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
commit
9bbdc8d42b
82 files changed
+8742
No files matched your search
@@ -0,0 +1 @@
|
||||
"""DLinear model implementation."""
|
||||
@@ -0,0 +1,203 @@
|
||||
"""DLinear: Decomposition Linear model for time series forecasting.
|
||||
|
||||
Reference: Are Transformers Effective for Time Series Forecasting? (DLinear)
|
||||
Paper: https://arxiv.org/abs/2205.13504
|
||||
"""
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from dataclasses import dataclass
|
||||
from typing import Literal
|
||||
|
||||
from ..base import BasePredictor, ModelConfig
|
||||
|
||||
|
||||
@dataclass
|
||||
class DLinearConfig(ModelConfig):
|
||||
"""Configuration for DLinear model."""
|
||||
|
||||
moving_avg_kernel: int = 25 # Kernel size for moving average decomposition
|
||||
individual: bool = False # Whether to use individual linear layers per feature
|
||||
|
||||
|
||||
class MovingAvg(nn.Module):
|
||||
"""Moving average block for trend extraction."""
|
||||
|
||||
def __init__(self, kernel_size: int):
|
||||
"""Initialize moving average layer.
|
||||
|
||||
Args:
|
||||
kernel_size: Size of the moving average window
|
||||
"""
|
||||
super().__init__()
|
||||
self.kernel_size = kernel_size
|
||||
self.avg = nn.AvgPool1d(
|
||||
kernel_size=kernel_size,
|
||||
stride=1,
|
||||
padding=kernel_size // 2,
|
||||
count_include_pad=False
|
||||
)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""Apply moving average.
|
||||
|
||||
Args:
|
||||
x: Input tensor of shape (batch_size, seq_len, features)
|
||||
|
||||
Returns:
|
||||
Smoothed tensor
|
||||
"""
|
||||
# Padding at both ends
|
||||
front = x[:, 0:1, :].repeat(1, self.kernel_size // 2, 1)
|
||||
end = x[:, -1:, :].repeat(1, (self.kernel_size - 1) // 2, 1)
|
||||
x_padded = torch.cat([front, x, end], dim=1)
|
||||
|
||||
# Apply average pooling (swap dimensions for AvgPool1d)
|
||||
x_padded = x_padded.permute(0, 2, 1) # (batch, features, seq_len)
|
||||
x_smoothed = self.avg(x_padded)
|
||||
x_smoothed = x_smoothed.permute(0, 2, 1) # (batch, seq_len, features)
|
||||
|
||||
# Trim to original length
|
||||
return x_smoothed[:, :x.size(1), :]
|
||||
|
||||
|
||||
class SeriesDecomposition(nn.Module):
|
||||
"""Series decomposition block to extract trend and seasonal components."""
|
||||
|
||||
def __init__(self, kernel_size: int):
|
||||
"""Initialize decomposition layer.
|
||||
|
||||
Args:
|
||||
kernel_size: Size of the moving average window for trend
|
||||
"""
|
||||
super().__init__()
|
||||
self.moving_avg = MovingAvg(kernel_size)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> tuple:
|
||||
"""Decompose series into trend and seasonal components.
|
||||
|
||||
Args:
|
||||
x: Input tensor of shape (batch_size, seq_len, features)
|
||||
|
||||
Returns:
|
||||
Tuple of (seasonal, trend) tensors
|
||||
"""
|
||||
trend = self.moving_avg(x)
|
||||
seasonal = x - trend
|
||||
return seasonal, trend
|
||||
|
||||
|
||||
class DLinear(BasePredictor):
|
||||
"""DLinear model: Decomposition + Linear layers for forecasting."""
|
||||
|
||||
def __init__(self, config: DLinearConfig):
|
||||
"""Initialize DLinear model.
|
||||
|
||||
Args:
|
||||
config: Model configuration
|
||||
"""
|
||||
super().__init__(config)
|
||||
|
||||
self.decomposition = SeriesDecomposition(config.moving_avg_kernel)
|
||||
|
||||
if config.individual:
|
||||
# Individual linear layer for each feature
|
||||
self.linear_seasonal = nn.ModuleList([
|
||||
nn.Linear(config.lookback, 1)
|
||||
for _ in range(config.input_dim)
|
||||
])
|
||||
self.linear_trend = nn.ModuleList([
|
||||
nn.Linear(config.lookback, 1)
|
||||
for _ in range(config.input_dim)
|
||||
])
|
||||
else:
|
||||
# Shared linear layers
|
||||
self.linear_seasonal = nn.Linear(config.lookback, 1)
|
||||
self.linear_trend = nn.Linear(config.lookback, 1)
|
||||
|
||||
# Final aggregation layer
|
||||
self.aggregation = nn.Sequential(
|
||||
nn.Linear(config.input_dim, config.hidden_dim),
|
||||
nn.ReLU(),
|
||||
nn.Dropout(config.dropout),
|
||||
nn.Linear(config.hidden_dim, 1)
|
||||
)
|
||||
|
||||
# Output activation
|
||||
# 注意：对于分类任务，使用Identity而非Sigmoid
|
||||
# 因为BCEWithLogitsLoss内部已包含Sigmoid，避免双重激活
|
||||
if config.task == 'classification':
|
||||
self.output_activation = nn.Identity() # 输出logits，由损失函数处理
|
||||
else:
|
||||
self.output_activation = nn.Identity()
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""Forward pass through DLinear.
|
||||
|
||||
Args:
|
||||
x: Input tensor of shape (batch_size, lookback, input_dim)
|
||||
|
||||
Returns:
|
||||
Output tensor of shape (batch_size, 1)
|
||||
"""
|
||||
# Decompose into seasonal and trend
|
||||
seasonal, trend = self.decomposition(x)
|
||||
|
||||
# Apply linear layers to each component
|
||||
if self.config.individual:
|
||||
seasonal_out = []
|
||||
trend_out = []
|
||||
for i in range(self.config.input_dim):
|
||||
# Extract feature i: (batch, seq_len)
|
||||
s_i = seasonal[:, :, i]
|
||||
t_i = trend[:, :, i]
|
||||
|
||||
# Apply linear transformation: (batch, seq_len) -> (batch, 1)
|
||||
seasonal_out.append(self.linear_seasonal[i](s_i))
|
||||
trend_out.append(self.linear_trend[i](t_i))
|
||||
|
||||
# Stack outputs: (batch, input_dim)
|
||||
seasonal_out = torch.stack(seasonal_out, dim=1).squeeze(-1)
|
||||
trend_out = torch.stack(trend_out, dim=1).squeeze(-1)
|
||||
else:
|
||||
# Shared linear layer
|
||||
# Swap dimensions: (batch, seq_len, features) -> (batch, features, seq_len)
|
||||
seasonal = seasonal.permute(0, 2, 1)
|
||||
trend = trend.permute(0, 2, 1)
|
||||
|
||||
# Apply linear: (batch, features, seq_len) -> (batch, features, 1)
|
||||
seasonal_out = self.linear_seasonal(seasonal).squeeze(-1)
|
||||
trend_out = self.linear_trend(trend).squeeze(-1)
|
||||
|
||||
# Combine seasonal and trend
|
||||
combined = seasonal_out + trend_out # (batch, input_dim)
|
||||
|
||||
# Aggregate and produce final prediction
|
||||
output = self.aggregation(combined) # (batch, 1)
|
||||
output = self.output_activation(output)
|
||||
|
||||
return output
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
# Test DLinear model
|
||||
config = DLinearConfig(
|
||||
input_dim=30,
|
||||
lookback=168,
|
||||
task='classification',
|
||||
hidden_dim=512,
|
||||
moving_avg_kernel=25
|
||||
)
|
||||
|
||||
model = DLinear(config)
|
||||
print(f"DLinear parameters: {model.count_parameters()}")
|
||||
|
||||
# Test forward pass
|
||||
x = torch.randn(32, 168, 30)
|
||||
output = model(x)
|
||||
print(f"Output shape: {output.shape}")
|
||||
print(f"Output range: [{output.min():.4f}, {output.max():.4f}]")
|
||||
|
||||
# Test ONNX export
|
||||
model.to_onnx('/tmp/dlinear_test.onnx')
|
||||
print("ONNX export successful!")
|
||||
Reference in new issue
Block a user