Initial commit: BTC 1h prediction model (53.10% accuracy)

🎯 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>
This commit is contained in:
crypto_predictorandClaude Sonnet 4.5 committed 2026-01-29 21:27:15 +08:00
commit 9bbdc8d42b
82 files changed
+8742

No files matched your search

+134
View File
@@ -0,0 +1,134 @@
"""简化版模型诊断(无需matplotlib)"""
import torch
import numpy as np
from models.dlinear.model import DLinear, DLinearConfig
from data.fetcher import BinanceDataFetcher
from data.preprocessor import CryptoPreprocessor, split_data
from data.dataset import create_dataloaders
def diagnose():
print("="*80)
print("🔍 模型预测偏差诊断报告")
print("="*80)
# 加载模型
print("\n[1/4] 加载已训练模型...")
model_path = './models/checkpoints/dlinear_btc_1h_balanced/best_model.pt'
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
checkpoint = torch.load(model_path, map_location=device, weights_only=False)
config = checkpoint['config']
model = DLinear(config)
model.load_state_dict(checkpoint['model_state_dict'])
model = model.to(device)
model.eval()
print(f"✓ 模型配置: task={config.task}, hidden_dim={config.hidden_dim}")
# 准备数据
print("\n[2/4] 加载测试数据...")
fetcher = BinanceDataFetcher()
df = fetcher.fetch_ohlcv('BTC', '1h', use_cache=True)
preprocessor = CryptoPreprocessor(lookback=168, forecast_horizon=24, price_change_threshold=0.0)
X, y, _ = preprocessor.preprocess_pipeline(df, task='classification', fit=True)
_, _, test_data = split_data(X, y, train_split=0.7, val_split=0.15)
_, _, test_loader = create_dataloaders(test_data, test_data, test_data, batch_size=64)
# 收集输出
print("\n[3/4] 分析模型输出...")
all_outputs = []
all_labels = []
with torch.no_grad():
for batch_X, batch_y in test_loader:
batch_X = batch_X.to(device)
outputs = model(batch_X).squeeze()
all_outputs.extend(outputs.cpu().numpy().flatten())
all_labels.extend(batch_y.numpy().flatten())
all_outputs = np.array(all_outputs)
all_labels = np.array(all_labels)
# 统计分析
print("\n" + "="*80)
print("📊 关键发现")
print("="*80)
print(f"\n1️⃣ 模型输出统计(已经过Sigmoid):")
print(f" 最小值: {all_outputs.min():.6f}")
print(f" 最大值: {all_outputs.max():.6f}")
print(f" 均值: {all_outputs.mean():.6f} ← 应该接近0.5")
print(f" 中位数: {np.median(all_outputs):.6f}")
print(f" 标准差: {all_outputs.std():.6f}")
print(f"\n2️⃣ 预测阈值分析:")
for threshold in [0.5, 0.4, 0.3, 0.2, 0.1]:
count = (all_outputs > threshold).sum()
pct = (all_outputs > threshold).mean() * 100
print(f" 输出 > {threshold}: {count:4d} 个 ({pct:5.2f}%)")
print(f"\n3️⃣ 真实标签分布:")
n_down = (all_labels == 0).sum()
n_up = (all_labels == 1).sum()
print(f" 下跌(0): {n_down} ({n_down/len(all_labels)*100:.1f}%)")
print(f" 上涨(1): {n_up} ({n_up/len(all_labels)*100:.1f}%)")
# 验证双重sigmoid问题
print("\n" + "="*80)
print("🔴 问题诊断:双重Sigmoid Bug")
print("="*80)
print("\n📋 当前架构:")
print(" 模型输出层: self.output_activation = nn.Sigmoid() (models/dlinear/model.py:128)")
print(" 损失函数: BCEWithLogitsLoss() (models/base.py:74)")
print("\n⚠️ 问题:")
print(" BCEWithLogitsLoss 内部包含 Sigmoid,期望输入是 logits(未激活)")
print(" 但模型已经输出了 sigmoid(logits)")
print(" → 损失函数再次应用: sigmoid(sigmoid(logits)) ← 双重激活!")
print("\n🧪 双重Sigmoid效果演示:")
test_logits = np.array([-3, -1, 0, 1, 3])
single = 1 / (1 + np.exp(-test_logits))
double = 1 / (1 + np.exp(-single))
print(f" 原始logits: {test_logits}")
print(f" 单次sigmoid: {np.round(single, 4)}")
print(f" 双重sigmoid: {np.round(double, 4)} ← 全部压缩到~0.5")
print("\n💥 后果:")
print(f" - 模型输出均值仅 {all_outputs.mean():.1%}(期望~50%)")
print(f" - {(all_outputs>0.5).sum()} 个样本输出>0.5(共{len(all_outputs)}个)")
print(f" - 模型永远预测类别0(下跌),因为没有输出超过阈值0.5")
print(f" - 梯度消失,模型无法有效学习")
# 解决方案
print("\n" + "="*80)
print("✅ 修复方案")
print("="*80)
print("\n方案1(推荐):移除模型的Sigmoid输出")
print(" 文件: models/dlinear/model.py")
print(" 位置: 第127-130行")
print(" 修改前: if config.task == 'classification':")
print(" self.output_activation = nn.Sigmoid()")
print(" 修改后: if config.task == 'classification':")
print(" self.output_activation = nn.Identity() # 输出logits")
print(" 说明: 保持损失函数为BCEWithLogitsLoss")
print("\n方案2:更换损失函数(不推荐)")
print(" 文件: models/base.py")
print(" 位置: 第70-76行")
print(" 修改: 将 BCEWithLogitsLoss 改为 BCELoss")
print(" 说明: 保持模型输出Sigmoid,但BCELoss数值稳定性较差")
print("\n" + "="*80)
print("📌 结论")
print("="*80)
print("✓ 预测偏差的根本原因是【双重Sigmoid Bug】")
print("✓ 与数据处理、类别平衡、特征工程无关")
print("✓ 修复后需要重新训练模型")
print("="*80)
if __name__ == "__main__":
diagnose()