GLoVE: GARCH-Loss Based Volatility Forecasting Experiment

How to reenact the result and further comparison

0. Set up Python virtual environment

source .venv/bin/activate
pip install -r requirements.txt

1. Download newest dataset from Releases

• unzip and rename the folder to "Dataset"

2. Run following code to complete data preprocessing

python src/preprocess.py --input Dataset/ml_dataset_alpha101_volatility.csv --output Dataset/clean.pkl --disabled_features close log_return u_hat_90 gjrgarch_var_90 tgarch_var_90 --use_log_target --target_col var_true_90 --garch_col garch_var_90 

• --use_log_target: 將target_col及garch_col數據取ln(1+p)，後續Eval Model時設定需與此相同
• Final data is in Dataset/clean.pkl

3. Build dataset

python src/dataset_builder.py --input Dataset/clean.pkl --output Dataset/ts_data.pkl --val_frac 0.2 --test_frac 0.1 --input_chunk_length 90 --static_mode ticker --target_col var_true_90 --garch_col garch_var_90 --scale_target zscore

• --scale_target: 預設為none，可調整為mean, zscore，對target_col及garch_col標準化，後續Eval Model時設定需與此相同

4. Execute deep learning scripts

腳本會自動化進行以下步驟:

1. 以optuna搜尋最佳超參數
2. 對各種lambda分別訓練模型及預測
3. 將各模型結果整合並繪製成圖表

TSMixer

bash tsmixer_pipeline.sh --parallel 4 --use_log_target --covariate_mode alpha --invert_train_scale zscore

LSTM

bash lstm_pipeline.sh --parallel 4 --use_log_target --covariate_mode alpha --invert_train_scale zscore

• --parallel: 平行化幾個
• --use_log_target: 若Step 2.處理資料時有進行log轉換，則需要在此轉換回來
• --covariate_mode: 預設為none，不使用alpha資料；設定為alpha會使用alpha資料
• --invert_train_scale: 若Step 3.處理資料時有進行標準化，則需要在此返還波動度為真實值
• 其餘參數可在.sh檔調整

Trained Model and Experimental Results

https://drive.google.com/drive/u/0/folders/1xbHQe9j2vbTLrdpKLcS7qy-QH8qNLHXI