amt_mean': group['total_amt'].mean(),
'amt_std': max(amt_std if pd.notna(amt_std) else group['total_amt'].mean() * 0.5, 1.0),
'rank_mean': group['rank_pct'].mean(),
'rank_std': max(rank_std if pd.notna(rank_std) else 0.2, 0.05),
'sample_count': len(group),
}
return baseline_dict
# ==================== 计算单日 Z-Score 特征（带滚动基线）====================
def compute_zscore_features_rolling(
trade_date: str,
concepts: List[dict],
all_stocks: List[str],
historical_raw_data: pd.DataFrame # 该日期之前 N 天的原始数据
) -> pd.DataFrame:
"""
计算单日的 Z-Score 特征（使用滚动窗口基线）
关键改进：
1. 基线只使用 trade_date 之前的数据（无未来泄露）
2. 动量基于 Z-Score 计算（消除波动率异构性）
"""
# 计算当日原始特征
raw_df = compute_raw_concept_features(trade_date, concepts, all_stocks)
if raw_df.empty:
return pd.DataFrame()
zscore_records = []
for concept_id, group in raw_df.groupby('concept_id'):
# 计算该概念的滚动基线（只用历史数据）
baseline_dict = compute_rolling_baseline(historical_raw_data, concept_id)
if not baseline_dict:
continue
# 按时间排序
group = group.sort_values('timestamp').reset_index(drop=True)
# Z-Score 历史（用于计算基于 Z-Score 的动量）
zscore_history = []
for idx, row in group.iterrows():
time_slot = row['time_slot']
if time_slot not in baseline_dict:
continue
bl = baseline_dict[time_slot]
# 计算 Z-Score
alpha_zscore = (row['alpha'] - bl['alpha_mean']) / bl['alpha_std']
amt_zscore = (row['total_amt'] - bl['amt_mean']) / bl['amt_std']
rank_zscore = (row['rank_pct'] - bl['rank_mean']) / bl['rank_std']
# 截断极端值
clip = CONFIG['zscore_clip']
alpha_zscore = np.clip(alpha_zscore, -clip, clip)
amt_zscore = np.clip(amt_zscore, -clip, clip)
rank_zscore = np.clip(rank_zscore, -clip, clip)
# 记录 Z-Score 历史
zscore_history.append(alpha_zscore)
# 基于 Z-Score 计算动量（消除波动率异构性）
momentum_3m = 0.0
momentum_5m = 0.0
if len(zscore_history) >= 3:
recent_3 = zscore_history[-3:]
older_3 = zscore_history[-6:-3] if len(zscore_history) >= 6 else [zscore_history[0]]
momentum_3m = np.mean(recent_3) - np.mean(older_3)
if len(zscore_history) >= 5:
recent_5 = zscore_history[-5:]
older_5 = zscore_history[-10:-5] if len(zscore_history) >= 10 else [zscore_history[0]]
momentum_5m = np.mean(recent_5) - np.mean(older_5)
zscore_records.append({
'concept_id': concept_id,
'timestamp': row['timestamp'],
'time_slot': time_slot,
'trade_date': trade_date,
# 原始特征
'alpha': row['alpha'],
'total_amt': row['total_amt'],
'limit_up_ratio': row['limit_up_ratio'],
'stock_count': row['stock_count'],
'rank_pct': row['rank_pct'],
# Z-Score 特征
'alpha_zscore': alpha_zscore,
'amt_zscore': amt_zscore,
'rank_zscore': rank_zscore,
# 基于 Z-Score 的动量
'momentum_3m': momentum_3m,
'momentum_5m': momentum_5m,
})
if not zscore_records:
return pd.DataFrame()
return pd.DataFrame(zscore_records)
# ==================== 多进程处理 ====================
def process_single_day_v2(args) -> Tuple[str, bool]:
"""处理单个交易日（多进程版本）"""
trade_date, day_index, concepts, all_stocks, all_trading_days = args
output_file = os.path.join(OUTPUT_DIR, f'features_v2_{trade_date}.parquet')
if os.path.exists(output_file):
return (trade_date, True)
try:
# 计算滚动窗口范围（该日期之前的 N 天）
baseline_days = CONFIG['baseline_days']
# 找出 trade_date 之前的交易日
start_idx = max(0, day_index - baseline_days)
end_idx = day_index # 不包含当天
if end_idx <= start_idx:
# 没有足够的历史数据
return (trade_date, False)
historical_days = all_trading_days[start_idx:end_idx]
# 加载历史原始数据
historical_dfs = []
for hist_date in historical_days:
cache_file = os.path.join(RAW_CACHE_DIR, f'raw_{hist_date}.parquet')
if os.path.exists(cache_file):
historical_dfs.append(pd.read_parquet(cache_file))
else:
# 需要计算
hist_df = compute_raw_concept_features(hist_date, concepts, all_stocks)
if not hist_df.empty:
historical_dfs.append(hist_df)
if not historical_dfs:
return (trade_date, False)
historical_raw_data = pd.concat(historical_dfs, ignore_index=True)
# 计算当日 Z-Score 特征（使用滚动基线）
df = compute_zscore_features_rolling(trade_date, concepts, all_stocks, historical_raw_data)
if df.empty:
return (trade_date, False)
df.to_parquet(output_file, index=False)
return (trade_date, True)
except Exception as e:
print(f"[{trade_date}] 处理失败: {e}")
import traceback
traceback.print_exc()
return (trade_date, False)
# ==================== 主流程 ====================
def main():
import argparse
parser = argparse.ArgumentParser(description='准备训练数据 V2（滚动窗口基线 + Z-Score 动量）')
parser.add_argument('--start', type=str, default='2022-01-01', help='开始日期')
parser.add_argument('--end', type=str, default=None, help='结束日期（默认今天）')
parser.add_argument('--workers', type=int, default=18, help='并行进程数')
parser.add_argument('--baseline-days', type=int, default=20, help='滚动基线窗口大小')
parser.add_argument('--force', action='store_true', help='强制重新计算（忽略缓存）')
args = parser.parse_args()
end_date = args.end or datetime.now().strftime('%Y-%m-%d')
CONFIG['baseline_days'] = args.baseline_days
print("=" * 60)
print("数据准备 V2 - 滚动窗口基线 + Z-Score 动量")
print("=" * 60)
print(f"日期范围: {args.start} ~ {end_date}")
print(f"并行进程数: {args.workers}")
print(f"滚动基线窗口: {args.baseline_days} 天")
# 初始化主进程连接
init_process_connections()
# 1. 获取概念列表
concepts = get_all_concepts()
all_stocks = list(set(s for c in concepts for s in c['stocks']))
print(f"股票总数: {len(all_stocks)}")
# 2. 获取交易日列表
trading_days = get_trading_days(args.start, end_date)
if not trading_days:
print("无交易日数据")
return
# 3. 第一阶段：预计算所有原始特征（用于缓存）
print(f"\n{'='*60}")
print("第一阶段：预计算原始特征（用于滚动基线）")
print(f"{'='*60}")
# 如果强制重新计算，删除缓存
if args.force:
import shutil
if os.path.exists(RAW_CACHE_DIR):
shutil.rmtree(RAW_CACHE_DIR)
os.makedirs(RAW_CACHE_DIR, exist_ok=True)
if os.path.exists(OUTPUT_DIR):
for f in os.listdir(OUTPUT_DIR):
if f.startswith('features_v2_'):
os.remove(os.path.join(OUTPUT_DIR, f))
# 单线程预计算原始特征（因为需要顺序缓存）
print(f"预计算 {len(trading_days)} 天的原始特征...")
for trade_date in tqdm(trading_days, desc="预计算原始特征"):
cache_file = os.path.join(RAW_CACHE_DIR, f'raw_{trade_date}.parquet')
if not os.path.exists(cache_file):
compute_raw_concept_features(trade_date, concepts, all_stocks)
# 4. 第二阶段：计算 Z-Score 特征（多进程）
print(f"\n{'='*60}")
print("第二阶段：计算 Z-Score 特征（滚动基线）")
print(f"{'='*60}")
# 从第 baseline_days 天开始（前面的没有足够历史）
start_idx = args.baseline_days
processable_days = trading_days[start_idx:]
if not processable_days:
print(f"错误：需要至少 {args.baseline_days + 1} 天的数据")
return
print(f"可处理日期: {processable_days[0]} ~ {processable_days[-1]} ({len(processable_days)} 天)")
print(f"跳过前 {start_idx} 天（基线预热期）")
# 构建任务
tasks = []
for i, trade_date in enumerate(trading_days):
if i >= start_idx:
tasks.append((trade_date, i, concepts, all_stocks, trading_days))
print(f"开始处理 {len(tasks)} 个交易日（{args.workers} 进程并行）...")
success_count = 0
failed_dates = []
# 使用进程池初始化器
with ProcessPoolExecutor(max_workers=args.workers, initializer=init_process_connections) as executor:
futures = {executor.submit(process_single_day_v2, task): task[0] for task in tasks}
with tqdm(total=len(futures), desc="处理进度", unit="天") as pbar:
for future in as_completed(futures):
trade_date = futures[future]
try:
result_date, success = future.result()
if success:
success_count += 1
else:
failed_dates.append(result_date)
except Exception as e:
print(f"\n[{trade_date}] 进程异常: {e}")
failed_dates.append(trade_date)
pbar.update(1)
print("\n" + "=" * 60)
print(f"处理完成: {success_count}/{len(tasks)} 个交易日")
if failed_dates:
print(f"失败日期: {failed_dates[:10]}{'...' if len(failed_dates) > 10 else ''}")
print(f"数据保存在: {OUTPUT_DIR}")
print("=" * 60)
if __name__ == "__main__":
main()

1520
ml/realtime_detector.py
View File

File diff suppressed because it is too large Load Diff

729
ml/realtime_detector_v2.py
View File

@@ -1,729 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
V2 实时异动检测器
使用方法：
# 作为模块导入
from ml.realtime_detector_v2 import RealtimeDetectorV2
detector = RealtimeDetectorV2()
alerts = detector.detect_realtime() # 检测当前时刻
# 或命令行测试
python ml/realtime_detector_v2.py --date 2025-12-09
"""
import os
import sys
import json
import pickle
from datetime import datetime, timedelta
from typing import Dict, List, Optional, Tuple
from collections import defaultdict, deque
import numpy as np
import pandas as pd
import torch
from sqlalchemy import create_engine, text
from elasticsearch import Elasticsearch
from clickhouse_driver import Client
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from ml.model import TransformerAutoencoder
# ==================== 配置 ====================
MYSQL_URL = "mysql+pymysql://root:Zzl5588161!@192.168.1.5:3306/stock"
ES_HOST = 'http://127.0.0.1:9200'
ES_INDEX = 'concept_library_v3'
CLICKHOUSE_CONFIG = {
'host': '127.0.0.1',
'port': 9000,
'user': 'default',
'password': 'Zzl33818!',
'database': 'stock'
}
REFERENCE_INDEX = '000001.SH'
BASELINE_FILE = 'ml/data_v2/baselines/realtime_baseline.pkl'
MODEL_DIR = 'ml/checkpoints_v2'
# 检测配置
CONFIG = {
'seq_len': 10, # LSTM 序列长度
'confirm_window': 5, # 持续确认窗口
'confirm_ratio': 0.6, # 确认比例
'rule_weight': 0.5,
'ml_weight': 0.5,
'rule_trigger': 60,
'ml_trigger': 70,
'fusion_trigger': 50,
'cooldown_minutes': 10,
'max_alerts_per_minute': 15,
'zscore_clip': 5.0,
'limit_up_threshold': 9.8,
}
FEATURES = ['alpha_zscore', 'amt_zscore', 'rank_zscore', 'momentum_3m', 'momentum_5m', 'limit_up_ratio']
# ==================== 数据库连接 ====================
_mysql_engine = None
_es_client = None
_ch_client = None
def get_mysql_engine():
global _mysql_engine
if _mysql_engine is None:
_mysql_engine = create_engine(MYSQL_URL, echo=False, pool_pre_ping=True)
return _mysql_engine
def get_es_client():
global _es_client
if _es_client is None:
_es_client = Elasticsearch([ES_HOST])
return _es_client
def get_ch_client():
global _ch_client
if _ch_client is None:
_ch_client = Client(**CLICKHOUSE_CONFIG)
return _ch_client
def code_to_ch_format(code: str) -> str:
if not code or len(code) != 6 or not code.isdigit():
return None
if code.startswith('6'):
return f"{code}.SH"
elif code.startswith('0') or code.startswith('3'):
return f"{code}.SZ"
return f"{code}.BJ"
def time_to_slot(ts) -> str:
if isinstance(ts, str):
return ts
return ts.strftime('%H:%M')
# ==================== 规则评分 ====================
def score_rules_zscore(features: Dict) -> Tuple[float, List[str]]:
"""基于 Z-Score 的规则评分"""
score = 0.0
triggered = []
alpha_z = abs(features.get('alpha_zscore', 0))
amt_z = features.get('amt_zscore', 0)
rank_z = abs(features.get('rank_zscore', 0))
mom_3m = features.get('momentum_3m', 0)
mom_5m = features.get('momentum_5m', 0)
limit_up = features.get('limit_up_ratio', 0)
# Alpha Z-Score
if alpha_z >= 4.0:
score += 25
triggered.append('alpha_extreme')
elif alpha_z >= 3.0:
score += 18
triggered.append('alpha_strong')
elif alpha_z >= 2.0:
score += 10
triggered.append('alpha_moderate')
# 成交额 Z-Score
if amt_z >= 4.0:
score += 20
triggered.append('amt_extreme')
elif amt_z >= 3.0:
score += 12
triggered.append('amt_strong')
elif amt_z >= 2.0:
score += 6
triggered.append('amt_moderate')
# 排名 Z-Score
if rank_z >= 3.0:
score += 15
triggered.append('rank_extreme')
elif rank_z >= 2.0:
score += 8
triggered.append('rank_strong')
# 动量（基于 Z-Score 的）
if mom_3m >= 1.0:
score += 12
triggered.append('momentum_3m_strong')
elif mom_3m >= 0.5:
score += 6
triggered.append('momentum_3m_moderate')
if mom_5m >= 1.5:
score += 10
triggered.append('momentum_5m_strong')
# 涨停比例
if limit_up >= 0.3:
score += 20
triggered.append('limit_up_extreme')
elif limit_up >= 0.15:
score += 12
triggered.append('limit_up_strong')
elif limit_up >= 0.08:
score += 5
triggered.append('limit_up_moderate')
# 组合规则
if alpha_z >= 2.0 and amt_z >= 2.0:
score += 15
triggered.append('combo_alpha_amt')
if alpha_z >= 2.0 and limit_up >= 0.1:
score += 12
triggered.append('combo_alpha_limitup')
return min(score, 100), triggered
# ==================== 实时检测器 ====================
class RealtimeDetectorV2:
"""V2 实时异动检测器"""
def __init__(self, model_dir: str = MODEL_DIR, baseline_file: str = BASELINE_FILE):
print("初始化 V2 实时检测器...")
# 加载概念
self.concepts = self._load_concepts()
self.concept_stocks = {c['concept_id']: set(c['stocks']) for c in self.concepts}
self.all_stocks = list(set(s for c in self.concepts for s in c['stocks']))
# 加载基线
self.baselines = self._load_baselines(baseline_file)
# 加载模型
self.model, self.thresholds, self.device = self._load_model(model_dir)
# 状态管理
self.zscore_history = defaultdict(lambda: deque(maxlen=CONFIG['seq_len']))
self.anomaly_candidates = defaultdict(lambda: deque(maxlen=CONFIG['confirm_window']))
self.cooldown = {}
print(f"初始化完成: {len(self.concepts)} 概念, {len(self.baselines)} 基线")
def _load_concepts(self) -> List[dict]:
"""从 ES 加载概念"""
es = get_es_client()
concepts = []
query = {"query": {"match_all": {}}, "size": 100, "_source": ["concept_id", "concept", "stocks"]}
resp = es.search(index=ES_INDEX, body=query, scroll='2m')
scroll_id = resp['_scroll_id']
hits = resp['hits']['hits']
while hits:
for hit in hits:
src = hit['_source']
stocks = [s['code'] for s in src.get('stocks', []) if isinstance(s, dict) and s.get('code')]
if stocks:
concepts.append({
'concept_id': src.get('concept_id'),
'concept_name': src.get('concept'),
'stocks': stocks
})
resp = es.scroll(scroll_id=scroll_id, scroll='2m')
scroll_id = resp['_scroll_id']
hits = resp['hits']['hits']
es.clear_scroll(scroll_id=scroll_id)
return concepts
def _load_baselines(self, baseline_file: str) -> Dict:
"""加载基线"""
if not os.path.exists(baseline_file):
print(f"警告: 基线文件不存在: {baseline_file}")
print("请先运行: python ml/update_baseline.py")
return {}
with open(baseline_file, 'rb') as f:
data = pickle.load(f)
print(f"基线日期范围: {data.get('date_range', 'unknown')}")
print(f"更新时间: {data.get('update_time', 'unknown')}")
return data.get('baselines', {})
def _load_model(self, model_dir: str):
"""加载模型"""
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
config_path = os.path.join(model_dir, 'config.json')
model_path = os.path.join(model_dir, 'best_model.pt')
threshold_path = os.path.join(model_dir, 'thresholds.json')
if not os.path.exists(model_path):
print(f"警告: 模型不存在: {model_path}")
return None, {}, device
with open(config_path) as f:
config = json.load(f)
model = TransformerAutoencoder(**config['model'])
checkpoint = torch.load(model_path, map_location=device)
model.load_state_dict(checkpoint['model_state_dict'])
model.to(device)
model.eval()
thresholds = {}
if os.path.exists(threshold_path):
with open(threshold_path) as f:
thresholds = json.load(f)
print(f"模型已加载: {model_path}")
return model, thresholds, device
def _get_realtime_data(self, trade_date: str) -> pd.DataFrame:
"""获取实时数据并计算原始特征"""
ch = get_ch_client()
# 获取股票数据
ch_codes = [code_to_ch_format(c) for c in self.all_stocks if code_to_ch_format(c)]
ch_codes_str = "','".join(ch_codes)
stock_query = f"""
SELECT code, timestamp, close, amt
FROM stock_minute
WHERE toDate(timestamp) = '{trade_date}'
AND code IN ('{ch_codes_str}')
ORDER BY timestamp
"""
stock_result = ch.execute(stock_query)
if not stock_result:
return pd.DataFrame()
stock_df = pd.DataFrame(stock_result, columns=['ch_code', 'timestamp', 'close', 'amt'])
# 映射回原始代码
ch_to_code = {code_to_ch_format(c): c for c in self.all_stocks if code_to_ch_format(c)}
stock_df['code'] = stock_df['ch_code'].map(ch_to_code)
stock_df = stock_df.dropna(subset=['code'])
# 获取指数数据
index_query = f"""
SELECT timestamp, close
FROM index_minute
WHERE toDate(timestamp) = '{trade_date}'
AND code = '{REFERENCE_INDEX}'
ORDER BY timestamp
"""
index_result = ch.execute(index_query)
if not index_result:
return pd.DataFrame()
index_df = pd.DataFrame(index_result, columns=['timestamp', 'close'])
# 获取昨收价
engine = get_mysql_engine()
codes_str = "','".join([c for c in self.all_stocks if c and len(c) == 6])
with engine.connect() as conn:
prev_result = conn.execute(text(f"""
SELECT SECCODE, F007N FROM ea_trade
WHERE SECCODE IN ('{codes_str}')
AND TRADEDATE = (SELECT MAX(TRADEDATE) FROM ea_trade WHERE TRADEDATE < '{trade_date}')
AND F007N > 0
"""))
prev_close = {row[0]: float(row[1]) for row in prev_result if row[1]}
idx_result = conn.execute(text("""
SELECT F006N FROM ea_exchangetrade
WHERE INDEXCODE = '000001' AND TRADEDATE < :today
ORDER BY TRADEDATE DESC LIMIT 1
"""), {'today': trade_date}).fetchone()
index_prev_close = float(idx_result[0]) if idx_result else None
if not prev_close or not index_prev_close:
return pd.DataFrame()
# 计算涨跌幅
stock_df['prev_close'] = stock_df['code'].map(prev_close)
stock_df = stock_df.dropna(subset=['prev_close'])
stock_df['change_pct'] = (stock_df['close'] - stock_df['prev_close']) / stock_df['prev_close'] * 100
index_df['change_pct'] = (index_df['close'] - index_prev_close) / index_prev_close * 100
index_map = dict(zip(index_df['timestamp'], index_df['change_pct']))
# 按时间聚合概念特征
results = []
for ts in sorted(stock_df['timestamp'].unique()):
ts_data = stock_df[stock_df['timestamp'] == ts]
idx_chg = index_map.get(ts, 0)
stock_chg = dict(zip(ts_data['code'], ts_data['change_pct']))
stock_amt = dict(zip(ts_data['code'], ts_data['amt']))
for cid, stocks in self.concept_stocks.items():
changes = [stock_chg[s] for s in stocks if s in stock_chg]
amts = [stock_amt.get(s, 0) for s in stocks if s in stock_chg]
if not changes:
continue
alpha = np.mean(changes) - idx_chg
total_amt = sum(amts)
limit_up_ratio = sum(1 for c in changes if c >= CONFIG['limit_up_threshold']) / len(changes)
results.append({
'concept_id': cid,
'timestamp': ts,
'time_slot': time_to_slot(ts),
'alpha': alpha,
'total_amt': total_amt,
'limit_up_ratio': limit_up_ratio,
'stock_count': len(changes),
})
if not results:
return pd.DataFrame()
df = pd.DataFrame(results)
# 计算排名
for ts in df['timestamp'].unique():
mask = df['timestamp'] == ts
df.loc[mask, 'rank_pct'] = df.loc[mask, 'alpha'].rank(pct=True)
return df
def _compute_zscore(self, concept_id: str, time_slot: str, alpha: float, total_amt: float, rank_pct: float) -> Optional[Dict]:
"""计算 Z-Score"""
if concept_id not in self.baselines:
return None
baseline = self.baselines[concept_id]
if time_slot not in baseline:
return None
bl = baseline[time_slot]
alpha_z = np.clip((alpha - bl['alpha_mean']) / bl['alpha_std'], -5, 5)
amt_z = np.clip((total_amt - bl['amt_mean']) / bl['amt_std'], -5, 5)
rank_z = np.clip((rank_pct - bl['rank_mean']) / bl['rank_std'], -5, 5)
# 动量（基于 Z-Score 历史）
history = list(self.zscore_history[concept_id])
mom_3m = 0.0
mom_5m = 0.0
if len(history) >= 3:
recent = [h['alpha_zscore'] for h in history[-3:]]
older = [h['alpha_zscore'] for h in history[-6:-3]] if len(history) >= 6 else [history[0]['alpha_zscore']]
mom_3m = np.mean(recent) - np.mean(older)
if len(history) >= 5:
recent = [h['alpha_zscore'] for h in history[-5:]]
older = [h['alpha_zscore'] for h in history[-10:-5]] if len(history) >= 10 else [history[0]['alpha_zscore']]
mom_5m = np.mean(recent) - np.mean(older)
return {
'alpha_zscore': float(alpha_z),
'amt_zscore': float(amt_z),
'rank_zscore': float(rank_z),
'momentum_3m': float(mom_3m),
'momentum_5m': float(mom_5m),
}
@torch.no_grad()
def _ml_score(self, sequences: np.ndarray) -> np.ndarray:
"""批量 ML 评分"""
if self.model is None or len(sequences) == 0:
return np.zeros(len(sequences))
x = torch.FloatTensor(sequences).to(self.device)
errors = self.model.compute_reconstruction_error(x, reduction='none')
last_errors = errors[:, -1].cpu().numpy()
# 转换为 0-100 分数
if self.thresholds:
p50 = self.thresholds.get('median', 0.001)
p99 = self.thresholds.get('p99', 0.05)
scores = 50 + (last_errors - p50) / (p99 - p50 + 1e-6) * 49
else:
scores = last_errors * 1000
return np.clip(scores, 0, 100)
def detect(self, trade_date: str = None) -> List[Dict]:
"""检测指定日期的异动"""
trade_date = trade_date or datetime.now().strftime('%Y-%m-%d')
print(f"\n检测 {trade_date} 的异动...")
# 重置状态
self.zscore_history.clear()
self.anomaly_candidates.clear()
self.cooldown.clear()
# 获取数据
raw_df = self._get_realtime_data(trade_date)
if raw_df.empty:
print("无数据")
return []
timestamps = sorted(raw_df['timestamp'].unique())
print(f"时间点数: {len(timestamps)}")
all_alerts = []
for ts in timestamps:
ts_data = raw_df[raw_df['timestamp'] == ts]
time_slot = time_to_slot(ts)
candidates = []
# 计算每个概念的 Z-Score
for _, row in ts_data.iterrows():
cid = row['concept_id']
zscore = self._compute_zscore(
cid, time_slot,
row['alpha'], row['total_amt'], row['rank_pct']
)
if zscore is None:
continue
# 完整特征
features = {
**zscore,
'alpha': row['alpha'],
'limit_up_ratio': row['limit_up_ratio'],
'total_amt': row['total_amt'],
}
# 更新历史
self.zscore_history[cid].append(zscore)
# 规则评分
rule_score, triggered = score_rules_zscore(features)
candidates.append((cid, features, rule_score, triggered))
if not candidates:
continue
# 批量 ML 评分
sequences = []
valid_candidates = []
for cid, features, rule_score, triggered in candidates:
history = list(self.zscore_history[cid])
if len(history) >= CONFIG['seq_len']:
seq = np.array([[h['alpha_zscore'], h['amt_zscore'], h['rank_zscore'],
h['momentum_3m'], h['momentum_5m'], features['limit_up_ratio']]
for h in history])
sequences.append(seq)
valid_candidates.append((cid, features, rule_score, triggered))
if not sequences:
continue
ml_scores = self._ml_score(np.array(sequences))
# 融合 + 确认
for i, (cid, features, rule_score, triggered) in enumerate(valid_candidates):
ml_score = ml_scores[i]
final_score = CONFIG['rule_weight'] * rule_score + CONFIG['ml_weight'] * ml_score
# 判断触发
is_triggered = (
rule_score >= CONFIG['rule_trigger'] or
ml_score >= CONFIG['ml_trigger'] or
final_score >= CONFIG['fusion_trigger']
)
self.anomaly_candidates[cid].append((ts, final_score))
if not is_triggered:
continue
# 冷却期
if cid in self.cooldown:
if (ts - self.cooldown[cid]).total_seconds() < CONFIG['cooldown_minutes'] * 60:
continue
# 持续确认
recent = list(self.anomaly_candidates[cid])
if len(recent) < CONFIG['confirm_window']:
continue
exceed = sum(1 for _, s in recent if s >= CONFIG['fusion_trigger'])
ratio = exceed / len(recent)
if ratio < CONFIG['confirm_ratio']:
continue
# 确认异动！
self.cooldown[cid] = ts
alpha = features['alpha']
alert_type = 'surge_up' if alpha >= 1.5 else 'surge_down' if alpha <= -1.5 else 'surge'
concept_name = next((c['concept_name'] for c in self.concepts if c['concept_id'] == cid), cid)
all_alerts.append({
'concept_id': cid,
'concept_name': concept_name,
'alert_time': ts,
'trade_date': trade_date,
'alert_type': alert_type,
'final_score': float(final_score),
'rule_score': float(rule_score),
'ml_score': float(ml_score),
'confirm_ratio': float(ratio),
'alpha': float(alpha),
'alpha_zscore': float(features['alpha_zscore']),
'amt_zscore': float(features['amt_zscore']),
'rank_zscore': float(features['rank_zscore']),
'momentum_3m': float(features['momentum_3m']),
'momentum_5m': float(features['momentum_5m']),
'limit_up_ratio': float(features['limit_up_ratio']),
'triggered_rules': triggered,
'trigger_reason': f"融合({final_score:.0f})+确认({ratio:.0%})",
})
print(f"检测到 {len(all_alerts)} 个异动")
return all_alerts
# ==================== 数据库存储 ====================
def create_v2_table():
"""创建 V2 异动表（如果不存在）"""
engine = get_mysql_engine()
with engine.begin() as conn:
conn.execute(text("""
CREATE TABLE IF NOT EXISTS concept_anomaly_v2 (
id INT AUTO_INCREMENT PRIMARY KEY,
concept_id VARCHAR(50) NOT NULL,
alert_time DATETIME NOT NULL,
trade_date DATE NOT NULL,
alert_type VARCHAR(20) NOT NULL,
final_score FLOAT,
rule_score FLOAT,
ml_score FLOAT,
trigger_reason VARCHAR(200),
confirm_ratio FLOAT,
alpha FLOAT,
alpha_zscore FLOAT,
amt_zscore FLOAT,
rank_zscore FLOAT,
momentum_3m FLOAT,
momentum_5m FLOAT,
limit_up_ratio FLOAT,
triggered_rules TEXT,
created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
UNIQUE KEY uk_concept_time (concept_id, alert_time),
INDEX idx_trade_date (trade_date),
INDEX idx_alert_type (alert_type)
) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4
"""))
print("concept_anomaly_v2 表已就绪")
def save_alerts_to_db(alerts: List[Dict]) -> int:
"""保存异动到数据库"""
if not alerts:
return 0
engine = get_mysql_engine()
saved = 0
with engine.begin() as conn:
for alert in alerts:
try:
insert_sql = text("""
INSERT IGNORE INTO concept_anomaly_v2
(concept_id, alert_time, trade_date, alert_type,
final_score, rule_score, ml_score, trigger_reason, confirm_ratio,
alpha, alpha_zscore, amt_zscore, rank_zscore,
momentum_3m, momentum_5m, limit_up_ratio, triggered_rules)
VALUES
(:concept_id, :alert_time, :trade_date, :alert_type,
:final_score, :rule_score, :ml_score, :trigger_reason, :confirm_ratio,
:alpha, :alpha_zscore, :amt_zscore, :rank_zscore,
:momentum_3m, :momentum_5m, :limit_up_ratio, :triggered_rules)
""")
result = conn.execute(insert_sql, {
'concept_id': alert['concept_id'],
'alert_time': alert['alert_time'],
'trade_date': alert['trade_date'],
'alert_type': alert['alert_type'],
'final_score': alert['final_score'],
'rule_score': alert['rule_score'],
'ml_score': alert['ml_score'],
'trigger_reason': alert['trigger_reason'],
'confirm_ratio': alert['confirm_ratio'],
'alpha': alert['alpha'],
'alpha_zscore': alert['alpha_zscore'],
'amt_zscore': alert['amt_zscore'],
'rank_zscore': alert['rank_zscore'],
'momentum_3m': alert['momentum_3m'],
'momentum_5m': alert['momentum_5m'],
'limit_up_ratio': alert['limit_up_ratio'],
'triggered_rules': json.dumps(alert.get('triggered_rules', []), ensure_ascii=False),
})
if result.rowcount > 0:
saved += 1
except Exception as e:
print(f"保存失败: {alert['concept_id']} - {e}")
return saved
def main():
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--date', type=str, default=None)
parser.add_argument('--no-save', action='store_true', help='不保存到数据库，只打印')
args = parser.parse_args()
# 确保表存在
if not args.no_save:
create_v2_table()
detector = RealtimeDetectorV2()
alerts = detector.detect(args.date)
print(f"\n{'='*60}")
print(f"检测结果 ({len(alerts)} 个异动)")
print('='*60)
for a in alerts[:20]:
print(f"[{a['alert_time'].strftime('%H:%M') if hasattr(a['alert_time'], 'strftime') else a['alert_time']}] "
f"{a['concept_name']} | {a['alert_type']} | "
f"分数={a['final_score']:.0f} 确认={a['confirm_ratio']:.0%} "
f"α={a['alpha']:.2f}% αZ={a['alpha_zscore']:.1f}")
if len(alerts) > 20:
print(f"... 共 {len(alerts)} 个")
# 保存到数据库
if not args.no_save and alerts:
saved = save_alerts_to_db(alerts)
print(f"\n✅ 已保存 {saved}/{len(alerts)} 条到 concept_anomaly_v2 表")
elif args.no_save:
print(f"\n⚠️ --no-save 模式，未保存到数据库")
if __name__ == "__main__":
main()

25
ml/requirements.txt
View File

@@ -1,25 +0,0 @@
# 概念异动检测 ML 模块依赖
# 安装: pip install -r ml/requirements.txt
# PyTorch (根据 CUDA 版本选择)
# 5090 显卡需要 CUDA 12.x
# pip install torch --index-url https://download.pytorch.org/whl/cu124
torch>=2.0.0
# 数据处理
numpy>=1.24.0
pandas>=2.0.0
pyarrow>=14.0.0
# 数据库
clickhouse-driver>=0.2.6
elasticsearch>=7.0.0,<8.0.0
sqlalchemy>=2.0.0
pymysql>=1.1.0
# 训练工具
tqdm>=4.65.0
# 可选: 可视化
# matplotlib>=3.7.0
# tensorboard>=2.14.0

99
ml/run_training.sh
View File

@@ -1,99 +0,0 @@
#!/bin/bash
# 概念异动检测模型训练脚本 (Linux)
#
# 使用方法:
# chmod +x run_training.sh
# ./run_training.sh
#
# 或指定参数:
# ./run_training.sh --start 2022-01-01 --epochs 100
set -e
echo "============================================================"
echo "概念异动检测模型训练流程"
echo "============================================================"
echo ""
# 获取脚本所在目录
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
cd "$SCRIPT_DIR/.."
echo "[1/4] 检查环境..."
python3 --version || { echo "Python3 未找到！"; exit 1; }
# 检查 GPU
if python3 -c "import torch; print(f'CUDA: {torch.cuda.is_available()}')" 2>/dev/null; then
echo "PyTorch GPU 检测完成"
else
echo "警告: PyTorch 未安装或无法检测 GPU"
fi
echo ""
echo "[2/4] 检查依赖..."
pip3 install -q torch pandas numpy pyarrow tqdm clickhouse-driver elasticsearch sqlalchemy pymysql
echo ""
echo "[3/4] 准备训练数据..."
echo "从 ClickHouse 提取历史数据，这可能需要较长时间..."
echo ""
# 解析参数
START_DATE="2022-01-01"
END_DATE=""
EPOCHS=100
BATCH_SIZE=256
TRAIN_END="2025-06-30"
VAL_END="2025-09-30"
while [[ $# -gt 0 ]]; do
case $1 in
--start)
START_DATE="$2"
shift 2
;;
--end)
END_DATE="$2"
shift 2
;;
--epochs)
EPOCHS="$2"
shift 2
;;
--batch_size)
BATCH_SIZE="$2"
shift 2
;;
--train_end)
TRAIN_END="$2"
shift 2
;;
--val_end)
VAL_END="$2"
shift 2
;;
*)
shift
;;
esac
done
# 数据准备
if [ -n "$END_DATE" ]; then
python3 ml/prepare_data.py --start "$START_DATE" --end "$END_DATE"
else
python3 ml/prepare_data.py --start "$START_DATE"
fi
echo ""
echo "[4/4] 训练模型..."
echo "使用 GPU 加速训练..."
echo ""
python3 ml/train.py --epochs "$EPOCHS" --batch_size "$BATCH_SIZE" --train_end "$TRAIN_END" --val_end "$VAL_END"
echo ""
echo "============================================================"
echo "训练完成！"
echo "模型保存在: ml/checkpoints/"
echo "============================================================"

808
ml/train.py
View File

@@ -1,808 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Transformer Autoencoder 训练脚本 (修复版)
修复问题：
1. 按概念分组构建序列，避免跨概念切片
2. 按时间（日期）切分数据集，避免数据泄露
3. 使用 RobustScaler + Clipping 处理非平稳性
4. 使用验证集计算阈值
训练流程：
1. 加载预处理好的特征数据（parquet 文件）
2. 按概念分组，在每个概念内部构建序列
3. 按日期划分训练/验证/测试集
4. 训练 Autoencoder（最小化重构误差）
5. 保存模型和阈值
使用方法：
python train.py --data_dir ml/data --epochs 100 --batch_size 256
"""
import os
import sys
import argparse
import json
from datetime import datetime
from pathlib import Path
from typing import List, Tuple, Dict
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from torch.optim import AdamW
from torch.optim.lr_scheduler import CosineAnnealingWarmRestarts
from tqdm import tqdm
from model import TransformerAutoencoder, AnomalyDetectionLoss, count_parameters
# 性能优化：启用 cuDNN benchmark（对固定输入尺寸自动选择最快算法）
torch.backends.cudnn.benchmark = True
# 启用 TF32（RTX 30/40 系列特有，提速约 3 倍）
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
# 可视化（可选）
try:
import matplotlib
matplotlib.use('Agg') # 无头模式，不需要显示器
import matplotlib.pyplot as plt
HAS_MATPLOTLIB = True
except ImportError:
HAS_MATPLOTLIB = False
# ==================== 配置 ====================
TRAIN_CONFIG = {
# 数据配置
'seq_len': 30, # 输入序列长度（30分钟）
'stride': 5, # 滑动窗口步长
# 时间切分（按日期）
'train_end_date': '2024-06-30', # 训练集截止日期
'val_end_date': '2024-09-30', # 验证集截止日期（之后为测试集）
# 特征配置
'features': [
'alpha', # 超额收益
'alpha_delta', # Alpha 变化率
'amt_ratio', # 成交额比率
'amt_delta', # 成交额变化率
'rank_pct', # Alpha 排名百分位
'limit_up_ratio', # 涨停比例
],
# 训练配置（针对 4x RTX 4090 优化）
'batch_size': 4096, # 256 -> 4096（大幅增加，充分利用显存）
'epochs': 100,
'learning_rate': 3e-4, # 1e-4 -> 3e-4（大 batch 需要更大学习率）
'weight_decay': 1e-5,
'gradient_clip': 1.0,
# 早停配置
'patience': 10,
'min_delta': 1e-6,
# 模型配置（LSTM Autoencoder，简洁有效）
'model': {
'n_features': 6,
'hidden_dim': 32, # LSTM 隐藏维度（小）
'latent_dim': 4, # 瓶颈维度（非常小！关键）
'num_layers': 1, # LSTM 层数
'dropout': 0.2,
'bidirectional': True, # 双向编码器
},
# 标准化配置
'use_instance_norm': True, # 模型内部使用 Instance Norm（推荐）
'clip_value': 10.0, # 简单截断极端值
# 阈值配置
'threshold_percentiles': [90, 95, 99],
}
# ==================== 数据加载（修复版）====================
def load_data_by_date(data_dir: str, features: List[str]) -> Dict[str, pd.DataFrame]:
"""
按日期加载数据，返回 {date: DataFrame} 字典
每个 DataFrame 包含该日所有概念的所有时间点数据
"""
data_path = Path(data_dir)
parquet_files = sorted(data_path.glob("features_*.parquet"))
if not parquet_files:
raise FileNotFoundError(f"未找到 parquet 文件: {data_dir}")
print(f"找到 {len(parquet_files)} 个数据文件")
date_data = {}
for pf in tqdm(parquet_files, desc="加载数据"):
# 提取日期
date = pf.stem.replace('features_', '')
df = pd.read_parquet(pf)
# 检查必要列
required_cols = features + ['concept_id', 'timestamp']
missing_cols = [c for c in required_cols if c not in df.columns]
if missing_cols:
print(f"警告: {date} 缺少列: {missing_cols}, 跳过")
continue
date_data[date] = df
print(f"成功加载 {len(date_data)} 天的数据")
return date_data
def split_data_by_date(
date_data: Dict[str, pd.DataFrame],
train_end: str,
val_end: str
) -> Tuple[Dict[str, pd.DataFrame], Dict[str, pd.DataFrame], Dict[str, pd.DataFrame]]:
"""
按日期严格划分数据集
- 训练集: <= train_end
- 验证集: train_end < date <= val_end
- 测试集: > val_end
"""
train_data = {}
val_data = {}
test_data = {}
for date, df in date_data.items():
if date <= train_end:
train_data[date] = df
elif date <= val_end:
val_data[date] = df
else:
test_data[date] = df
print(f"数据集划分（按日期）:")
print(f" 训练集: {len(train_data)} 天 (<= {train_end})")
print(f" 验证集: {len(val_data)} 天 ({train_end} ~ {val_end})")
print(f" 测试集: {len(test_data)} 天 (> {val_end})")
return train_data, val_data, test_data
def build_sequences_by_concept(
date_data: Dict[str, pd.DataFrame],
features: List[str],
seq_len: int,
stride: int
) -> np.ndarray:
"""
按概念分组构建序列（性能优化版）
使用 groupby 一次性分组，避免重复扫描大数组
1. 将所有日期的数据合并
2. 使用 groupby 按 concept_id 分组
3. 在每个概念内部，按时间排序并滑动窗口
4. 合并所有序列
"""
# 合并所有日期的数据
all_dfs = []
for date, df in sorted(date_data.items()):
df = df.copy()
df['date'] = date
all_dfs.append(df)
if not all_dfs:
return np.array([])
combined = pd.concat(all_dfs, ignore_index=True)
# 预先排序（按概念、日期、时间），这样 groupby 会更快
combined = combined.sort_values(['concept_id', 'date', 'timestamp'])
# 使用 groupby 一次性分组（性能关键！）
all_sequences = []
grouped = combined.groupby('concept_id', sort=False)
n_concepts = len(grouped)
for concept_id, concept_df in tqdm(grouped, desc="构建序列", total=n_concepts, leave=False):
# 已经排序过了，直接提取特征
feature_data = concept_df[features].values
# 处理缺失值
feature_data = np.nan_to_num(feature_data, nan=0.0, posinf=0.0, neginf=0.0)
# 在该概念内部滑动窗口
n_points = len(feature_data)
for start in range(0, n_points - seq_len + 1, stride):
seq = feature_data[start:start + seq_len]
all_sequences.append(seq)
if not all_sequences:
return np.array([])
sequences = np.array(all_sequences)
print(f" 构建序列: {len(sequences):,} 条 (来自 {n_concepts} 个概念)")
return sequences
# ==================== 数据集 ====================
class SequenceDataset(Dataset):
"""序列数据集（已经构建好的序列）"""
def __init__(self, sequences: np.ndarray):
self.sequences = torch.FloatTensor(sequences)
def __len__(self) -> int:
return len(self.sequences)
def __getitem__(self, idx: int) -> torch.Tensor:
return self.sequences[idx]
# ==================== 训练器 ====================
class EarlyStopping:
"""早停机制"""
def __init__(self, patience: int = 10, min_delta: float = 1e-6):
self.patience = patience
self.min_delta = min_delta
self.counter = 0
self.best_loss = float('inf')
self.early_stop = False
def __call__(self, val_loss: float) -> bool:
if val_loss < self.best_loss - self.min_delta:
self.best_loss = val_loss
self.counter = 0
else:
self.counter += 1
if self.counter >= self.patience:
self.early_stop = True
return self.early_stop
class Trainer:
"""模型训练器（支持 AMP 混合精度加速）"""
def __init__(
self,
model: nn.Module,
train_loader: DataLoader,
val_loader: DataLoader,
config: Dict,
device: torch.device,
save_dir: str = 'ml/checkpoints'
):
self.model = model.to(device)
self.train_loader = train_loader
self.val_loader = val_loader
self.config = config
self.device = device
self.save_dir = Path(save_dir)
self.save_dir.mkdir(parents=True, exist_ok=True)
# 优化器
self.optimizer = AdamW(
model.parameters(),
lr=config['learning_rate'],
weight_decay=config['weight_decay']
)
# 学习率调度器
self.scheduler = CosineAnnealingWarmRestarts(
self.optimizer,
T_0=10,
T_mult=2,
eta_min=1e-6
)
# 损失函数（简化版，只用 MSE）
self.criterion = AnomalyDetectionLoss()
# 早停
self.early_stopping = EarlyStopping(
patience=config['patience'],
min_delta=config['min_delta']
)
# AMP 混合精度训练（大幅提速 + 省显存）
self.use_amp = torch.cuda.is_available()
self.scaler = torch.cuda.amp.GradScaler() if self.use_amp else None
if self.use_amp:
print(" ✓ 启用 AMP 混合精度训练")
# 训练历史
self.history = {
'train_loss': [],
'val_loss': [],
'learning_rate': [],
}
self.best_val_loss = float('inf')
def train_epoch(self) -> float:
"""训练一个 epoch（使用 AMP 混合精度）"""
self.model.train()
total_loss = 0.0
n_batches = 0
pbar = tqdm(self.train_loader, desc="Training", leave=False)
for batch in pbar:
batch = batch.to(self.device, non_blocking=True) # 异步传输
self.optimizer.zero_grad(set_to_none=True) # 更快的梯度清零
# AMP 混合精度前向传播
if self.use_amp:
with torch.cuda.amp.autocast():
output, latent = self.model(batch)
loss, loss_dict = self.criterion(output, batch, latent)
# AMP 反向传播
self.scaler.scale(loss).backward()
# 梯度裁剪（需要 unscale）
self.scaler.unscale_(self.optimizer)
torch.nn.utils.clip_grad_norm_(
self.model.parameters(),
self.config['gradient_clip']
)
self.scaler.step(self.optimizer)
self.scaler.update()
else:
# 非 AMP 模式
output, latent = self.model(batch)
loss, loss_dict = self.criterion(output, batch, latent)
loss.backward()
torch.nn.utils.clip_grad_norm_(
self.model.parameters(),
self.config['gradient_clip']
)
self.optimizer.step()
total_loss += loss.item()
n_batches += 1
pbar.set_postfix({'loss': f"{loss.item():.4f}"})
return total_loss / n_batches
@torch.no_grad()
def validate(self) -> float:
"""验证（使用 AMP）"""
self.model.eval()
total_loss = 0.0
n_batches = 0
for batch in self.val_loader:
batch = batch.to(self.device, non_blocking=True)
if self.use_amp:
with torch.cuda.amp.autocast():
output, latent = self.model(batch)
loss, _ = self.criterion(output, batch, latent)
else:
output, latent = self.model(batch)
loss, _ = self.criterion(output, batch, latent)
total_loss += loss.item()
n_batches += 1
return total_loss / n_batches
def save_checkpoint(self, epoch: int, val_loss: float, is_best: bool = False):
"""保存检查点"""
# 处理 DataParallel 包装
model_to_save = self.model.module if hasattr(self.model, 'module') else self.model
checkpoint = {
'epoch': epoch,
'model_state_dict': model_to_save.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'scheduler_state_dict': self.scheduler.state_dict(),
'val_loss': val_loss,
'config': self.config,
}
# 保存最新检查点
torch.save(checkpoint, self.save_dir / 'last_checkpoint.pt')
# 保存最佳模型
if is_best:
torch.save(checkpoint, self.save_dir / 'best_model.pt')
print(f" ✓ 保存最佳模型 (val_loss: {val_loss:.6f})")
def train(self, epochs: int):
"""完整训练流程"""
print(f"\n开始训练 ({epochs} epochs)...")
print(f"设备: {self.device}")
print(f"模型参数量: {count_parameters(self.model):,}")
for epoch in range(1, epochs + 1):
print(f"\nEpoch {epoch}/{epochs}")
# 训练
train_loss = self.train_epoch()
# 验证
val_loss = self.validate()
# 更新学习率
self.scheduler.step()
current_lr = self.optimizer.param_groups[0]['lr']
# 记录历史
self.history['train_loss'].append(train_loss)
self.history['val_loss'].append(val_loss)
self.history['learning_rate'].append(current_lr)
# 打印进度
print(f" Train Loss: {train_loss:.6f}")
print(f" Val Loss: {val_loss:.6f}")
print(f" LR: {current_lr:.2e}")
# 保存检查点
is_best = val_loss < self.best_val_loss
if is_best:
self.best_val_loss = val_loss
self.save_checkpoint(epoch, val_loss, is_best)
# 早停检查
if self.early_stopping(val_loss):
print(f"\n早停触发！验证损失已 {self.early_stopping.patience} 个 epoch 未改善")
break
print(f"\n训练完成！最佳验证损失: {self.best_val_loss:.6f}")
# 保存训练历史
self.save_history()
return self.history
def save_history(self):
"""保存训练历史"""
history_path = self.save_dir / 'training_history.json'
with open(history_path, 'w') as f:
json.dump(self.history, f, indent=2)
print(f"训练历史已保存: {history_path}")
# 绘制训练曲线
self.plot_training_curves()
def plot_training_curves(self):
"""绘制训练曲线"""
if not HAS_MATPLOTLIB:
print("matplotlib 未安装，跳过绘图")
return
fig, axes = plt.subplots(1, 2, figsize=(14, 5))
epochs = range(1, len(self.history['train_loss']) + 1)
# 1. Loss 曲线
ax1 = axes[0]
ax1.plot(epochs, self.history['train_loss'], 'b-', label='Train Loss', linewidth=2)
ax1.plot(epochs, self.history['val_loss'], 'r-', label='Val Loss', linewidth=2)
ax1.set_xlabel('Epoch', fontsize=12)
ax1.set_ylabel('Loss', fontsize=12)
ax1.set_title('Training & Validation Loss', fontsize=14)
ax1.legend(fontsize=11)
ax1.grid(True, alpha=0.3)
# 标记最佳点
best_epoch = np.argmin(self.history['val_loss']) + 1
best_val_loss = min(self.history['val_loss'])
ax1.axvline(x=best_epoch, color='g', linestyle='--', alpha=0.7, label=f'Best Epoch: {best_epoch}')
ax1.scatter([best_epoch], [best_val_loss], color='g', s=100, zorder=5)
ax1.annotate(f'Best: {best_val_loss:.6f}', xy=(best_epoch, best_val_loss),
xytext=(best_epoch + 2, best_val_loss + 0.0005),
fontsize=10, color='green')
# 2. 学习率曲线
ax2 = axes[1]
ax2.plot(epochs, self.history['learning_rate'], 'g-', linewidth=2)
ax2.set_xlabel('Epoch', fontsize=12)
ax2.set_ylabel('Learning Rate', fontsize=12)
ax2.set_title('Learning Rate Schedule', fontsize=14)
ax2.set_yscale('log')
ax2.grid(True, alpha=0.3)
plt.tight_layout()
# 保存图片
plot_path = self.save_dir / 'training_curves.png'
plt.savefig(plot_path, dpi=150, bbox_inches='tight')
plt.close()
print(f"训练曲线已保存: {plot_path}")
# ==================== 阈值计算（使用验证集）====================
@torch.no_grad()
def compute_thresholds(
model: nn.Module,
data_loader: DataLoader,
device: torch.device,
percentiles: List[float] = [90, 95, 99]
) -> Dict[str, float]:
"""
在验证集上计算重构误差的百分位数阈值
注：使用验证集而非测试集，避免数据泄露
"""
model.eval()
all_errors = []
print("计算异动阈值（使用验证集）...")
for batch in tqdm(data_loader, desc="Computing thresholds"):
batch = batch.to(device)
errors = model.compute_reconstruction_error(batch, reduction='none')
# 取每个序列的最后一个时刻误差（预测当前时刻）
seq_errors = errors[:, -1] # (batch,)
all_errors.append(seq_errors.cpu().numpy())
all_errors = np.concatenate(all_errors)
thresholds = {}
for p in percentiles:
threshold = np.percentile(all_errors, p)
thresholds[f'p{p}'] = float(threshold)
print(f" P{p}: {threshold:.6f}")
# 额外统计
thresholds['mean'] = float(np.mean(all_errors))
thresholds['std'] = float(np.std(all_errors))
thresholds['median'] = float(np.median(all_errors))
print(f" Mean: {thresholds['mean']:.6f}")
print(f" Median: {thresholds['median']:.6f}")
print(f" Std: {thresholds['std']:.6f}")
return thresholds
# ==================== 主函数 ====================
def main():
parser = argparse.ArgumentParser(description='训练概念异动检测模型')
parser.add_argument('--data_dir', type=str, default='ml/data',
help='数据目录路径')
parser.add_argument('--epochs', type=int, default=100,
help='训练轮数')
parser.add_argument('--batch_size', type=int, default=4096,
help='批次大小（4x RTX 4090 推荐 4096~8192）')
parser.add_argument('--lr', type=float, default=3e-4,
help='学习率（大 batch 推荐 3e-4）')
parser.add_argument('--device', type=str, default='auto',
help='设备 (auto/cuda/cpu)')
parser.add_argument('--save_dir', type=str, default='ml/checkpoints',
help='模型保存目录')
parser.add_argument('--train_end', type=str, default='2024-06-30',
help='训练集截止日期')
parser.add_argument('--val_end', type=str, default='2024-09-30',
help='验证集截止日期')
args = parser.parse_args()
# 更新配置
config = TRAIN_CONFIG.copy()
config['batch_size'] = args.batch_size
config['epochs'] = args.epochs
config['learning_rate'] = args.lr
config['train_end_date'] = args.train_end
config['val_end_date'] = args.val_end
# 设备选择
if args.device == 'auto':
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
else:
device = torch.device(args.device)
print("=" * 60)
print("概念异动检测模型训练（修复版）")
print("=" * 60)
print(f"配置:")
print(f" 数据目录: {args.data_dir}")
print(f" 设备: {device}")
print(f" 批次大小: {config['batch_size']}")
print(f" 学习率: {config['learning_rate']}")
print(f" 训练轮数: {config['epochs']}")
print(f" 训练集截止: {config['train_end_date']}")
print(f" 验证集截止: {config['val_end_date']}")
print("=" * 60)
# 1. 按日期加载数据
print("\n[1/6] 加载数据...")
date_data = load_data_by_date(args.data_dir, config['features'])
# 2. 按日期划分
print("\n[2/6] 按日期划分数据集...")
train_data, val_data, test_data = split_data_by_date(
date_data,
config['train_end_date'],
config['val_end_date']
)
# 3. 按概念构建序列
print("\n[3/6] 按概念构建序列...")
print("训练集:")
train_sequences = build_sequences_by_concept(
train_data, config['features'], config['seq_len'], config['stride']
)
print("验证集:")
val_sequences = build_sequences_by_concept(
val_data, config['features'], config['seq_len'], config['stride']
)
print("测试集:")
test_sequences = build_sequences_by_concept(
test_data, config['features'], config['seq_len'], config['stride']
)
if len(train_sequences) == 0:
print("错误: 训练集为空！请检查数据和日期范围")
return
# 4. 数据预处理（简单截断极端值，标准化在模型内部通过 Instance Norm 完成）
print("\n[4/6] 数据预处理...")
print(" 注意: 使用 Instance Norm，每个序列在模型内部单独标准化")
print(" 这样可以处理不同概念波动率差异（银行 vs 半导体）")
clip_value = config['clip_value']
print(f" 截断极端值: ±{clip_value}")
# 简单截断极端值（防止异常数据影响训练）
train_sequences = np.clip(train_sequences, -clip_value, clip_value)
if len(val_sequences) > 0:
val_sequences = np.clip(val_sequences, -clip_value, clip_value)
if len(test_sequences) > 0:
test_sequences = np.clip(test_sequences, -clip_value, clip_value)
# 保存配置
save_dir = Path(args.save_dir)
save_dir.mkdir(parents=True, exist_ok=True)
preprocess_params = {
'features': config['features'],
'normalization': 'instance_norm', # 在模型内部完成
'clip_value': clip_value,
'note': '标准化在模型内部通过 InstanceNorm1d 完成，无需外部 Scaler'
}
with open(save_dir / 'normalization_stats.json', 'w') as f:
json.dump(preprocess_params, f, indent=2)
print(f" 预处理参数已保存")
# 5. 创建数据集和加载器
print("\n[5/6] 创建数据加载器...")
train_dataset = SequenceDataset(train_sequences)
val_dataset = SequenceDataset(val_sequences) if len(val_sequences) > 0 else None
test_dataset = SequenceDataset(test_sequences) if len(test_sequences) > 0 else None
print(f" 训练序列: {len(train_dataset):,}")
print(f" 验证序列: {len(val_dataset) if val_dataset else 0:,}")
print(f" 测试序列: {len(test_dataset) if test_dataset else 0:,}")
# 多卡时增加 num_workers（Linux 上可以用更多）
n_gpus = torch.cuda.device_count() if torch.cuda.is_available() else 1
num_workers = min(32, 8 * n_gpus) if sys.platform != 'win32' else 0
print(f" DataLoader workers: {num_workers}")
print(f" Batch size: {config['batch_size']}")
# 大 batch + 多 worker + prefetch 提速
train_loader = DataLoader(
train_dataset,
batch_size=config['batch_size'],
shuffle=True,
num_workers=num_workers,
pin_memory=True,
prefetch_factor=4 if num_workers > 0 else None, # 预取更多 batch
persistent_workers=True if num_workers > 0 else False, # 保持 worker 存活
drop_last=True # 丢弃不完整的最后一批，避免 batch 大小不一致
)
val_loader = DataLoader(
val_dataset,
batch_size=config['batch_size'] * 2, # 验证时可以用更大 batch（无梯度）
shuffle=False,
num_workers=num_workers,
pin_memory=True,
prefetch_factor=4 if num_workers > 0 else None,
persistent_workers=True if num_workers > 0 else False,
) if val_dataset else None
test_loader = DataLoader(
test_dataset,
batch_size=config['batch_size'] * 2,
shuffle=False,
num_workers=num_workers,
pin_memory=True,
prefetch_factor=4 if num_workers > 0 else None,
persistent_workers=True if num_workers > 0 else False,
) if test_dataset else None
# 6. 训练
print("\n[6/6] 训练模型...")
model_config = config['model'].copy()
model = TransformerAutoencoder(**model_config)
# 多卡并行
if torch.cuda.device_count() > 1:
print(f" 使用 {torch.cuda.device_count()} 张 GPU 并行训练")
model = nn.DataParallel(model)
if val_loader is None:
print("警告: 验证集为空，将使用训练集的一部分作为验证")
# 简单处理：用训练集的后 10% 作为验证
split_idx = int(len(train_dataset) * 0.9)
train_subset = torch.utils.data.Subset(train_dataset, range(split_idx))
val_subset = torch.utils.data.Subset(train_dataset, range(split_idx, len(train_dataset)))
train_loader = DataLoader(train_subset, batch_size=config['batch_size'], shuffle=True, num_workers=num_workers, pin_memory=True)
val_loader = DataLoader(val_subset, batch_size=config['batch_size'], shuffle=False, num_workers=num_workers, pin_memory=True)
trainer = Trainer(
model=model,
train_loader=train_loader,
val_loader=val_loader,
config=config,
device=device,
save_dir=args.save_dir
)
history = trainer.train(config['epochs'])
# 7. 计算阈值（使用验证集）
print("\n[额外] 计算异动阈值...")
# 加载最佳模型
best_checkpoint = torch.load(
save_dir / 'best_model.pt',
map_location=device
)
model.load_state_dict(best_checkpoint['model_state_dict'])
model.to(device)
# 使用验证集计算阈值（避免数据泄露）
thresholds = compute_thresholds(
model,
val_loader,
device,
config['threshold_percentiles']
)
# 保存阈值
with open(save_dir / 'thresholds.json', 'w') as f:
json.dump(thresholds, f, indent=2)
print(f"阈值已保存")
# 保存完整配置
with open(save_dir / 'config.json', 'w') as f:
json.dump(config, f, indent=2)
print("\n" + "=" * 60)
print("训练完成！")
print("=" * 60)
print(f"模型保存位置: {args.save_dir}")
print(f" - best_model.pt: 最佳模型权重")
print(f" - thresholds.json: 异动阈值")
print(f" - normalization_stats.json: 标准化参数")
print(f" - config.json: 训练配置")
print("=" * 60)
if __name__ == "__main__":
main()

622
ml/train_v2.py
View File

@@ -1,622 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
训练脚本 V2 - 基于 Z-Score 特征的 LSTM Autoencoder
改进点：
1. 使用 Z-Score 特征（相对于同时间片历史的偏离）
2. 短序列：10分钟（不需要30分钟预热）
3. 开盘即可检测：9:30 直接有特征
模型输入：
- 过去10分钟的 Z-Score 特征序列
- 特征：alpha_zscore, amt_zscore, rank_zscore, momentum_3m, momentum_5m, limit_up_ratio
模型学习：
- 学习 Z-Score 序列的"正常演化模式"
- 异动 = Z-Score 序列的异常演化（重构误差大）
"""
import os
import sys
import argparse
import json
from datetime import datetime
from pathlib import Path
from typing import List, Tuple, Dict
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from torch.optim import AdamW
from torch.optim.lr_scheduler import CosineAnnealingWarmRestarts
from tqdm import tqdm
from model import TransformerAutoencoder, AnomalyDetectionLoss, count_parameters
# 性能优化
torch.backends.cudnn.benchmark = True
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
try:
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
HAS_MATPLOTLIB = True
except ImportError:
HAS_MATPLOTLIB = False
# ==================== 配置 ====================
TRAIN_CONFIG = {
# 数据配置（改进！）
'seq_len': 10, # 10分钟序列（不是30分钟！）
'stride': 2, # 步长2分钟
# 时间切分
'train_end_date': '2024-06-30',
'val_end_date': '2024-09-30',
# V2 特征（Z-Score 为主）
'features': [
'alpha_zscore', # Alpha 的 Z-Score
'amt_zscore', # 成交额的 Z-Score
'rank_zscore', # 排名的 Z-Score
'momentum_3m', # 3分钟动量
'momentum_5m', # 5分钟动量
'limit_up_ratio', # 涨停占比
],
# 训练配置
'batch_size': 4096,
'epochs': 100,
'learning_rate': 3e-4,
'weight_decay': 1e-5,
'gradient_clip': 1.0,
# 早停配置
'patience': 15,
'min_delta': 1e-6,
# 模型配置（小型 LSTM）
'model': {
'n_features': 6,
'hidden_dim': 32,
'latent_dim': 4,
'num_layers': 1,
'dropout': 0.2,
'bidirectional': True,
},
# 标准化配置
'clip_value': 5.0, # Z-Score 已经标准化，clip 5.0 足够
# 阈值配置
'threshold_percentiles': [90, 95, 99],
}
# ==================== 数据加载 ====================
def load_data_by_date(data_dir: str, features: List[str]) -> Dict[str, pd.DataFrame]:
"""按日期加载 V2 数据"""
data_path = Path(data_dir)
parquet_files = sorted(data_path.glob("features_v2_*.parquet"))
if not parquet_files:
raise FileNotFoundError(f"未找到 V2 数据文件: {data_dir}")
print(f"找到 {len(parquet_files)} 个 V2 数据文件")
date_data = {}
for pf in tqdm(parquet_files, desc="加载数据"):
date = pf.stem.replace('features_v2_', '')
df = pd.read_parquet(pf)
required_cols = features + ['concept_id', 'timestamp']
missing_cols = [c for c in required_cols if c not in df.columns]
if missing_cols:
print(f"警告: {date} 缺少列: {missing_cols}, 跳过")
continue
date_data[date] = df
print(f"成功加载 {len(date_data)} 天的数据")
return date_data
def split_data_by_date(
date_data: Dict[str, pd.DataFrame],
train_end: str,
val_end: str
) -> Tuple[Dict[str, pd.DataFrame], Dict[str, pd.DataFrame], Dict[str, pd.DataFrame]]:
"""按日期划分数据集"""
train_data = {}
val_data = {}
test_data = {}
for date, df in date_data.items():
if date <= train_end:
train_data[date] = df
elif date <= val_end:
val_data[date] = df
else:
test_data[date] = df
print(f"数据集划分:")
print(f" 训练集: {len(train_data)} 天 (<= {train_end})")
print(f" 验证集: {len(val_data)} 天 ({train_end} ~ {val_end})")
print(f" 测试集: {len(test_data)} 天 (> {val_end})")
return train_data, val_data, test_data
def build_sequences_by_concept(
date_data: Dict[str, pd.DataFrame],
features: List[str],
seq_len: int,
stride: int
) -> np.ndarray:
"""按概念分组构建序列"""
all_dfs = []
for date, df in sorted(date_data.items()):
df = df.copy()
df['date'] = date
all_dfs.append(df)
if not all_dfs:
return np.array([])
combined = pd.concat(all_dfs, ignore_index=True)
combined = combined.sort_values(['concept_id', 'date', 'timestamp'])
all_sequences = []
grouped = combined.groupby('concept_id', sort=False)
n_concepts = len(grouped)
for concept_id, concept_df in tqdm(grouped, desc="构建序列", total=n_concepts, leave=False):
feature_data = concept_df[features].values
feature_data = np.nan_to_num(feature_data, nan=0.0, posinf=0.0, neginf=0.0)
n_points = len(feature_data)
for start in range(0, n_points - seq_len + 1, stride):
seq = feature_data[start:start + seq_len]
all_sequences.append(seq)
if not all_sequences:
return np.array([])
sequences = np.array(all_sequences)
print(f" 构建序列: {len(sequences):,} 条 (来自 {n_concepts} 个概念)")
return sequences
# ==================== 数据集 ====================
class SequenceDataset(Dataset):
def __init__(self, sequences: np.ndarray):
self.sequences = torch.FloatTensor(sequences)
def __len__(self) -> int:
return len(self.sequences)
def __getitem__(self, idx: int) -> torch.Tensor:
return self.sequences[idx]
# ==================== 训练器 ====================
class EarlyStopping:
def __init__(self, patience: int = 10, min_delta: float = 1e-6):
self.patience = patience
self.min_delta = min_delta
self.counter = 0
self.best_loss = float('inf')
self.early_stop = False
def __call__(self, val_loss: float) -> bool:
if val_loss < self.best_loss - self.min_delta:
self.best_loss = val_loss
self.counter = 0
else:
self.counter += 1
if self.counter >= self.patience:
self.early_stop = True
return self.early_stop
class Trainer:
def __init__(
self,
model: nn.Module,
train_loader: DataLoader,
val_loader: DataLoader,
config: Dict,
device: torch.device,
save_dir: str = 'ml/checkpoints_v2'
):
self.model = model.to(device)
self.train_loader = train_loader
self.val_loader = val_loader
self.config = config
self.device = device
self.save_dir = Path(save_dir)
self.save_dir.mkdir(parents=True, exist_ok=True)
self.optimizer = AdamW(
model.parameters(),
lr=config['learning_rate'],
weight_decay=config['weight_decay']
)
self.scheduler = CosineAnnealingWarmRestarts(
self.optimizer, T_0=10, T_mult=2, eta_min=1e-6
)
self.criterion = AnomalyDetectionLoss()
self.early_stopping = EarlyStopping(
patience=config['patience'],
min_delta=config['min_delta']
)
self.use_amp = torch.cuda.is_available()
self.scaler = torch.cuda.amp.GradScaler() if self.use_amp else None
if self.use_amp:
print(" ✓ 启用 AMP 混合精度训练")
self.history = {'train_loss': [], 'val_loss': [], 'learning_rate': []}
self.best_val_loss = float('inf')
def train_epoch(self) -> float:
self.model.train()
total_loss = 0.0
n_batches = 0
pbar = tqdm(self.train_loader, desc="Training", leave=False)
for batch in pbar:
batch = batch.to(self.device, non_blocking=True)
self.optimizer.zero_grad(set_to_none=True)
if self.use_amp:
with torch.cuda.amp.autocast():
output, latent = self.model(batch)
loss, _ = self.criterion(output, batch, latent)
self.scaler.scale(loss).backward()
self.scaler.unscale_(self.optimizer)
torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.config['gradient_clip'])
self.scaler.step(self.optimizer)
self.scaler.update()
else:
output, latent = self.model(batch)
loss, _ = self.criterion(output, batch, latent)
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.config['gradient_clip'])
self.optimizer.step()
total_loss += loss.item()
n_batches += 1
pbar.set_postfix({'loss': f"{loss.item():.4f}"})
return total_loss / n_batches
@torch.no_grad()
def validate(self) -> float:
self.model.eval()
total_loss = 0.0
n_batches = 0
for batch in self.val_loader:
batch = batch.to(self.device, non_blocking=True)
if self.use_amp:
with torch.cuda.amp.autocast():
output, latent = self.model(batch)
loss, _ = self.criterion(output, batch, latent)
else:
output, latent = self.model(batch)
loss, _ = self.criterion(output, batch, latent)
total_loss += loss.item()
n_batches += 1
return total_loss / n_batches
def save_checkpoint(self, epoch: int, val_loss: float, is_best: bool = False):
model_to_save = self.model.module if hasattr(self.model, 'module') else self.model
checkpoint = {
'epoch': epoch,
'model_state_dict': model_to_save.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'scheduler_state_dict': self.scheduler.state_dict(),
'val_loss': val_loss,
'config': self.config,
}
torch.save(checkpoint, self.save_dir / 'last_checkpoint.pt')
if is_best:
torch.save(checkpoint, self.save_dir / 'best_model.pt')
print(f" ✓ 保存最佳模型 (val_loss: {val_loss:.6f})")
def train(self, epochs: int):
print(f"\n开始训练 ({epochs} epochs)...")
print(f"设备: {self.device}")
print(f"模型参数量: {count_parameters(self.model):,}")
for epoch in range(1, epochs + 1):
print(f"\nEpoch {epoch}/{epochs}")
train_loss = self.train_epoch()
val_loss = self.validate()
self.scheduler.step()
current_lr = self.optimizer.param_groups[0]['lr']
self.history['train_loss'].append(train_loss)
self.history['val_loss'].append(val_loss)
self.history['learning_rate'].append(current_lr)
print(f" Train Loss: {train_loss:.6f}")
print(f" Val Loss: {val_loss:.6f}")
print(f" LR: {current_lr:.2e}")
is_best = val_loss < self.best_val_loss
if is_best:
self.best_val_loss = val_loss
self.save_checkpoint(epoch, val_loss, is_best)
if self.early_stopping(val_loss):
print(f"\n早停触发！")
break
print(f"\n训练完成！最佳验证损失: {self.best_val_loss:.6f}")
self.save_history()
return self.history
def save_history(self):
history_path = self.save_dir / 'training_history.json'
with open(history_path, 'w') as f:
json.dump(self.history, f, indent=2)
print(f"训练历史已保存: {history_path}")
if HAS_MATPLOTLIB:
self.plot_training_curves()
def plot_training_curves(self):
fig, axes = plt.subplots(1, 2, figsize=(14, 5))
epochs = range(1, len(self.history['train_loss']) + 1)
ax1 = axes[0]
ax1.plot(epochs, self.history['train_loss'], 'b-', label='Train Loss', linewidth=2)
ax1.plot(epochs, self.history['val_loss'], 'r-', label='Val Loss', linewidth=2)
ax1.set_xlabel('Epoch')
ax1.set_ylabel('Loss')
ax1.set_title('Training & Validation Loss (V2)')
ax1.legend()
ax1.grid(True, alpha=0.3)
best_epoch = np.argmin(self.history['val_loss']) + 1
best_val_loss = min(self.history['val_loss'])
ax1.axvline(x=best_epoch, color='g', linestyle='--', alpha=0.7)
ax1.scatter([best_epoch], [best_val_loss], color='g', s=100, zorder=5)
ax2 = axes[1]
ax2.plot(epochs, self.history['learning_rate'], 'g-', linewidth=2)
ax2.set_xlabel('Epoch')
ax2.set_ylabel('Learning Rate')
ax2.set_title('Learning Rate Schedule')
ax2.set_yscale('log')
ax2.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(self.save_dir / 'training_curves.png', dpi=150, bbox_inches='tight')
plt.close()
print(f"训练曲线已保存")
# ==================== 阈值计算 ====================
@torch.no_grad()
def compute_thresholds(
model: nn.Module,
data_loader: DataLoader,
device: torch.device,
percentiles: List[float] = [90, 95, 99]
) -> Dict[str, float]:
"""在验证集上计算阈值"""
model.eval()
all_errors = []
print("计算异动阈值...")
for batch in tqdm(data_loader, desc="Computing thresholds"):
batch = batch.to(device)
errors = model.compute_reconstruction_error(batch, reduction='none')
seq_errors = errors[:, -1] # 最后一个时刻
all_errors.append(seq_errors.cpu().numpy())
all_errors = np.concatenate(all_errors)
thresholds = {}
for p in percentiles:
threshold = np.percentile(all_errors, p)
thresholds[f'p{p}'] = float(threshold)
print(f" P{p}: {threshold:.6f}")
thresholds['mean'] = float(np.mean(all_errors))
thresholds['std'] = float(np.std(all_errors))
thresholds['median'] = float(np.median(all_errors))
return thresholds
# ==================== 主函数 ====================
def main():
parser = argparse.ArgumentParser(description='训练 V2 模型')
parser.add_argument('--data_dir', type=str, default='ml/data_v2', help='V2 数据目录')
parser.add_argument('--epochs', type=int, default=100)
parser.add_argument('--batch_size', type=int, default=4096)
parser.add_argument('--lr', type=float, default=3e-4)
parser.add_argument('--device', type=str, default='auto')
parser.add_argument('--save_dir', type=str, default='ml/checkpoints_v2')
parser.add_argument('--train_end', type=str, default='2024-06-30')
parser.add_argument('--val_end', type=str, default='2024-09-30')
parser.add_argument('--seq_len', type=int, default=10, help='序列长度（分钟）')
args = parser.parse_args()
config = TRAIN_CONFIG.copy()
config['batch_size'] = args.batch_size
config['epochs'] = args.epochs
config['learning_rate'] = args.lr
config['train_end_date'] = args.train_end
config['val_end_date'] = args.val_end
config['seq_len'] = args.seq_len
if args.device == 'auto':
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
else:
device = torch.device(args.device)
print("=" * 60)
print("概念异动检测模型训练 V2（Z-Score 特征）")
print("=" * 60)
print(f"数据目录: {args.data_dir}")
print(f"设备: {device}")
print(f"序列长度: {config['seq_len']} 分钟")
print(f"批次大小: {config['batch_size']}")
print(f"特征: {config['features']}")
print("=" * 60)
# 1. 加载数据
print("\n[1/6] 加载 V2 数据...")
date_data = load_data_by_date(args.data_dir, config['features'])
# 2. 划分数据集
print("\n[2/6] 划分数据集...")
train_data, val_data, test_data = split_data_by_date(
date_data, config['train_end_date'], config['val_end_date']
)
# 3. 构建序列
print("\n[3/6] 构建序列...")
print("训练集:")
train_sequences = build_sequences_by_concept(
train_data, config['features'], config['seq_len'], config['stride']
)
print("验证集:")
val_sequences = build_sequences_by_concept(
val_data, config['features'], config['seq_len'], config['stride']
)
if len(train_sequences) == 0:
print("错误: 训练集为空！")
return
# 4. 预处理
print("\n[4/6] 数据预处理...")
clip_value = config['clip_value']
print(f" Z-Score 特征已标准化，截断: ±{clip_value}")
train_sequences = np.clip(train_sequences, -clip_value, clip_value)
if len(val_sequences) > 0:
val_sequences = np.clip(val_sequences, -clip_value, clip_value)
# 保存配置
save_dir = Path(args.save_dir)
save_dir.mkdir(parents=True, exist_ok=True)
with open(save_dir / 'config.json', 'w') as f:
json.dump(config, f, indent=2)
# 5. 创建数据加载器
print("\n[5/6] 创建数据加载器...")
train_dataset = SequenceDataset(train_sequences)
val_dataset = SequenceDataset(val_sequences) if len(val_sequences) > 0 else None
print(f" 训练序列: {len(train_dataset):,}")
print(f" 验证序列: {len(val_dataset) if val_dataset else 0:,}")
n_gpus = torch.cuda.device_count() if torch.cuda.is_available() else 1
num_workers = min(32, 8 * n_gpus) if sys.platform != 'win32' else 0
train_loader = DataLoader(
train_dataset,
batch_size=config['batch_size'],
shuffle=True,
num_workers=num_workers,
pin_memory=True,
prefetch_factor=4 if num_workers > 0 else None,
persistent_workers=True if num_workers > 0 else False,
drop_last=True
)
val_loader = DataLoader(
val_dataset,
batch_size=config['batch_size'] * 2,
shuffle=False,
num_workers=num_workers,
pin_memory=True,
) if val_dataset else None
# 6. 训练
print("\n[6/6] 训练模型...")
model = TransformerAutoencoder(**config['model'])
if torch.cuda.device_count() > 1:
print(f" 使用 {torch.cuda.device_count()} 张 GPU 并行训练")
model = nn.DataParallel(model)
if val_loader is None:
print("警告: 验证集为空，使用训练集的 10% 作为验证")
split_idx = int(len(train_dataset) * 0.9)
train_subset = torch.utils.data.Subset(train_dataset, range(split_idx))
val_subset = torch.utils.data.Subset(train_dataset, range(split_idx, len(train_dataset)))
train_loader = DataLoader(train_subset, batch_size=config['batch_size'], shuffle=True, num_workers=num_workers, pin_memory=True)
val_loader = DataLoader(val_subset, batch_size=config['batch_size'], shuffle=False, num_workers=num_workers, pin_memory=True)
trainer = Trainer(
model=model,
train_loader=train_loader,
val_loader=val_loader,
config=config,
device=device,
save_dir=args.save_dir
)
trainer.train(config['epochs'])
# 计算阈值
print("\n[额外] 计算异动阈值...")
best_checkpoint = torch.load(save_dir / 'best_model.pt', map_location=device)
# 创建新的单 GPU 模型用于计算阈值（避免 DataParallel 问题）
threshold_model = TransformerAutoencoder(**config['model'])
threshold_model.load_state_dict(best_checkpoint['model_state_dict'])
threshold_model.to(device)
threshold_model.eval()
thresholds = compute_thresholds(threshold_model, val_loader, device, config['threshold_percentiles'])
with open(save_dir / 'thresholds.json', 'w') as f:
json.dump(thresholds, f, indent=2)
print("\n" + "=" * 60)
print("训练完成！")
print(f"模型保存位置: {args.save_dir}")
print("=" * 60)
if __name__ == "__main__":
main()

132
ml/update_baseline.py
View File

@@ -1,132 +0,0 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
每日盘后运行：更新滚动基线
使用方法：
python ml/update_baseline.py
建议加入 crontab，每天 15:30 后运行：
30 15 * * 1-5 cd /path/to/project && python ml/update_baseline.py
"""
import os
import sys
import pickle
import pandas as pd
import numpy as np
from datetime import datetime, timedelta
from pathlib import Path
from tqdm import tqdm
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from ml.prepare_data_v2 import (
get_all_concepts, get_trading_days, compute_raw_concept_features,
init_process_connections, CONFIG, RAW_CACHE_DIR, BASELINE_DIR
)
def update_rolling_baseline(baseline_days: int = 20):
"""
更新滚动基线（用于实盘检测）
基线 = 最近 N 个交易日每个时间片的统计量
"""
print("=" * 60)
print("更新滚动基线（用于实盘）")
print("=" * 60)
# 初始化连接
init_process_connections()
# 获取概念列表
concepts = get_all_concepts()
all_stocks = list(set(s for c in concepts for s in c['stocks']))
# 获取最近的交易日
today = datetime.now().strftime('%Y-%m-%d')
start_date = (datetime.now() - timedelta(days=60)).strftime('%Y-%m-%d') # 多取一些
trading_days = get_trading_days(start_date, today)
if len(trading_days) < baseline_days:
print(f"错误：交易日不足 {baseline_days} 天")
return
# 只取最近 N 天
recent_days = trading_days[-baseline_days:]
print(f"使用 {len(recent_days)} 天数据: {recent_days[0]} ~ {recent_days[-1]}")
# 加载原始数据
all_data = []
for trade_date in tqdm(recent_days, desc="加载数据"):
cache_file = os.path.join(RAW_CACHE_DIR, f'raw_{trade_date}.parquet')
if os.path.exists(cache_file):
df = pd.read_parquet(cache_file)
else:
df = compute_raw_concept_features(trade_date, concepts, all_stocks)
if not df.empty:
all_data.append(df)
if not all_data:
print("错误：无数据")
return
combined = pd.concat(all_data, ignore_index=True)
print(f"总数据量: {len(combined):,} 条")
# 按概念计算基线
baselines = {}
for concept_id, group in tqdm(combined.groupby('concept_id'), desc="计算基线"):
baseline_dict = {}
for time_slot, slot_group in group.groupby('time_slot'):
if len(slot_group) < CONFIG['min_baseline_samples']:
continue
alpha_std = slot_group['alpha'].std()
amt_std = slot_group['total_amt'].std()
rank_std = slot_group['rank_pct'].std()
baseline_dict[time_slot] = {
'alpha_mean': float(slot_group['alpha'].mean()),
'alpha_std': float(max(alpha_std if pd.notna(alpha_std) else 1.0, 0.1)),
'amt_mean': float(slot_group['total_amt'].mean()),
'amt_std': float(max(amt_std if pd.notna(amt_std) else slot_group['total_amt'].mean() * 0.5, 1.0)),
'rank_mean': float(slot_group['rank_pct'].mean()),
'rank_std': float(max(rank_std if pd.notna(rank_std) else 0.2, 0.05)),
'sample_count': len(slot_group),
}
if baseline_dict:
baselines[concept_id] = baseline_dict
print(f"计算了 {len(baselines)} 个概念的基线")
# 保存
os.makedirs(BASELINE_DIR, exist_ok=True)
baseline_file = os.path.join(BASELINE_DIR, 'realtime_baseline.pkl')
with open(baseline_file, 'wb') as f:
pickle.dump({
'baselines': baselines,
'update_time': datetime.now().isoformat(),
'date_range': [recent_days[0], recent_days[-1]],
'baseline_days': baseline_days,
}, f)
print(f"基线已保存: {baseline_file}")
print("=" * 60)
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--days', type=int, default=20, help='基线天数')
args = parser.parse_args()
update_rolling_baseline(args.days)

175
src/views/StockOverview/components/HotspotOverview/components/ConceptAlertList.js
View File

@@ -23,11 +23,17 @@ import {
Th,
Td,
TableContainer,
Popover,
PopoverTrigger,
PopoverContent,
PopoverBody,
Portal,
} from '@chakra-ui/react';
import { FaArrowUp, FaArrowDown, FaFire, FaChevronDown, FaChevronRight } from 'react-icons/fa';
import { useNavigate } from 'react-router-dom';
import axios from 'axios';
import { getAlertTypeLabel, formatScore, getScoreColor } from '../utils/chartHelpers';
import MiniTimelineChart from '@components/Charts/Stock/MiniTimelineChart';
/**
* 紧凑型异动卡片
@@ -38,6 +44,8 @@ const AlertCard = ({ alert, isExpanded, onToggle, stocks, loadingStocks }) => {
const hoverBg = useColorModeValue('gray.50', '#252525');
const borderColor = useColorModeValue('gray.200', '#333');
const expandedBg = useColorModeValue('purple.50', '#1e1e2e');
const tableBg = useColorModeValue('gray.50', '#151520');
const popoverBg = useColorModeValue('white', '#1a1a1a');
const isUp = alert.alert_type !== 'surge_down';
const typeColor = isUp ? 'red' : 'green';
@@ -175,7 +183,7 @@ const AlertCard = ({ alert, isExpanded, onToggle, stocks, loadingStocks }) => {
borderTopWidth="1px"
borderColor={borderColor}
p={3}
bg={useColorModeValue('gray.50', '#151520')}
bg={tableBg}
>
{loadingStocks ? (
<HStack justify="center" py={4}>
@@ -183,62 +191,115 @@ const AlertCard = ({ alert, isExpanded, onToggle, stocks, loadingStocks }) => {
<Text fontSize="sm" color="gray.500">加载相关股票...</Text>
</HStack>
) : stocks && stocks.length > 0 ? (
<TableContainer maxH="200px" overflowY="auto">
<Table size="sm" variant="simple">
<Thead position="sticky" top={0} bg={useColorModeValue('gray.50', '#151520')} zIndex={1}>
<Tr>
<Th px={2} py={1} fontSize="xs" color="gray.500">股票</Th>
<Th px={2} py={1} fontSize="xs" color="gray.500" isNumeric>涨跌</Th>
<Th px={2} py={1} fontSize="xs" color="gray.500" maxW="120px">原因</Th>
</Tr>
</Thead>
<Tbody>
{stocks.slice(0, 10).map((stock, idx) => {
const changePct = stock.change_pct;
const hasChange = changePct != null && !isNaN(changePct);
return (
<Tr
key={idx}
cursor="pointer"
_hover={{ bg: hoverBg }}
onClick={(e) => handleStockClick(e, stock.code || stock.stock_code)}
>
<Td px={2} py={1.5}>
<Text fontSize="xs" color="cyan.400" fontWeight="medium">
{stock.name || stock.stock_name || '-'}
</Text>
</Td>
<Td px={2} py={1.5} isNumeric>
<Text
fontSize="xs"
fontWeight="bold"
color={
hasChange && changePct > 0 ? 'red.400' :
hasChange && changePct < 0 ? 'green.400' : 'gray.400'
}
>
{hasChange
? `${changePct > 0 ? '+' : ''}${changePct.toFixed(2)}%`
: '-'
}
</Text>
</Td>
<Td px={2} py={1.5} maxW="120px">
<Text fontSize="xs" color="gray.500" noOfLines={1}>
{stock.reason || '-'}
</Text>
</Td>
</Tr>
);
})}
</Tbody>
</Table>
{stocks.length > 10 && (
<Text fontSize="xs" color="gray.500" textAlign="center" mt={2}>
共 {stocks.length} 只相关股票，显示前 10 只
</Text>
)}
</TableContainer>
<>
{/* 概念涨跌幅统计 */}
{(() => {
const validStocks = stocks.filter(s => s.change_pct != null && !isNaN(s.change_pct));
if (validStocks.length === 0) return null;
const avgChange = validStocks.reduce((sum, s) => sum + s.change_pct, 0) / validStocks.length;
const upCount = validStocks.filter(s => s.change_pct > 0).length;
const downCount = validStocks.filter(s => s.change_pct < 0).length;
return (
<HStack spacing={4} mb={2} fontSize="xs" color="gray.500">
<HStack>
<Text>概念均涨:</Text>
<Text fontWeight="bold" color={avgChange >= 0 ? 'red.400' : 'green.400'}>
{avgChange >= 0 ? '+' : ''}{avgChange.toFixed(2)}%
</Text>
</HStack>
<HStack spacing={1}>
<Text color="red.400">{upCount}涨</Text>
<Text>/</Text>
<Text color="green.400">{downCount}跌</Text>
</HStack>
</HStack>
);
})()}
<TableContainer maxH="200px" overflowY="auto">
<Table size="sm" variant="simple">
<Thead position="sticky" top={0} bg={tableBg} zIndex={1}>
<Tr>
<Th px={2} py={1} fontSize="xs" color="gray.500">股票</Th>
<Th px={2} py={1} fontSize="xs" color="gray.500" isNumeric>涨跌</Th>
<Th px={2} py={1} fontSize="xs" color="gray.500" maxW="120px">原因</Th>
</Tr>
</Thead>
<Tbody>
{stocks.slice(0, 10).map((stock, idx) => {
const changePct = stock.change_pct;
const hasChange = changePct != null && !isNaN(changePct);
const stockCode = stock.code || stock.stock_code;
const stockName = stock.name || stock.stock_name || '-';
return (
<Tr
key={idx}
cursor="pointer"
_hover={{ bg: hoverBg }}
onClick={(e) => handleStockClick(e, stockCode)}
>
<Td px={2} py={1.5}>
<Popover trigger="hover" placement="right" isLazy>
<PopoverTrigger>
<Text
fontSize="xs"
color="cyan.400"
fontWeight="medium"
_hover={{ textDecoration: 'underline' }}
>
{stockName}
</Text>
</PopoverTrigger>
<Portal>
<PopoverContent
w="180px"
h="80px"
bg={popoverBg}
borderColor={borderColor}
boxShadow="lg"
onClick={(e) => e.stopPropagation()}
>
<PopoverBody p={2}>
<Text fontSize="xs" color="gray.500" mb={1}>{stockName} 分时</Text>
<Box h="50px">
<MiniTimelineChart stockCode={stockCode} />
</Box>
</PopoverBody>
</PopoverContent>
</Portal>
</Popover>
</Td>
<Td px={2} py={1.5} isNumeric>
<Text
fontSize="xs"
fontWeight="bold"
color={
hasChange && changePct > 0 ? 'red.400' :
hasChange && changePct < 0 ? 'green.400' : 'gray.400'
}
>
{hasChange
? `${changePct > 0 ? '+' : ''}${changePct.toFixed(2)}%`
: '-'
}
</Text>
</Td>
<Td px={2} py={1.5} maxW="120px">
<Text fontSize="xs" color="gray.500" noOfLines={1}>
{stock.reason || '-'}
</Text>
</Td>
</Tr>
);
})}
</Tbody>
</Table>
{stocks.length > 10 && (
<Text fontSize="xs" color="gray.500" textAlign="center" mt={2}>
共 {stocks.length} 只相关股票，显示前 10 只
</Text>
)}
</TableContainer>
</>
) : (
<Text fontSize="sm" color="gray.500" textAlign="center" py={2}>
暂无相关股票数据

16
src/views/StockOverview/index.js
View File

@@ -843,8 +843,22 @@ const StockOverview = () => {
</Box>
{/* 热点概览 - 大盘走势 + 概念异动 */}
{/* 只在 selectedDate 确定后渲染，避免 null → 日期 的双重请求 */}
<Box mb={10}>
<HotspotOverview selectedDate={selectedDate} />
{selectedDate ? (
<HotspotOverview selectedDate={selectedDate} />
) : (
<Card bg={cardBg} borderWidth="1px" borderColor={borderColor}>
<CardBody>
<Center h="400px">
<VStack spacing={4}>
<Spinner size="xl" color="purple.500" thickness="4px" />
<Text color={subTextColor}>加载热点概览数据...</Text>
</VStack>
</Center>
</CardBody>
</Card>
)}
</Box>
{/* 灵活屏 - 实时行情监控 */}