"""按天统计双模型和三模型的准确率""" import sys import os sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) import torch import pickle import numpy as np import pandas as pd import pandas_ta as ta from sklearn.preprocessing import StandardScaler from data.fetcher import BinanceDataFetcher # ==================== 预处理器定义 ==================== class ClassificationPreprocessor: """DLinear分类预处理器""" def __init__(self, lookback=168, forecast_horizon=1): self.lookback = lookback self.forecast_horizon = forecast_horizon self.scaler = StandardScaler() self.is_fitted = False def add_features(self, df): df = df.copy() df['rsi'] = ta.rsi(df['close'], length=14) macd = ta.macd(df['close']) if macd is not None and not macd.empty: df['macd'] = macd.iloc[:, 0] df['macd_hist'] = macd.iloc[:, 1] df['macd_signal'] = macd.iloc[:, 2] return df def create_labels(self, df): df = df.copy() df['future_price'] = df['close'].shift(-self.forecast_horizon) df['future_return'] = (df['future_price'] - df['close']) / df['close'] df['label'] = (df['future_return'] > 0).astype(int) return df[:-self.forecast_horizon] def preprocess(self, df, fit=True): df = self.add_features(df) df = self.create_labels(df) df = df.dropna().reset_index(drop=True) feature_cols = ['open', 'high', 'low', 'close', 'volume', 'rsi', 'macd', 'macd_hist', 'macd_signal'] if fit: scaled_features = self.scaler.fit_transform(df[feature_cols]) self.is_fitted = True self.feature_columns = feature_cols else: scaled_features = self.scaler.transform(df[feature_cols]) X, y = [], [] for i in range(len(scaled_features) - self.lookback): X.append(scaled_features[i:i + self.lookback]) y.append(df['label'].iloc[i + self.lookback]) return np.array(X), np.array(y), df class NBeatsPreprocessor: """N-BEATS预处理器""" def __init__(self, lookback=168, forecast_horizon=1): self.lookback = lookback self.forecast_horizon = forecast_horizon self.price_scaler = StandardScaler() self.is_fitted = False def preprocess(self, df, fit=True): df = df.copy() df['future_price'] = df['close'].shift(-self.forecast_horizon) df = df[:-self.forecast_horizon].reset_index(drop=True) prices = df['close'].values.reshape(-1, 1) future_prices = df['future_price'].values.reshape(-1, 1) if fit: scaled_prices = self.price_scaler.fit_transform(prices) scaled_future = self.price_scaler.transform(future_prices) self.is_fitted = True else: scaled_prices = self.price_scaler.transform(prices) scaled_future = self.price_scaler.transform(future_prices) X, y = [], [] for i in range(len(scaled_prices) - self.lookback): X.append(scaled_prices[i:i + self.lookback, 0]) y.append(scaled_future[i + self.lookback, 0]) return np.array(X), np.array(y), df def inverse_transform_price(self, scaled_price): return self.price_scaler.inverse_transform(scaled_price.reshape(-1, 1)).flatten() def main(): print("="*80) print("按天统计：双模型交集 vs 三模型投票 每日准确率分析") print("="*80) # ========== 加载模型 ========== print("\n[1/4] 加载模型...") device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu') # 加载DLinear from models.dlinear.model import DLinear dlinear_checkpoint = torch.load( './models/checkpoints/dlinear_classification_2026/best_model.pt', map_location=device ) dlinear_model = DLinear(dlinear_checkpoint['config']) dlinear_model.load_state_dict(dlinear_checkpoint['model_state_dict']) dlinear_model = dlinear_model.to(device) dlinear_model.eval() with open('./models/checkpoints/dlinear_classification_2026/preprocessor.pkl', 'rb') as f: dlinear_preprocessor = pickle.load(f) # 加载N-BEATS from models.nbeats.model import NBeats nbeats_checkpoint = torch.load( './models/checkpoints/nbeats_btc_1h/best_model.pt', map_location=device ) nbeats_model = NBeats(nbeats_checkpoint['config']) nbeats_model.load_state_dict(nbeats_checkpoint['model_state_dict']) nbeats_model = nbeats_model.to(device) nbeats_model.eval() with open('./models/checkpoints/nbeats_btc_1h/preprocessor.pkl', 'rb') as f: nbeats_preprocessor = pickle.load(f) # 加载TTM-R2 from tsfm_public.models.tinytimemixer import TinyTimeMixerForPrediction ttm_model = TinyTimeMixerForPrediction.from_pretrained( "ibm-granite/granite-timeseries-ttm-r2", trust_remote_code=True ) ttm_model = ttm_model.to(device) ttm_model.eval() print(f"✓ 三个模型已加载") # ========== 获取数据 ========== print("\n[2/4] 获取2026年测试数据...") fetcher = BinanceDataFetcher() try: df_2026 = fetcher.fetch_ohlcv('BTC', '1h', start_date='2025-12-26', end_date='2026-02-02', use_cache=False) except: df_2026 = fetcher.fetch_ohlcv('BTC', '1h', start_date='2025-12-01', use_cache=False) print(f"✓ 测试数据: {len(df_2026)} 行") # ========== 预处理 ========== print("\n[3/4] 预处理数据...") X_dlinear, y_dlinear, df_dlinear = dlinear_preprocessor.preprocess(df_2026, fit=False) X_nbeats, y_nbeats, df_nbeats = nbeats_preprocessor.preprocess(df_2026, fit=False) min_samples = min(len(X_dlinear), len(X_nbeats)) X_dlinear = X_dlinear[:min_samples] y_dlinear = y_dlinear[:min_samples] X_nbeats = X_nbeats[:min_samples] y_nbeats = y_nbeats[:min_samples] # ========== 预测 ========== print("\n[4/4] 执行预测...") # DLinear预测 X_dlinear_tensor = torch.FloatTensor(X_dlinear).to(device) with torch.no_grad(): dlinear_logits = dlinear_model(X_dlinear_tensor).cpu().numpy() dlinear_probs = 1 / (1 + np.exp(-dlinear_logits)) dlinear_preds = (dlinear_probs > 0.5).astype(int).flatten() # N-BEATS预测 X_nbeats_tensor = torch.FloatTensor(X_nbeats).to(device) with torch.no_grad(): nbeats_pred_scaled = nbeats_model(X_nbeats_tensor).cpu().numpy() nbeats_pred_prices = nbeats_preprocessor.inverse_transform_price(nbeats_pred_scaled) current_open_prices = df_nbeats['open'].iloc[168:168+len(nbeats_pred_prices)].values nbeats_preds = (nbeats_pred_prices > current_open_prices).astype(int) # TTM-R2预测 context_length = 512 ttm_preds_list = [] all_prices = df_2026['close'].values all_open_prices = df_2026['open'].values for i in range(min_samples): start_idx = i end_idx = start_idx + context_length if end_idx > len(all_prices): context_data = all_prices[-context_length:].reshape(1, -1, 1) current_open_price = all_open_prices[min(168 + i, len(all_open_prices) - 1)] else: context_data = all_prices[start_idx:end_idx].reshape(1, -1, 1) current_open_price = all_open_prices[end_idx] if end_idx < len(all_open_prices) else all_open_prices[-1] context_tensor = torch.FloatTensor(context_data).to(device) with torch.no_grad(): output = ttm_model(context_tensor) pred_price = output.prediction_outputs.cpu().numpy()[0, 0, 0] ttm_preds_list.append(1 if pred_price > current_open_price else 0) ttm_preds = np.array(ttm_preds_list) print(f"✓ 预测完成") # ========== 构建结果DataFrame ========== # 获取对应的时间戳（第i个样本对应df_2026的第168+i根K线） timestamps = df_2026['timestamp'].iloc[168:168+min_samples].values # 转换为美国东部时间 (ET) timestamps_utc = pd.to_datetime(timestamps) timestamps_et = timestamps_utc.tz_localize('UTC').tz_convert('America/New_York') dates = timestamps_et.date # 双模型交集 two_model_agree_up = (dlinear_preds == 1) & (nbeats_preds == 1) two_model_agree_down = (dlinear_preds == 0) & (nbeats_preds == 0) two_model_preds = np.full(len(dlinear_preds), -1) two_model_preds[two_model_agree_up] = 1 two_model_preds[two_model_agree_down] = 0 # 三模型多数投票 vote_sum = dlinear_preds + nbeats_preds + ttm_preds three_model_preds = np.where(vote_sum >= 2, 1, 0) # 真实标签 true_labels = y_dlinear # 创建DataFrame results_df = pd.DataFrame({ 'date': dates, 'timestamp': timestamps, 'timestamp_et': timestamps_et, 'true_label': true_labels, 'dlinear_pred': dlinear_preds, 'nbeats_pred': nbeats_preds, 'ttm_pred': ttm_preds, 'two_model_pred': two_model_preds, 'three_model_pred': three_model_preds, 'two_model_correct': (two_model_preds == true_labels) & (two_model_preds != -1), 'three_model_correct': (three_model_preds == true_labels) }) # 添加周几信息（美国东部时间） results_df['day_of_week'] = pd.to_datetime(results_df['timestamp_et']).dt.day_name() results_df['is_weekend'] = pd.to_datetime(results_df['timestamp_et']).dt.dayofweek >= 5 # 5=Saturday, 6=Sunday # 2026年1月美国重要经济数据发布日期（美国东部时间） # 数据来源：美联储、劳工部等官方日历 us_economic_events = { '2026-01-08': ['初请失业金人数'], '2026-01-09': ['非农就业报告 (NFP)', '失业率'], '2026-01-13': ['CPI通胀数据'], '2026-01-14': ['PPI生产者物价指数'], '2026-01-15': ['零售销售数据'], '2026-01-16': ['密歇根大学消费者信心指数'], '2026-01-22': ['初请失业金人数'], '2026-01-27': ['耐用品订单'], '2026-01-28': ['FOMC利率决议', 'GDP初值'], '2026-01-29': ['PCE物价指数', '个人收入与支出'], } # ========== 按天统计 ========== print("\n" + "="*80) print("📊 每日准确率统计 (美国东部时间)") print("="*80) daily_stats = results_df.groupby('date').agg({ 'true_label': 'count', # 总样本数 'two_model_pred': lambda x: (x != -1).sum(), # 双模型有效预测数 'two_model_correct': 'sum', # 双模型正确数 'three_model_correct': 'sum', # 三模型正确数 'is_weekend': 'first', # 是否周末 'day_of_week': 'first' # 星期几 }) daily_stats.columns = ['total_samples', 'two_model_valid', 'two_model_correct', 'three_model_correct', 'is_weekend', 'day_of_week'] # 计算准确率 daily_stats['two_model_accuracy'] = daily_stats['two_model_correct'] / daily_stats['two_model_valid'] * 100 daily_stats['three_model_accuracy'] = daily_stats['three_model_correct'] / daily_stats['total_samples'] * 100 daily_stats['two_model_coverage'] = daily_stats['two_model_valid'] / daily_stats['total_samples'] * 100 # 填充NaN（当天没有有效预测） daily_stats['two_model_accuracy'] = daily_stats['two_model_accuracy'].fillna(0) # 添加美国经济数据发布标记 daily_stats['us_events'] = daily_stats.index.map( lambda x: us_economic_events.get(str(x), []) ) daily_stats['has_us_data'] = daily_stats['us_events'].apply(lambda x: len(x) > 0) print(f"\n每日详细统计：") print(f"{'日期':<12} {'星期':<10} {'周末':<4} {'美国数据':<4} {'样本':>4} " f"{'双模型%':>8} {'覆盖%':>6} {'三模型%':>8} {'差异':>7}") print("-" * 95) for date, row in daily_stats.iterrows(): diff = row['two_model_accuracy'] - row['three_model_accuracy'] diff_str = f"{diff:+.1f}" if row['two_model_valid'] > 0 else "N/A" weekend_mark = "✓" if row['is_weekend'] else "" us_data_mark = "📊" if row['has_us_data'] else "" # 根据准确率设置颜色标记 two_acc = row['two_model_accuracy'] three_acc = row['three_model_accuracy'] two_mark = "🔥" if two_acc >= 70 else ("✅" if two_acc >= 60 else ("⚠️" if two_acc < 40 else "")) three_mark = "🔥" if three_acc >= 70 else ("✅" if three_acc >= 60 else ("⚠️" if three_acc < 40 else "")) print(f"{date} {row['day_of_week'][:3]:>8} {weekend_mark:>4} {us_data_mark:>4} " f"{row['total_samples']:>4.0f} " f"{row['two_model_accuracy']:>7.1f}{two_mark} " f"{row['two_model_coverage']:>5.1f} " f"{row['three_model_accuracy']:>7.1f}{three_mark} " f"{diff_str:>7}") # 如果有美国经济数据发布，显示详情 if row['has_us_data']: events_str = ", ".join(row['us_events']) print(f" └─ 📊 {events_str}") # ========== 周末 vs 工作日分析 ========== print("\n" + "="*80) print("📅 周末 vs 工作日准确率对比") print("="*80) weekend_stats = daily_stats[daily_stats['is_weekend']] weekday_stats = daily_stats[~daily_stats['is_weekend']] if len(weekend_stats) > 0: print(f"\n🌴 周末 ({len(weekend_stats)}天):") print(f" 双模型平均准确率: {weekend_stats['two_model_accuracy'].mean():.2f}%") print(f" 三模型平均准确率: {weekend_stats['three_model_accuracy'].mean():.2f}%") print(f" 平均覆盖率: {weekend_stats['two_model_coverage'].mean():.2f}%") if len(weekday_stats) > 0: print(f"\n💼 工作日 ({len(weekday_stats)}天):") print(f" 双模型平均准确率: {weekday_stats['two_model_accuracy'].mean():.2f}%") print(f" 三模型平均准确率: {weekday_stats['three_model_accuracy'].mean():.2f}%") print(f" 平均覆盖率: {weekday_stats['two_model_coverage'].mean():.2f}%") # ========== 美国数据发布日 vs 普通日分析 ========== print("\n" + "="*80) print("📊 美国经济数据发布日 vs 普通日准确率对比") print("="*80) us_data_days = daily_stats[daily_stats['has_us_data']] normal_days = daily_stats[~daily_stats['has_us_data']] if len(us_data_days) > 0: print(f"\n📊 美国数据发布日 ({len(us_data_days)}天):") print(f" 双模型平均准确率: {us_data_days['two_model_accuracy'].mean():.2f}%") print(f" 三模型平均准确率: {us_data_days['three_model_accuracy'].mean():.2f}%") print(f" 平均覆盖率: {us_data_days['two_model_coverage'].mean():.2f}%") print(f"\n 具体日期:") for date in us_data_days.index: events = ", ".join(us_data_days.loc[date, 'us_events']) two_acc = us_data_days.loc[date, 'two_model_accuracy'] three_acc = us_data_days.loc[date, 'three_model_accuracy'] print(f" {date}: 双{two_acc:.1f}% / 三{three_acc:.1f}% - {events}") if len(normal_days) > 0: print(f"\n📅 普通日 ({len(normal_days)}天):") print(f" 双模型平均准确率: {normal_days['two_model_accuracy'].mean():.2f}%") print(f" 三模型平均准确率: {normal_days['three_model_accuracy'].mean():.2f}%") print(f" 平均覆盖率: {normal_days['two_model_coverage'].mean():.2f}%") # ========== 分类统计 ========== print("\n" + "="*80) print("📈 准确率分类统计") print("="*80) # 双模型交集分类 two_model_high = daily_stats[daily_stats['two_model_accuracy'] >= 60] two_model_medium = daily_stats[(daily_stats['two_model_accuracy'] >= 45) & (daily_stats['two_model_accuracy'] < 60)] two_model_low = daily_stats[(daily_stats['two_model_accuracy'] > 0) & (daily_stats['two_model_accuracy'] < 45)] # 三模型投票分类 three_model_high = daily_stats[daily_stats['three_model_accuracy'] >= 60] three_model_medium = daily_stats[(daily_stats['three_model_accuracy'] >= 45) & (daily_stats['three_model_accuracy'] < 60)] three_model_low = daily_stats[daily_stats['three_model_accuracy'] < 45] print(f"\n🔵 双模型交集：") print(f" 高准确率 (>=60%): {len(two_model_high)}天") if len(two_model_high) > 0: print(f" 日期: {', '.join([str(d) for d in two_model_high.index])}") print(f" 平均准确率: {two_model_high['two_model_accuracy'].mean():.1f}%") print(f"\n 中等准确率 (45-60%): {len(two_model_medium)}天") if len(two_model_medium) > 0: print(f" 日期: {', '.join([str(d) for d in two_model_medium.index])}") print(f" 平均准确率: {two_model_medium['two_model_accuracy'].mean():.1f}%") print(f"\n 低准确率 (<45%): {len(two_model_low)}天") if len(two_model_low) > 0: print(f" 日期: {', '.join([str(d) for d in two_model_low.index])}") print(f" 平均准确率: {two_model_low['two_model_accuracy'].mean():.1f}%") print(f"\n🟢 三模型投票：") print(f" 高准确率 (>=60%): {len(three_model_high)}天") if len(three_model_high) > 0: print(f" 日期: {', '.join([str(d) for d in three_model_high.index])}") print(f" 平均准确率: {three_model_high['three_model_accuracy'].mean():.1f}%") print(f"\n 中等准确率 (45-60%): {len(three_model_medium)}天") if len(three_model_medium) > 0: print(f" 日期: {', '.join([str(d) for d in three_model_medium.index])}") print(f" 平均准确率: {three_model_medium['three_model_accuracy'].mean():.1f}%") print(f"\n 低准确率 (<45%): {len(three_model_low)}天") if len(three_model_low) > 0: print(f" 日期: {', '.join([str(d) for d in three_model_low.index])}") print(f" 平均准确率: {three_model_low['three_model_accuracy'].mean():.1f}%") # ========== 总结 ========== print("\n" + "="*80) print("📊 总体统计") print("="*80) print(f"\n双模型交集：") print(f" 总体准确率: {(daily_stats['two_model_correct'].sum() / daily_stats['two_model_valid'].sum() * 100):.2f}%") print(f" 总体覆盖率: {(daily_stats['two_model_valid'].sum() / daily_stats['total_samples'].sum() * 100):.2f}%") print(f" 最高日准确率: {daily_stats['two_model_accuracy'].max():.1f}% ({daily_stats['two_model_accuracy'].idxmax()})") print(f" 最低日准确率: {daily_stats[daily_stats['two_model_accuracy'] > 0]['two_model_accuracy'].min():.1f}% ({daily_stats[daily_stats['two_model_accuracy'] > 0]['two_model_accuracy'].idxmin()})") print(f"\n三模型投票：") print(f" 总体准确率: {(daily_stats['three_model_correct'].sum() / daily_stats['total_samples'].sum() * 100):.2f}%") print(f" 覆盖率: 100%") print(f" 最高日准确率: {daily_stats['three_model_accuracy'].max():.1f}% ({daily_stats['three_model_accuracy'].idxmax()})") print(f" 最低日准确率: {daily_stats['three_model_accuracy'].min():.1f}% ({daily_stats['three_model_accuracy'].idxmin()})") print("\n" + "="*80) print("✓ 每日准确率统计完成!") print("="*80) if __name__ == "__main__": main()