策略類型
機器學習交易演算法
面向量化金融的開源機器學習模型
機器學習交易演算法將統計學習模型（梯度提升、神經網路、Transformer）應用於金融時間序列資料，以預測價格走勢或產生交易訊號。這些開源實作涵蓋迴歸（價格目標）與分類（方向）兩種表述。
12 個演算法2 個函式庫
演算法網路
ML 演算法如何跨函式庫連接
ML Algorithms
Freqtrade9 algos
Microsoft Qlib6 algos
LightGBMRegressorintermediate
LightGBMClassifierintermediate
XGBoostRegressorintermediate
XGBoostClassifierintermediate
PyTorchMLPRegressoradvanced
PyTorchTransformerRegressoradvanced
LGBModelintermediate
XGBModelintermediate
DNNModeladvanced
ALSTMadvanced
TFTModeladvanced
GATsadvanced
交易決策流程
ML 演算法如何在交易系統中協同運作
1
Market Data Input
Price, volume, indicators
OHLCV price data
Technical indicators
Feature engineering
2
ML Signal Generation
Model predictions
LightGBM/XGBoost regression
Transformer sequence modeling
MLP classification
3
Entry Decision
Signal threshold filtering
Confidence score > 0.6
Direction: Long/Short
4
Exit Strategy
Profit taking & stop loss
Take profit (fixed % or trailing)
Stop loss (ATR-based)
5
Risk Management
Position sizing & portfolio
Position sizing (Kelly criterion)
Portfolio rebalancing
演算法比較
在關鍵維度上比較 ML 演算法
演算法對比矩陣點擊欄以展開詳情
| 指標 | LightGBMRegressorFreqtrade | LightGBMClassifierFreqtrade | XGBoostRegressorFreqtrade | XGBoostClassifierFreqtrade | PyTorchMLPRegressorFreqtrade | PyTorchTransformerRegressorFreqtrade | LGBModelQlib (Microsoft) | XGBModelQlib (Microsoft) | DNNModelQlib (Microsoft) | ALSTMQlib (Microsoft) | TFTModelQlib (Microsoft) | GATsQlib (Microsoft) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 複雜度 | ⭐⭐⭐intermediate | ⭐⭐⭐intermediate | ⭐⭐⭐intermediate | ⭐⭐⭐intermediate | ⭐⭐⭐⭐advanced | ⭐⭐⭐⭐advanced | ⭐⭐⭐intermediate | ⭐⭐⭐intermediate | ⭐⭐⭐⭐advanced | ⭐⭐⭐⭐advanced | ⭐⭐⭐⭐advanced | ⭐⭐⭐⭐advanced |
| 預測類型 | 迴歸 | 分類 | 迴歸 | 分類 | 迴歸 | 迴歸 | 混合 | 混合 | 混合 | 序列 | 混合 | 混合 |
| 訓練速度 | ⚡⚡⚡ | ⚡⚡⚡ | ⚡⚡⚡ | ⚡⚡⚡ | ⚡⚡ | ⚡ | ⚡⚡ | ⚡⚡ | ⚡⚡ | ⚡ | ⚡⚡ | ⚡⚡ |
| 準確度 | 📊📊📊 | 📊📊📊 | 📊📊📊 | 📊📊📊 | 📊📊📊 | 📊📊📊📊 | 📊📊 | 📊📊 | 📊📊 | 📊📊📊 | 📊📊 | 📊📊📊📊 |
| 最適合 | 表格資料 | 表格資料 | 表格資料 | 表格資料 | 非線性模式 | 時間序列模式 | 通用 | 通用 | 通用 | 序列資料 | 通用 | 圖關係 |
Freqtrade
Gradient boosting regression model for price movement prediction using LightGBM.
原始碼：
freqtrade/freqai/prediction_models/LightGBMRegressor.pyLightGBM classification model for user-defined discrete labels.
原始碼：
freqtrade/freqai/prediction_models/LightGBMClassifier.pyXGBoost-based regression model for continuous value prediction.
原始碼：
freqtrade/freqai/prediction_models/XGBoostRegressor.pyXGBoost classification model for user-defined discrete labels.
原始碼：
freqtrade/freqai/prediction_models/XGBoostClassifier.pyMulti-layer perceptron neural network for regression-based price prediction.
原始碼：
freqtrade/freqai/prediction_models/PyTorchMLPRegressor.pyQlib (Microsoft)
LightGBM model for stock return prediction using technical and fundamental features.
關鍵參數
| num_leaves | 31 | Maximum number of leaves |
| learning_rate | 0.1 | Boosting learning rate |
原始碼：
qlib/contrib/model/gbdt.pyDeep neural network for nonlinear feature extraction and return prediction.
關鍵參數
| hidden_size | 256 | Hidden layer size |
原始碼：
qlib/contrib/model/pytorch_nn.pyAttention-based LSTM for sequential stock data modeling with attention mechanism.
關鍵參數
| hidden_size | 64 | LSTM hidden size |
| num_layers | 2 | Number of LSTM layers |
原始碼：
qlib/contrib/model/pytorch_alstm.pyTemporal Fusion Transformer combining static and temporal features for multi-horizon prediction.
原始碼：
qlib/contrib/model/pytorch_tft.py機器學習交易演算法，演算法參考
- LightGBMRegressor (Freqtrade)
- Gradient boosting regression model for price movement prediction using LightGBM. 原始碼： https://github.com/freqtrade/freqtrade/blob/c0a991f5632cb900075ff2864b6671ed71a319d7/freqtrade/freqai/prediction_models/LightGBMRegressor.py.
- LightGBMClassifier (Freqtrade)
- LightGBM classification model for user-defined discrete labels. 原始碼： https://github.com/freqtrade/freqtrade/blob/c0a991f5632cb900075ff2864b6671ed71a319d7/freqtrade/freqai/prediction_models/LightGBMClassifier.py.
- XGBoostRegressor (Freqtrade)
- XGBoost-based regression model for continuous value prediction. 原始碼： https://github.com/freqtrade/freqtrade/blob/c0a991f5632cb900075ff2864b6671ed71a319d7/freqtrade/freqai/prediction_models/XGBoostRegressor.py.
- XGBoostClassifier (Freqtrade)
- XGBoost classification model for user-defined discrete labels. 原始碼： https://github.com/freqtrade/freqtrade/blob/c0a991f5632cb900075ff2864b6671ed71a319d7/freqtrade/freqai/prediction_models/XGBoostClassifier.py.
- PyTorchMLPRegressor (Freqtrade)
- Multi-layer perceptron neural network for regression-based price prediction. 原始碼： https://github.com/freqtrade/freqtrade/blob/c0a991f5632cb900075ff2864b6671ed71a319d7/freqtrade/freqai/prediction_models/PyTorchMLPRegressor.py.
- PyTorchTransformerRegressor (Freqtrade)
- Transformer architecture for time-series regression using self-attention mechanism. 原始碼： https://github.com/freqtrade/freqtrade/blob/c0a991f5632cb900075ff2864b6671ed71a319d7/freqtrade/freqai/prediction_models/PyTorchTransformerRegressor.py.
- LGBModel (Qlib (Microsoft))
- LightGBM model for stock return prediction using technical and fundamental features. 關鍵參數： num_leaves (Maximum number of leaves), learning_rate (Boosting learning rate).原始碼： https://github.com/microsoft/qlib/blob/main/qlib/contrib/model/gbdt.py.
- XGBModel (Qlib (Microsoft))
- XGBoost model for stock return prediction. 原始碼： https://github.com/microsoft/qlib/blob/main/qlib/contrib/model/xgboost.py.
- DNNModel (Qlib (Microsoft))
- Deep neural network for nonlinear feature extraction and return prediction. 關鍵參數： hidden_size (Hidden layer size).原始碼： https://github.com/microsoft/qlib/blob/main/qlib/contrib/model/pytorch_nn.py.
- ALSTM (Qlib (Microsoft))
- Attention-based LSTM for sequential stock data modeling with attention mechanism. 關鍵參數： hidden_size (LSTM hidden size), num_layers (Number of LSTM layers).原始碼： https://github.com/microsoft/qlib/blob/main/qlib/contrib/model/pytorch_alstm.py.
- TFTModel (Qlib (Microsoft))
- Temporal Fusion Transformer combining static and temporal features for multi-horizon prediction. 原始碼： https://github.com/microsoft/qlib/blob/main/qlib/contrib/model/pytorch_tft.py.
- GATs (Qlib (Microsoft))
- Graph Attention Networks modeling stock relationships for cross-stock prediction. 原始碼： https://github.com/microsoft/qlib/blob/main/qlib/contrib/model/pytorch_gats.py.