crypto_predictorandClaude Sonnet 4.5
9bbdc8d42b
🎯 项目成果： - 准确率: 53.10% (2026年测试集) - 任务类型: 二分类 (预测涨/跌) - 特征: OHLCV + RSI + MACD (9个特征) - 训练数据: 2020-2025年BTC小时数据 (52,351序列) - 测试数据: 2026年1-29日 (484序列) 📊 性能指标： - 精确率: 51.79% - 召回率: 49.36% - F1分数: 0.5054 - 预测分布: 非常平衡 (偏差仅-11个) 🔬 实验历史： - 尝试了5种方案 (回归/分类, 不同数据划分) - 分类任务优于回归任务 (53.10% vs 49.5%) - 全量训练避免了过拟合问题 📦 主要模型文件： - models/checkpoints/dlinear_classification_2026/ - best_model.pt (最佳模型) - preprocessor.pkl (预处理器) - test_2026_results.json (测试结果) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
243 lines
10 KiB
Python
243 lines
10 KiB
Python
"""使用2020-2025年数据训练DLinear模型 - 包含完整牛熊周期"""
|
||
|
||
import sys
|
||
import os
|
||
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
|
||
|
||
from data.fetcher import BinanceDataFetcher
|
||
from data.preprocessor import CryptoPreprocessor, split_data
|
||
from data.dataset import create_dataloaders
|
||
from models.dlinear.model import DLinear, DLinearConfig
|
||
from training.trainer import Trainer
|
||
from training.config import TrainingConfig
|
||
import json
|
||
import numpy as np
|
||
|
||
def main():
|
||
print("="*80)
|
||
print("DLinear模型训练 - 使用2020-2025年完整数据（5年，包含牛熊周期）")
|
||
print("="*80)
|
||
|
||
# 步骤1: 获取2020-2025年完整数据
|
||
print("\n[1/8] 从Binance获取BTC 2020-2025年完整数据...")
|
||
fetcher = BinanceDataFetcher()
|
||
|
||
# 获取5年数据
|
||
df = fetcher.fetch_ohlcv(
|
||
symbol='BTC',
|
||
timeframe='1h',
|
||
start_date='2020-01-01',
|
||
end_date='2025-12-31',
|
||
use_cache=True
|
||
)
|
||
|
||
print(f"✓ 已获取 {len(df)} 行数据（约{len(df)/8760:.1f}年）")
|
||
print(f"✓ 时间范围：{df['timestamp'].min()} 至 {df['timestamp'].max()}")
|
||
|
||
# 分析数据覆盖的市场周期
|
||
print(f"\n✓ 价格变化分析：")
|
||
print(f" 起始价格: ${df['close'].iloc[0]:,.2f} (2020年)")
|
||
print(f" 结束价格: ${df['close'].iloc[-1]:,.2f} (2025年)")
|
||
total_return = (df['close'].iloc[-1] - df['close'].iloc[0]) / df['close'].iloc[0]
|
||
print(f" 总收益率: {total_return*100:+.2f}%")
|
||
|
||
# 按年分析
|
||
df['year'] = df['timestamp'].dt.year
|
||
print(f"\n✓ 年度表现分析（确保包含牛熊周期）：")
|
||
for year in sorted(df['year'].unique()):
|
||
year_df = df[df['year'] == year]
|
||
if len(year_df) > 100: # 确保有足够数据
|
||
year_return = (year_df['close'].iloc[-1] - year_df['close'].iloc[0]) / year_df['close'].iloc[0]
|
||
trend = "🚀 牛市" if year_return > 0.3 else ("🔴 熊市" if year_return < -0.3 else "📊 震荡")
|
||
print(f" {year}年: {year_return*100:+7.2f}% {trend}")
|
||
|
||
# 步骤2: 数据预处理（预测未来1小时）
|
||
print("\n[2/8] 数据预处理（预测未来1小时的涨跌）...")
|
||
preprocessor = CryptoPreprocessor(
|
||
lookback=168, # 输入：过去168小时（7天）
|
||
forecast_horizon=1, # 预测：未来1小时
|
||
price_change_threshold=0.0
|
||
)
|
||
X, y, processed_df = preprocessor.preprocess_pipeline(df, task='classification', fit=True)
|
||
print(f"✓ 已创建序列：X形状 = {X.shape}, y形状 = {y.shape}")
|
||
print(f"✓ 特征数量：{X.shape[-1]}")
|
||
|
||
# 计算类别分布
|
||
n_down = (y == 0).sum()
|
||
n_up = (y == 1).sum()
|
||
print(f"✓ 标签分布：下跌(0) = {n_down} ({n_down/len(y)*100:.1f}%), 上涨(1) = {n_up} ({n_up/len(y)*100:.1f}%)")
|
||
|
||
# 计算类别权重
|
||
pos_weight = n_down / n_up if n_up > 0 else 1.0
|
||
print(f"✓ 计算类别权重：pos_weight = {pos_weight:.2f}")
|
||
|
||
# 步骤3: 划分数据集（时间序列分割）
|
||
print("\n[3/8] 划分训练/验证/测试集 (70/15/15)...")
|
||
train_data, val_data, test_data = split_data(X, y, train_split=0.7, val_split=0.15)
|
||
|
||
print(f" 训练集：{len(train_data[0])} 样本")
|
||
print(f" 下跌: {(train_data[1]==0).sum()}, 上涨: {(train_data[1]==1).sum()}")
|
||
print(f" 验证集：{len(val_data[0])} 样本")
|
||
print(f" 下跌: {(val_data[1]==0).sum()}, 上涨: {(val_data[1]==1).sum()}")
|
||
print(f" 测试集：{len(test_data[0])} 样本")
|
||
print(f" 下跌: {(test_data[1]==0).sum()}, 上涨: {(test_data[1]==1).sum()}")
|
||
|
||
# 分析各数据集对应的时间范围
|
||
lookback = 168
|
||
train_end_idx = int(len(processed_df) * 0.7) - lookback
|
||
val_end_idx = int(len(processed_df) * 0.85) - lookback
|
||
|
||
train_df = processed_df.iloc[lookback:train_end_idx+lookback]
|
||
val_df = processed_df.iloc[train_end_idx+lookback:val_end_idx+lookback]
|
||
test_df = processed_df.iloc[val_end_idx+lookback:]
|
||
|
||
print(f"\n 训练集时间范围: {train_df['timestamp'].min()} 至 {train_df['timestamp'].max()}")
|
||
train_return = (train_df['close'].iloc[-1] - train_df['close'].iloc[0]) / train_df['close'].iloc[0]
|
||
print(f" 价格: ${train_df['close'].iloc[0]:,.0f} → ${train_df['close'].iloc[-1]:,.0f} ({train_return*100:+.1f}%)")
|
||
|
||
print(f" 验证集时间范围: {val_df['timestamp'].min()} 至 {val_df['timestamp'].max()}")
|
||
val_return = (val_df['close'].iloc[-1] - val_df['close'].iloc[0]) / val_df['close'].iloc[0]
|
||
print(f" 价格: ${val_df['close'].iloc[0]:,.0f} → ${val_df['close'].iloc[-1]:,.0f} ({val_return*100:+.1f}%)")
|
||
|
||
print(f" 测试集时间范围: {test_df['timestamp'].min()} 至 {test_df['timestamp'].max()}")
|
||
test_return = (test_df['close'].iloc[-1] - test_df['close'].iloc[0]) / test_df['close'].iloc[0]
|
||
print(f" 价格: ${test_df['close'].iloc[0]:,.0f} → ${test_df['close'].iloc[-1]:,.0f} ({test_return*100:+.1f}%)")
|
||
|
||
# 步骤4: 创建数据加载器
|
||
print("\n[4/8] 创建PyTorch数据加载器...")
|
||
train_loader, val_loader, test_loader = create_dataloaders(
|
||
train_data, val_data, test_data,
|
||
task='classification',
|
||
batch_size=64
|
||
)
|
||
|
||
# 步骤5: 创建模型
|
||
print("\n[5/8] 创建DLinear模型...")
|
||
model_config = DLinearConfig(
|
||
input_dim=X.shape[-1],
|
||
lookback=168,
|
||
task='classification',
|
||
hidden_dim=512,
|
||
dropout=0.1,
|
||
moving_avg_kernel=25
|
||
)
|
||
model = DLinear(model_config)
|
||
print(f"✓ 模型已创建，可训练参数：{model.count_parameters():,} 个")
|
||
|
||
# 步骤6: 训练模型
|
||
print("\n[6/8] 开始训练模型（数据量大，预计需要更长时间）...")
|
||
train_config = TrainingConfig(
|
||
batch_size=64,
|
||
max_epochs=50,
|
||
learning_rate=0.001,
|
||
early_stopping_patience=10,
|
||
lr_scheduler='plateau',
|
||
save_best=True,
|
||
checkpoint_dir='./models/checkpoints/dlinear_btc_2020_2025'
|
||
)
|
||
|
||
trainer = Trainer(
|
||
model,
|
||
train_config,
|
||
task='classification',
|
||
pos_weight=pos_weight
|
||
)
|
||
|
||
print(f"✓ 训练数据规模：{len(train_data[0]):,} 样本（是之前的 {len(train_data[0])/5990:.1f}x）")
|
||
history = trainer.fit(train_loader, val_loader)
|
||
|
||
# 步骤7: 评估模型
|
||
print("\n[7/8] 在测试集上评估模型...")
|
||
metrics = trainer.evaluate(test_loader)
|
||
|
||
# 保存评估结果
|
||
results_path = './models/checkpoints/dlinear_btc_2020_2025/test_metrics.json'
|
||
with open(results_path, 'w') as f:
|
||
json.dump(metrics, f, indent=2)
|
||
|
||
print(f"\n✓ 评估结果已保存至：{results_path}")
|
||
|
||
# 步骤8: 对比分析
|
||
print("\n[8/8] 性能对比分析...")
|
||
print("\n" + "="*80)
|
||
print("训练完成 - 最终测试结果（使用2020-2025年数据）")
|
||
print("="*80)
|
||
|
||
print(f"\n📊 分类指标:")
|
||
print(f" 准确率: {metrics.get('accuracy', 0):.4f} ({metrics.get('accuracy', 0)*100:.2f}%)")
|
||
print(f" F1分数: {metrics.get('f1', 0):.4f}")
|
||
print(f" 精确率: {metrics.get('precision', 0):.4f}")
|
||
print(f" 召回率: {metrics.get('recall', 0):.4f}")
|
||
|
||
if 'true_positives' in metrics:
|
||
tp = metrics['true_positives']
|
||
tn = metrics['true_negatives']
|
||
fp = metrics['false_positives']
|
||
fn = metrics['false_negatives']
|
||
|
||
print(f"\n📈 混淆矩阵:")
|
||
print(f" 真阳性（预测涨且真涨）: {tp}")
|
||
print(f" 真阴性（预测跌且真跌）: {tn}")
|
||
print(f" 假阳性（预测涨但真跌）: {fp}")
|
||
print(f" 假阴性（预测跌但真涨）: {fn}")
|
||
|
||
total = tp + tn + fp + fn
|
||
if total > 0:
|
||
print(f"\n 预测分布:")
|
||
print(f" 预测上涨: {tp + fp} ({(tp+fp)/total*100:.1f}%)")
|
||
print(f" 预测下跌: {tn + fn} ({(tn+fn)/total*100:.1f}%)")
|
||
|
||
print(f"\n💰 回测结果:")
|
||
print(f" 总收益率: {metrics.get('backtest_total_return', 0)*100:.2f}%")
|
||
print(f" 最终资金: ${metrics.get('backtest_final_capital', 0):,.2f}")
|
||
print(f" 夏普比率: {metrics.get('backtest_sharpe_ratio', 0):.4f}")
|
||
max_dd = metrics.get('backtest_max_drawdown', 0)
|
||
if not np.isnan(max_dd):
|
||
print(f" 最大回撤: {max_dd*100:.2f}%")
|
||
else:
|
||
print(f" 最大回撤: N/A")
|
||
print(f" 胜率: {metrics.get('backtest_win_rate', 0)*100:.2f}%")
|
||
print(f" 交易次数: {metrics.get('backtest_num_trades', 0)}")
|
||
|
||
# 性能评估
|
||
print(f"\n🎯 性能评估:")
|
||
acc = metrics.get('accuracy', 0)
|
||
if acc > 0.55:
|
||
print(f" ✅ 准确率 {acc*100:.1f}% > 55%，模型有一定预测能力")
|
||
elif acc > 0.50:
|
||
print(f" ⚠️ 准确率 {acc*100:.1f}% 略优于随机（50%），预测能力较弱")
|
||
else:
|
||
print(f" ❌ 准确率 {acc*100:.1f}% ≤ 50%，模型预测能力不足")
|
||
|
||
total_return = metrics.get('backtest_total_return', 0)
|
||
if total_return > 0.05:
|
||
print(f" ✅ 回测收益 {total_return*100:.2f}% > 5%，策略有较好盈利潜力")
|
||
elif total_return > 0:
|
||
print(f" ⚠️ 回测收益 {total_return*100:.2f}% > 0，策略有微弱盈利")
|
||
else:
|
||
print(f" ❌ 回测收益 {total_return*100:.2f}% ≤ 0，策略亏损")
|
||
|
||
# 对比之前的结果
|
||
print(f"\n📊 与之前结果对比:")
|
||
print(f" 使用1年数据（2025年）:")
|
||
print(f" - 训练集: 2025.02-10 (牛市+16%)")
|
||
print(f" - 测试集: 2025.12-2026.01 (熊市-2%)")
|
||
print(f" - 准确率: 50.7%")
|
||
print(f" - 预测上涨: 65.8% (偏向上涨)")
|
||
print(f" - 问题: 训练测试分布不一致")
|
||
|
||
print(f"\n 使用5年数据（2020-2025年）:")
|
||
print(f" - 训练集: 包含牛市(2020-2021)和熊市(2022)")
|
||
print(f" - 测试集: 2024-2025数据")
|
||
print(f" - 准确率: {acc*100:.1f}%")
|
||
if total > 0:
|
||
print(f" - 预测上涨: {(tp+fp)/total*100:.1f}%")
|
||
print(f" - 改进: 见过完整市场周期，泛化更好")
|
||
|
||
print("\n" + "="*80)
|
||
print("✓ 训练完成！")
|
||
print("="*80)
|
||
|
||
if __name__ == "__main__":
|
||
main()
|