crypto_predictor
1bd9d25f60
- 删除测试脚本: test_2026_classification.py, test_lagllama.py, test_timesfm.py, test_ttm.py - 删除2019相关文件: analyze_2019_*, apply_2019_*, dlinear_2019_*, three_models_2019_* - 添加集成模型相关代码和分析脚本 - 添加N-BEATS模型和CatBoost分类模型
396 lines
16 KiB
Python
396 lines
16 KiB
Python
"""四模型集成每日准确率分析 - 多时间段测试（2019年1月、2020年1月、2026年1月）"""
|
|
|
|
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 predict_period(period_name, start_date, end_date, dlinear_model, nbeats_model, ttm_model, timesfm_model,
|
|
dlinear_preprocessor, nbeats_preprocessor, device):
|
|
"""预测单个时间段的数据"""
|
|
|
|
print(f"\n{'='*80}")
|
|
print(f"📅 {period_name} ({start_date} 至 {end_date})")
|
|
print(f"{'='*80}")
|
|
|
|
# 获取数据
|
|
fetcher = BinanceDataFetcher()
|
|
df = fetcher.fetch_ohlcv('BTC', '1h', start_date=start_date, end_date=end_date, use_cache=False)
|
|
|
|
print(f"✓ 数据: {len(df)} 行")
|
|
|
|
# 预处理
|
|
X_dlinear, y_dlinear, df_dlinear = dlinear_preprocessor.preprocess(df, fit=False)
|
|
X_nbeats, y_nbeats, df_nbeats = nbeats_preprocessor.preprocess(df, 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]
|
|
|
|
print(f"✓ 样本数: {min_samples}")
|
|
|
|
# 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['close'].values
|
|
all_open_prices = df['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)
|
|
|
|
# TimesFM预测
|
|
timesfm_preds_list = []
|
|
|
|
for i in range(min_samples):
|
|
start_idx = i
|
|
end_idx = start_idx + 512
|
|
|
|
if end_idx > len(all_prices):
|
|
context_prices = all_prices[-512:]
|
|
current_open_price = all_open_prices[min(168 + i, len(all_open_prices) - 1)]
|
|
else:
|
|
context_prices = all_prices[start_idx:end_idx]
|
|
current_open_price = all_open_prices[end_idx] if end_idx < len(all_open_prices) else all_open_prices[-1]
|
|
|
|
try:
|
|
point_forecast, _ = timesfm_model.forecast(horizon=1, inputs=[context_prices])
|
|
pred_price = point_forecast[0, 0]
|
|
timesfm_preds_list.append(1 if pred_price > current_open_price else 0)
|
|
except:
|
|
timesfm_preds_list.append(np.random.randint(0, 2))
|
|
|
|
timesfm_preds = np.array(timesfm_preds_list)
|
|
|
|
# 集成策略
|
|
true_labels = y_dlinear
|
|
|
|
# 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
|
|
|
|
# 三模型投票
|
|
vote_sum_3 = dlinear_preds + nbeats_preds + ttm_preds
|
|
three_model_preds = np.where(vote_sum_3 >= 2, 1, 0)
|
|
|
|
# 四模型投票
|
|
vote_sum_4 = dlinear_preds + nbeats_preds + ttm_preds + timesfm_preds
|
|
four_model_preds = np.where(vote_sum_4 >= 3, 1, 0)
|
|
|
|
# 构建每日结果DataFrame
|
|
timestamps = df['timestamp'].iloc[168:168+min_samples].values
|
|
timestamps_utc = pd.to_datetime(timestamps)
|
|
timestamps_et = timestamps_utc.tz_localize('UTC').tz_convert('America/New_York')
|
|
dates = timestamps_et.date
|
|
|
|
results_df = pd.DataFrame({
|
|
'date': dates,
|
|
'timestamp_et': timestamps_et,
|
|
'true_label': true_labels,
|
|
'dlinear_pred': dlinear_preds,
|
|
'nbeats_pred': nbeats_preds,
|
|
'ttm_pred': ttm_preds,
|
|
'timesfm_pred': timesfm_preds,
|
|
'two_model_pred': two_model_preds,
|
|
'three_model_pred': three_model_preds,
|
|
'four_model_pred': four_model_preds,
|
|
})
|
|
|
|
# 按天统计
|
|
daily_stats = results_df.groupby('date').agg({
|
|
'true_label': 'count',
|
|
'dlinear_pred': lambda x: (x == results_df.loc[x.index, 'true_label']).sum(),
|
|
'nbeats_pred': lambda x: (x == results_df.loc[x.index, 'true_label']).sum(),
|
|
'ttm_pred': lambda x: (x == results_df.loc[x.index, 'true_label']).sum(),
|
|
'timesfm_pred': lambda x: (x == results_df.loc[x.index, 'true_label']).sum(),
|
|
'two_model_pred': lambda x: ((x == results_df.loc[x.index, 'true_label']) & (x != -1)).sum(),
|
|
'three_model_pred': lambda x: (x == results_df.loc[x.index, 'true_label']).sum(),
|
|
'four_model_pred': lambda x: (x == results_df.loc[x.index, 'true_label']).sum(),
|
|
})
|
|
|
|
daily_stats.columns = ['total', 'dlinear_correct', 'nbeats_correct', 'ttm_correct',
|
|
'timesfm_correct', 'two_model_correct', 'three_model_correct', 'four_model_correct']
|
|
|
|
# 计算每天的有效预测数（双模型）
|
|
two_model_valid = results_df.groupby('date')['two_model_pred'].apply(lambda x: (x != -1).sum())
|
|
daily_stats['two_model_valid'] = two_model_valid
|
|
|
|
# 计算准确率
|
|
daily_stats['dlinear_acc'] = daily_stats['dlinear_correct'] / daily_stats['total'] * 100
|
|
daily_stats['nbeats_acc'] = daily_stats['nbeats_correct'] / daily_stats['total'] * 100
|
|
daily_stats['ttm_acc'] = daily_stats['ttm_correct'] / daily_stats['total'] * 100
|
|
daily_stats['timesfm_acc'] = daily_stats['timesfm_correct'] / daily_stats['total'] * 100
|
|
daily_stats['two_model_acc'] = daily_stats['two_model_correct'] / daily_stats['two_model_valid'] * 100
|
|
daily_stats['three_model_acc'] = daily_stats['three_model_correct'] / daily_stats['total'] * 100
|
|
daily_stats['four_model_acc'] = daily_stats['four_model_correct'] / daily_stats['total'] * 100
|
|
|
|
daily_stats['two_model_acc'] = daily_stats['two_model_acc'].fillna(0)
|
|
|
|
return daily_stats, results_df
|
|
|
|
|
|
def main():
|
|
print("="*80)
|
|
print("四模型集成每日准确率分析 - 多时间段测试")
|
|
print("测试时段: 2019年1月、2020年1月、2026年1月")
|
|
print("="*80)
|
|
|
|
# ========== 加载模型 ==========
|
|
print("\n[1/2] 加载模型...")
|
|
|
|
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
|
|
|
|
from models.dlinear.model import DLinear
|
|
dlinear_checkpoint = torch.load(
|
|
'./models/checkpoints/dlinear_classification_2026/best_model.pt',
|
|
map_location=device, weights_only=False
|
|
)
|
|
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)
|
|
|
|
from models.nbeats.model import NBeats
|
|
nbeats_checkpoint = torch.load(
|
|
'./models/checkpoints/nbeats_btc_1h/best_model.pt',
|
|
map_location=device, weights_only=False
|
|
)
|
|
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)
|
|
|
|
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()
|
|
|
|
import timesfm
|
|
timesfm_model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
|
|
"google/timesfm-2.5-200m-pytorch",
|
|
torch_compile=False
|
|
)
|
|
timesfm_model.compile(
|
|
timesfm.ForecastConfig(
|
|
max_context=512,
|
|
max_horizon=1,
|
|
normalize_inputs=True,
|
|
use_continuous_quantile_head=True,
|
|
force_flip_invariance=True,
|
|
infer_is_positive=True,
|
|
fix_quantile_crossing=True,
|
|
)
|
|
)
|
|
|
|
print("✓ 四个模型已加载 (DLinear, N-BEATS, TTM-R2, TimesFM)")
|
|
|
|
# ========== 多时间段预测 ==========
|
|
print("\n[2/2] 多时间段预测...")
|
|
|
|
periods = [
|
|
("2019年1月", "2019-01-01", "2019-01-31"),
|
|
("2020年1月", "2020-01-01", "2020-01-31"),
|
|
("2026年1月", "2026-01-01", "2026-01-31"),
|
|
]
|
|
|
|
all_period_stats = {}
|
|
|
|
for period_name, start_date, end_date in periods:
|
|
daily_stats, results_df = predict_period(
|
|
period_name, start_date, end_date,
|
|
dlinear_model, nbeats_model, ttm_model, timesfm_model,
|
|
dlinear_preprocessor, nbeats_preprocessor, device
|
|
)
|
|
all_period_stats[period_name] = daily_stats
|
|
|
|
# 显示每日准确率
|
|
print(f"\n每日准确率明细:")
|
|
print(f"{'日期':<12} {'样本':>4} {'DLinear':>8} {'N-BEATS':>8} {'TTM-R2':>8} {'TimesFM':>8} {'双模型':>8} {'三模型':>8} {'四模型':>8}")
|
|
print("-" * 95)
|
|
|
|
for date, row in daily_stats.iterrows():
|
|
print(f"{date} {row['total']:>4.0f} "
|
|
f"{row['dlinear_acc']:>7.1f}% "
|
|
f"{row['nbeats_acc']:>7.1f}% "
|
|
f"{row['ttm_acc']:>7.1f}% "
|
|
f"{row['timesfm_acc']:>7.1f}% "
|
|
f"{row['two_model_acc']:>7.1f}% "
|
|
f"{row['three_model_acc']:>7.1f}% "
|
|
f"{row['four_model_acc']:>7.1f}%")
|
|
|
|
# 统计总体
|
|
print(f"\n{period_name} 总体统计:")
|
|
print(f" DLinear: {daily_stats['dlinear_correct'].sum() / daily_stats['total'].sum() * 100:.2f}%")
|
|
print(f" N-BEATS: {daily_stats['nbeats_correct'].sum() / daily_stats['total'].sum() * 100:.2f}%")
|
|
print(f" TTM-R2: {daily_stats['ttm_correct'].sum() / daily_stats['total'].sum() * 100:.2f}%")
|
|
print(f" TimesFM: {daily_stats['timesfm_correct'].sum() / daily_stats['total'].sum() * 100:.2f}%")
|
|
print(f" 双模型: {daily_stats['two_model_correct'].sum() / daily_stats['two_model_valid'].sum() * 100:.2f}% (覆盖率: {daily_stats['two_model_valid'].sum() / daily_stats['total'].sum() * 100:.1f}%)")
|
|
print(f" 三模型: {daily_stats['three_model_correct'].sum() / daily_stats['total'].sum() * 100:.2f}%")
|
|
print(f" 四模型: {daily_stats['four_model_correct'].sum() / daily_stats['total'].sum() * 100:.2f}%")
|
|
|
|
# ========== 跨时期对比 ==========
|
|
print("\n" + "="*80)
|
|
print("📊 跨时期对比总结")
|
|
print("="*80)
|
|
|
|
print(f"\n{'时期':<12} {'DLinear':>9} {'N-BEATS':>9} {'TTM-R2':>9} {'TimesFM':>9} {'双模型':>9} {'三模型':>9} {'四模型':>9}")
|
|
print("-" * 95)
|
|
|
|
for period_name in ["2019年1月", "2020年1月", "2026年1月"]:
|
|
stats = all_period_stats[period_name]
|
|
print(f"{period_name:<12} "
|
|
f"{stats['dlinear_correct'].sum() / stats['total'].sum() * 100:>8.2f}% "
|
|
f"{stats['nbeats_correct'].sum() / stats['total'].sum() * 100:>8.2f}% "
|
|
f"{stats['ttm_correct'].sum() / stats['total'].sum() * 100:>8.2f}% "
|
|
f"{stats['timesfm_correct'].sum() / stats['total'].sum() * 100:>8.2f}% "
|
|
f"{stats['two_model_correct'].sum() / stats['two_model_valid'].sum() * 100:>8.2f}% "
|
|
f"{stats['three_model_correct'].sum() / stats['total'].sum() * 100:>8.2f}% "
|
|
f"{stats['four_model_correct'].sum() / stats['total'].sum() * 100:>8.2f}%")
|
|
|
|
print("\n" + "="*80)
|
|
print("✓ 多时间段每日准确率分析完成!")
|
|
print("="*80)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|