ensemble file

ensemble_dlinear_jan_2026.py

@@ -0,0 +1,289 @@
+"""2026年1月BTC数据 - DLinear 准确率测试
+
+输入数据：预测K线的前一根（即用截止到t的数据预测t+1）
+价格对比：预测结果与预测K线(t+1)的收盘价做对比
+时间：币安数据UTC，输出转为北京时间(UTC+8)，按北京时间分天
+DLinear预测逻辑：整个df一次性预处理，按索引遍历窗口
+输出：每小时详细数据、每天准确率、总准确率、美股开盘标记
+"""
+
+import sys
+import os
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+
+import torch
+import numpy as np
+import pandas as pd
+import pickle
+import pandas_ta as ta
+from datetime import timedelta
+from sklearn.preprocessing import StandardScaler
+from data.fetcher import BinanceDataFetcher
+
+# ==================== 时区与美股判断 ====================
+
+def utc_to_beijing(ts):
+    """UTC时间转北京时间(UTC+8)"""
+    return ts + timedelta(hours=8)
+
+def is_us_stock_open_by_utc(utc_ts):
+    """根据UTC时间判断美股当天是否开盘
+
+    美股交易日基于美东时间(ET)，UTC-5(冬令时)
+    2026年1月在冬令时范围内（夏令时3月第二个周日开始）
+    所以美东 = UTC - 5h
+    """
+    et_ts = utc_ts - timedelta(hours=5)
+    et_date = pd.Timestamp(et_ts.date())
+
+    us_holidays_jan_2026 = {
+        pd.Timestamp('2026-01-01'),  # 元旦 New Year's Day
+        pd.Timestamp('2026-01-19'),  # 马丁·路德·金日 MLK Day
+    }
+
+    if et_date.weekday() >= 5:
+        return False
+    if et_date in us_holidays_jan_2026:
+        return False
+    return True
+
+# ==================== pickle反序列化占位类 ====================
+
+class ClassificationPreprocessor:
+    """pickle.load时需要能找到这个类名"""
+    def __init__(self, lookback=168, forecast_horizon=1):
+        self.lookback = lookback
+        self.forecast_horizon = forecast_horizon
+        self.scaler = StandardScaler()
+        self.is_fitted = False
+
+# ==================== 工具函数 ====================
+
+def fmt_pred(val):
+    """格式化预测结果"""
+    if val is None or (isinstance(val, float) and np.isnan(val)):
+        return "N/A"
+    return "涨" if int(val) == 1 else "跌"
+
+def fmt_result(pred, actual):
+    """格式化对错"""
+    if pred is None or (isinstance(pred, float) and np.isnan(pred)):
+        return "  "
+    return "✓" if int(pred) == int(actual) else "✗"
+
+
+def main():
+    print("=" * 90)
+    print("2026年1月 BTC 1h 涨跌预测准确率测试 (DLinear)")
+    print("输入：预测K线的前一根 | 对比：预测结果 vs 预测K线收盘价 | 时间：北京时间(UTC+8)")
+    print("=" * 90)
+
+    device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
+    print(f"使用设备: {device}")
+
+    # ========== 步骤1: 加载模型 ==========
+    print("\n[1/4] 加载模型...")
+
+    from models.dlinear.model import DLinear
+    checkpoint = torch.load('./models/checkpoints/dlinear_classification_2026/best_model.pt',
+                            map_location=device, weights_only=False)
+    model = DLinear(checkpoint['config'])
+    model.load_state_dict(checkpoint['model_state_dict'])
+    model = model.to(device).eval()
+    with open('./models/checkpoints/dlinear_classification_2026/preprocessor.pkl', 'rb') as f:
+        preprocessor = pickle.load(f)
+    print("  ✓ DLinear已加载")
+
+    # ========== 步骤2: 获取测试数据 ==========
+    print("\n[2/4] 获取数据...")
+    fetcher = BinanceDataFetcher()
+    df = fetcher.fetch_ohlcv('BTC', '1h', start_date='2025-12-01', end_date='2026-03-01', use_cache=False)
+
+    print(f"数据范围(UTC): {df['timestamp'].min()} -> {df['timestamp'].max()}")
+    print(f"总行数: {len(df)}")
+
+    # ========== 步骤3: 预处理与预测 ==========
+    print("\n[3/4] 运行预测...")
+
+    # DLinear 整体预处理（无create_labels，整个df一次性处理）
+    lookback = preprocessor.lookback
+    df_dl = df.copy()
+    df_dl['rsi'] = ta.rsi(df_dl['close'], length=14)
+    macd = ta.macd(df_dl['close'])
+    if macd is not None and not macd.empty:
+        df_dl['macd'] = macd.iloc[:, 0]
+        df_dl['macd_hist'] = macd.iloc[:, 1]
+        df_dl['macd_signal'] = macd.iloc[:, 2]
+    df_dl = df_dl.dropna().reset_index(drop=True)
+
+    feature_cols = ['open', 'high', 'low', 'close', 'volume',
+                   'rsi', 'macd', 'macd_hist', 'macd_signal']
+    scaled = preprocessor.scaler.transform(df_dl[feature_cols])
+
+    X_dl = []
+    idx_dl = []
+    for i in range(len(scaled) - lookback):
+        X_dl.append(scaled[i:i + lookback])
+        idx_dl.append(i + lookback - 1)
+    X_dl = np.array(X_dl)
+
+    # 构建 timestamp -> 窗口索引映射
+    dl_ts_map = {}
+    for i, idx in enumerate(idx_dl):
+        ts = df_dl.iloc[idx]['timestamp']
+        dl_ts_map[ts] = i
+
+    # 找出1月（北京时间）的所有预测目标K线索引
+    jan_targets = []
+    for target_idx in range(1, len(df)):
+        target_time_utc = df.iloc[target_idx]['timestamp']
+        target_time_bj = utc_to_beijing(target_time_utc)
+        if target_time_bj.year == 2026 and target_time_bj.month == 2:
+            jan_targets.append(target_idx)
+
+    print(f"1月（北京时间）目标K线数: {len(jan_targets)}")
+
+    results = []
+
+    for target_idx in jan_targets:
+        target_time_utc = df.iloc[target_idx]['timestamp']
+        target_time_bj = utc_to_beijing(target_time_utc)
+
+        # 输入K线索引（前一根）
+        input_idx = target_idx - 1
+        current_close = df.iloc[input_idx]['close']
+        target_close = df.iloc[target_idx]['close']
+        actual_label = 1 if target_close > current_close else 0
+
+        # --- DLinear 预测 ---
+        dl_pred = None
+        dl_prob = None
+        input_ts = df.iloc[input_idx]['timestamp']
+        dl_i = dl_ts_map.get(input_ts)
+        if dl_i is not None and dl_i < len(X_dl):
+            x_tensor = torch.FloatTensor(X_dl[dl_i:dl_i+1]).to(device)
+            with torch.no_grad():
+                logits = model(x_tensor).cpu().numpy()
+                prob = 1 / (1 + np.exp(-logits))
+                dl_pred = int(prob.item() > 0.5)
+                dl_prob = float(prob.item())
+
+        results.append({
+            'time_utc': target_time_utc,
+            'time_bj': target_time_bj,
+            'current_close': current_close,
+            'target_close': target_close,
+            'actual': actual_label,
+            'dlinear': dl_pred,
+            'dl_prob': dl_prob,
+        })
+
+    # ========== 步骤4: 输出结果 ==========
+    if not results:
+        print("没有找到符合条件的测试样本")
+        return
+
+    df_res = pd.DataFrame(results)
+    df_res['date_bj'] = df_res['time_bj'].dt.date
+    df_res['us_open'] = df_res['time_utc'].apply(is_us_stock_open_by_utc)
+
+    total_samples = len(df_res)
+
+    print(f"\n[4/4] 准确率统计 (2026年1月，北京时间)")
+    print(f"总样本数: {total_samples}")
+
+    # ========== 每小时详细数据 ==========
+    print("\n" + "=" * 90)
+    print("每小时详细数据")
+    print("=" * 90)
+
+    hourly_header = (
+        f"{'北京时间':<18} {'前收盘':>10} {'现收盘':>10} {'实际':>4} "
+        f"{'DL':>3} {'':>1} {'概率':>6}"
+    )
+
+    for date_bj, day_group in df_res.groupby('date_bj'):
+        us_open = day_group['us_open'].iloc[0]
+        us_str = "美股开盘" if us_open else "美股休市"
+
+        dl_day_correct = (day_group['dlinear'] == day_group['actual']).sum()
+        dl_day_total = day_group['dlinear'].notna().sum()
+        dl_day_acc = dl_day_correct / dl_day_total if dl_day_total > 0 else 0
+
+        print(f"\n--- {date_bj} ({us_str}) --- DL准确率:{dl_day_acc:.1%} ({dl_day_correct}/{dl_day_total})")
+        print(hourly_header)
+        print("-" * 90)
+
+        for _, row in day_group.iterrows():
+            time_str = row['time_bj'].strftime('%m-%d %H:%M')
+            actual_str = fmt_pred(row['actual'])
+            dl_str = fmt_pred(row['dlinear'])
+            dl_r = fmt_result(row['dlinear'], row['actual'])
+            prob_str = f"{row['dl_prob']:.3f}" if row['dl_prob'] is not None else "  N/A"
+
+            print(
+                f"{time_str:<18} {row['current_close']:>10.2f} {row['target_close']:>10.2f} {actual_str:>4} "
+                f"{dl_str:>3} {dl_r:>1} {prob_str:>6}"
+            )
+
+    # ========== 每日准确率汇总表 ==========
+    print("\n\n" + "=" * 90)
+    print("每日准确率汇总 (北京时间)")
+    print("=" * 90)
+
+    header = f"{'日期':<14} {'样本':>4} {'美股':>6}  {'DLinear':>8}"
+    print(header)
+    print("-" * 90)
+
+    for date_bj, group in df_res.groupby('date_bj'):
+        total = len(group)
+        us_open = group['us_open'].iloc[0]
+        us_str = "开盘" if us_open else "休市"
+
+        dl_valid = group.dropna(subset=['dlinear'])
+        dl_acc = (dl_valid['dlinear'] == dl_valid['actual']).mean() if len(dl_valid) > 0 else 0
+        dl_str = f"{dl_acc:.1%}"
+
+        print(f"{str(date_bj):<14} {total:>4} {us_str:>6}  {dl_str:>8}")
+
+    print("-" * 90)
+
+    # --- 总准确率 ---
+    dl_valid_all = df_res.dropna(subset=['dlinear'])
+    dl_total = (dl_valid_all['dlinear'] == dl_valid_all['actual']).mean() if len(dl_valid_all) > 0 else 0
+
+    print(f"{'总计':<14} {total_samples:>4} {'':>6}  {dl_total:.2%}")
+    print("=" * 90)
+
+    # --- 美股开盘日 vs 休市日对比 ---
+    print("\n[美股开盘日 vs 休市日 准确率对比]")
+    print("-" * 60)
+    for label, flag in [("开盘日", True), ("休市日", False)]:
+        subset = df_res[df_res['us_open'] == flag]
+        if len(subset) == 0:
+            continue
+        n = len(subset)
+        dl_s = subset.dropna(subset=['dlinear'])
+        dl_a = (dl_s['dlinear'] == dl_s['actual']).mean() if len(dl_s) > 0 else 0
+        print(f"  {label} ({n}样本): DLinear={dl_a:.2%}")
+
+    print("-" * 60)
+
+    # --- 预测分布统计 ---
+    print("\n[预测分布统计]")
+    print("-" * 60)
+    actual_up = (df_res['actual'] == 1).sum()
+    actual_down = (df_res['actual'] == 0).sum()
+    print(f"  实际: 上涨={actual_up} ({actual_up/total_samples:.1%})  下跌={actual_down} ({actual_down/total_samples:.1%})")
+
+    dl_valid = df_res['dlinear'].dropna()
+    if len(dl_valid) > 0:
+        up = (dl_valid == 1).sum()
+        down = (dl_valid == 0).sum()
+        print(f"  DLinear: 预测上涨={up} ({up/len(dl_valid):.1%})  预测下跌={down} ({down/len(dl_valid):.1%})")
+
+    print("=" * 90)
+
+
+if __name__ == "__main__":
+    main()