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
401 lines
17 KiB
Python
401 lines
17 KiB
Python
"""三模型集成预测 - DLinear + N-BEATS + TTM-R2 交集策略"""
|
||
|
||
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 sklearn.metrics import confusion_matrix
|
||
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("三模型集成预测：DLinear + N-BEATS + TTM-R2 交集策略")
|
||
print("="*80)
|
||
|
||
# ========== 步骤1: 加载三个模型 ==========
|
||
print("\n[1/6] 加载三个模型...")
|
||
|
||
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
|
||
|
||
# 1. 加载DLinear分类模型
|
||
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)
|
||
|
||
print(f"✓ DLinear分类模型已加载 (准确率: 53.10%)")
|
||
|
||
# 2. 加载N-BEATS回归模型
|
||
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)
|
||
|
||
print(f"✓ N-BEATS回归模型已加载 (准确率: 50.83%)")
|
||
|
||
# 3. 加载TTM-R2模型
|
||
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"✓ TTM-R2模型已加载 (准确率: 37.50%)")
|
||
|
||
# ========== 步骤2: 获取2026年测试数据 ==========
|
||
print("\n[2/6] 获取2026年测试数据...")
|
||
fetcher = BinanceDataFetcher()
|
||
|
||
try:
|
||
df_2026 = fetcher.fetch_ohlcv('BTC', '1h', start_date='2026-01-01',
|
||
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")
|
||
print(f"✓ 时间: {df_2026['timestamp'].min()} 至 {df_2026['timestamp'].max()}")
|
||
|
||
# ========== 步骤3: 预处理数据 ==========
|
||
print("\n[3/6] 预处理数据...")
|
||
|
||
# DLinear预处理
|
||
X_dlinear, y_dlinear, df_dlinear = dlinear_preprocessor.preprocess(df_2026, fit=False)
|
||
|
||
# N-BEATS预处理
|
||
X_nbeats, y_nbeats, df_nbeats = nbeats_preprocessor.preprocess(df_2026, fit=False)
|
||
|
||
# 确保样本数量一致
|
||
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]
|
||
|
||
print(f"✓ 对齐样本数: {min_samples}")
|
||
|
||
# ========== 步骤4: 三个模型分别预测 ==========
|
||
print("\n[4/6] 三个模型分别预测...")
|
||
|
||
# 4.1 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)) # sigmoid
|
||
dlinear_preds = (dlinear_probs > 0.5).astype(int).flatten() # 1=涨, 0=跌
|
||
|
||
print(f"✓ DLinear预测完成")
|
||
print(f" 预测上涨: {(dlinear_preds==1).sum()} ({(dlinear_preds==1).sum()/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" 预测下跌: {(dlinear_preds==0).sum()} ({(dlinear_preds==0).sum()/len(dlinear_preds)*100:.1f}%)")
|
||
|
||
# 4.2 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()
|
||
|
||
# 还原真实价格
|
||
nbeats_pred_prices = nbeats_preprocessor.inverse_transform_price(nbeats_pred_scaled)
|
||
|
||
# 【修改】获取当前K线的open价格（未来可交易价格）用于判断方向
|
||
# 第i个样本对应的是第168+i根K线，其open价格是真正可以交易的价格
|
||
current_open_prices = df_nbeats['open'].iloc[168:168+len(nbeats_pred_prices)].values
|
||
nbeats_preds = (nbeats_pred_prices > current_open_prices).astype(int) # 1=涨, 0=跌
|
||
|
||
print(f"✓ N-BEATS预测完成 (使用当前K线open价格判断)")
|
||
print(f" 预测上涨: {(nbeats_preds==1).sum()} ({(nbeats_preds==1).sum()/len(nbeats_preds)*100:.1f}%)")
|
||
print(f" 预测下跌: {(nbeats_preds==0).sum()} ({(nbeats_preds==0).sum()/len(nbeats_preds)*100:.1f}%)")
|
||
|
||
# 4.3 TTM-R2预测
|
||
print(f"✓ TTM-R2预测中...")
|
||
|
||
# TTM-R2需要特定格式: (batch_size, context_length, num_features)
|
||
context_length = 512
|
||
ttm_preds_list = []
|
||
|
||
# 从df_2026获取价格数据
|
||
all_prices = df_2026['close'].values
|
||
all_open_prices = df_2026['open'].values
|
||
|
||
# 为每个样本创建预测
|
||
for i in range(min_samples):
|
||
# 计算当前样本在原始数据中的位置
|
||
# DLinear的lookback是168，所以第i个样本对应df_2026的第168+i根K线
|
||
start_idx = i
|
||
end_idx = start_idx + context_length
|
||
|
||
if end_idx > len(all_prices):
|
||
# 如果超出范围，使用最后512个价格
|
||
context_data = all_prices[-context_length:].reshape(1, -1, 1)
|
||
# 【修改】使用对应K线的open价格作为当前可交易价格
|
||
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)
|
||
# 【修改】使用第end_idx根K线的open价格（即当前K线的开盘价）
|
||
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)
|
||
# 预测第一个时间步的值
|
||
pred_price = output.prediction_outputs.cpu().numpy()[0, 0, 0]
|
||
|
||
# 【修改】判断方向: 预测价格 vs 当前K线open价格
|
||
ttm_preds_list.append(1 if pred_price > current_open_price else 0)
|
||
|
||
ttm_preds = np.array(ttm_preds_list)
|
||
|
||
print(f"✓ TTM-R2预测完成 (使用当前K线open价格判断)")
|
||
print(f" 预测上涨: {(ttm_preds==1).sum()} ({(ttm_preds==1).sum()/len(ttm_preds)*100:.1f}%)")
|
||
print(f" 预测下跌: {(ttm_preds==0).sum()} ({(ttm_preds==0).sum()/len(ttm_preds)*100:.1f}%)")
|
||
|
||
# ========== 步骤5: 三模型多数投票策略（两方相同） ==========
|
||
print("\n[5/6] 应用三模型多数投票策略（两方相同）...")
|
||
|
||
# 真实标签
|
||
true_labels = y_dlinear
|
||
|
||
# 多数投票：至少两个模型预测上涨 → 上涨，至少两个预测下跌 → 下跌
|
||
vote_sum = dlinear_preds + nbeats_preds + ttm_preds
|
||
|
||
# vote_sum >= 2 表示至少两个模型预测上涨（1）
|
||
# vote_sum <= 1 表示至少两个模型预测下跌（0）
|
||
majority_up = vote_sum >= 2
|
||
majority_down = vote_sum <= 1
|
||
|
||
# 统计各种情况
|
||
all_agree_up = (dlinear_preds == 1) & (nbeats_preds == 1) & (ttm_preds == 1)
|
||
all_agree_down = (dlinear_preds == 0) & (nbeats_preds == 0) & (ttm_preds == 0)
|
||
|
||
# 恰好两方同意（不包括三方都同意）
|
||
exactly_two_up = majority_up & ~all_agree_up
|
||
exactly_two_down = majority_down & ~all_agree_down
|
||
|
||
n_all_agree_up = all_agree_up.sum()
|
||
n_all_agree_down = all_agree_down.sum()
|
||
n_exactly_two_up = exactly_two_up.sum()
|
||
n_exactly_two_down = exactly_two_down.sum()
|
||
|
||
print(f"\n📊 预测投票分析:")
|
||
print(f" 三方都预测上涨: {n_all_agree_up} ({n_all_agree_up/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" 三方都预测下跌: {n_all_agree_down} ({n_all_agree_down/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" 恰好两方预测上涨: {n_exactly_two_up} ({n_exactly_two_up/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" 恰好两方预测下跌: {n_exactly_two_down} ({n_exactly_two_down/len(dlinear_preds)*100:.1f}%)")
|
||
|
||
# 详细显示是哪两方同意
|
||
dl_nb_up = (dlinear_preds == 1) & (nbeats_preds == 1) & (ttm_preds == 0)
|
||
dl_ttm_up = (dlinear_preds == 1) & (nbeats_preds == 0) & (ttm_preds == 1)
|
||
nb_ttm_up = (dlinear_preds == 0) & (nbeats_preds == 1) & (ttm_preds == 1)
|
||
|
||
dl_nb_down = (dlinear_preds == 0) & (nbeats_preds == 0) & (ttm_preds == 1)
|
||
dl_ttm_down = (dlinear_preds == 0) & (nbeats_preds == 1) & (ttm_preds == 0)
|
||
nb_ttm_down = (dlinear_preds == 1) & (nbeats_preds == 0) & (ttm_preds == 0)
|
||
|
||
print(f"\n 详细组合:")
|
||
print(f" DLinear+NBEATS同意看涨: {dl_nb_up.sum()} ({dl_nb_up.sum()/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" DLinear+TTM同意看涨: {dl_ttm_up.sum()} ({dl_ttm_up.sum()/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" NBEATS+TTM同意看涨: {nb_ttm_up.sum()} ({nb_ttm_up.sum()/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" DLinear+NBEATS同意看跌: {dl_nb_down.sum()} ({dl_nb_down.sum()/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" DLinear+TTM同意看跌: {dl_ttm_down.sum()} ({dl_ttm_down.sum()/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" NBEATS+TTM同意看跌: {nb_ttm_down.sum()} ({nb_ttm_down.sum()/len(dlinear_preds)*100:.1f}%)")
|
||
|
||
# 多数投票结果
|
||
ensemble_preds = np.where(majority_up, 1, 0)
|
||
|
||
print(f"\n📊 三模型多数投票策略结果:")
|
||
print(f" 最终预测上涨: {(ensemble_preds==1).sum()} ({(ensemble_preds==1).sum()/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" 最终预测下跌: {(ensemble_preds==0).sum()} ({(ensemble_preds==0).sum()/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" 覆盖率: 100% (多数投票无弃权)")
|
||
|
||
# ========== 步骤6: 性能对比 ==========
|
||
print("\n[6/6] 性能对比分析...")
|
||
|
||
# 计算多数投票准确率
|
||
ensemble_accuracy = (ensemble_preds == true_labels).mean()
|
||
print(f"\n✨ 三模型多数投票准确率: {ensemble_accuracy*100:.2f}%")
|
||
|
||
# 混淆矩阵
|
||
cm = confusion_matrix(true_labels, ensemble_preds)
|
||
tn, fp, fn, tp = cm.ravel()
|
||
|
||
print(f"\n详细分析:")
|
||
print(f" ✓ 预测涨且真涨: {tp} ({tp/len(ensemble_preds)*100:.1f}%)")
|
||
print(f" ✓ 预测跌且真跌: {tn} ({tn/len(ensemble_preds)*100:.1f}%)")
|
||
print(f" ✗ 预测涨但真跌: {fp} ({fp/len(ensemble_preds)*100:.1f}%)")
|
||
print(f" ✗ 预测跌但真涨: {fn} ({fn/len(ensemble_preds)*100:.1f}%)")
|
||
|
||
# 对比分析
|
||
print("\n" + "="*80)
|
||
print("📊 性能对比")
|
||
print("="*80)
|
||
|
||
dlinear_acc = (dlinear_preds == true_labels).mean()
|
||
nbeats_acc = (nbeats_preds == true_labels).mean()
|
||
ttm_acc = (ttm_preds == true_labels).mean()
|
||
|
||
print(f"\n单模型性能:")
|
||
print(f" DLinear分类: {dlinear_acc*100:.2f}%")
|
||
print(f" N-BEATS回归: {nbeats_acc*100:.2f}%")
|
||
print(f" TTM-R2零样本: {ttm_acc*100:.2f}%")
|
||
|
||
# 两模型交集（DLinear + N-BEATS）
|
||
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
|
||
two_model_valid = two_model_preds != -1
|
||
two_model_acc = (two_model_preds[two_model_valid] == true_labels[two_model_valid]).mean()
|
||
|
||
print(f"\n集成策略性能:")
|
||
print(f" 双模型交集 (DLinear+NBEATS): {two_model_acc*100:.2f}% (覆盖率: {two_model_valid.sum()/len(dlinear_preds)*100:.1f}%)")
|
||
|
||
if len(valid_preds) > 0:
|
||
print(f" 三模型交集 (ALL): {ensemble_accuracy*100:.2f}% (覆盖率: {valid_mask.sum()/len(dlinear_preds)*100:.1f}%)")
|
||
|
||
# 性能评估
|
||
print("\n" + "="*80)
|
||
print("🎯 策略评估")
|
||
print("="*80)
|
||
|
||
best_single = max(dlinear_acc, nbeats_acc, ttm_acc)
|
||
|
||
if ensemble_accuracy > best_single:
|
||
improvement = ensemble_accuracy - best_single
|
||
print(f"\n ✅ 三模型交集优于所有单模型!")
|
||
print(f" 准确率提升: {improvement*100:.2f}个百分点")
|
||
print(f" 代价: 弃权{(~valid_mask).sum()}次预测 ({(~valid_mask).sum()/len(dlinear_preds)*100:.1f}%)")
|
||
elif ensemble_accuracy > two_model_acc:
|
||
improvement = ensemble_accuracy - two_model_acc
|
||
print(f"\n ✅ 三模型交集优于双模型交集!")
|
||
print(f" 准确率提升: {improvement*100:.2f}个百分点")
|
||
print(f" 但覆盖率下降: {(two_model_valid.sum() - valid_mask.sum())/len(dlinear_preds)*100:.1f}%")
|
||
else:
|
||
print(f"\n ⚠️ 三模型交集未能提升准确率")
|
||
print(f" 可能原因: TTM-R2表现太差({ttm_acc*100:.2f}%)，拖累了整体性能")
|
||
print(f" 建议: 使用双模型交集(DLinear+NBEATS)")
|
||
|
||
# 实际收益分析
|
||
print(f"\n💰 实际收益分析:")
|
||
print(f" 如果只在三个模型都同意时交易:")
|
||
print(f" - 交易次数: {valid_mask.sum()} ({valid_mask.sum()/len(dlinear_preds)*100:.1f}%)")
|
||
print(f" - 准确率: {ensemble_accuracy*100:.2f}%")
|
||
print(f" - 避免交易: {(~valid_mask).sum()}次 (规避不确定性)")
|
||
|
||
print("\n" + "="*80)
|
||
print("✓ 三模型集成预测分析完成!")
|
||
print("="*80)
|
||
|
||
|
||
if __name__ == "__main__":
|
||
main()
|