CIKM2024-MATCC: A Novel Approach for Robust Stock Price Prediction Incorporating Market Trends and Cross-time Correlations

This repository is the official implementation of MATCC: A Novel Approach for Robust Stock Price Prediction Incorporating Market Trends and Cross-time Correlations.

MATCC is a novel framework for robust stock price prediction, which explicitly extracts market trends as guiding information, decomposes stock data into trend and fluctuation components, and employs a carefully designed structure for mining cross-time correlation.

Requirements

1. Install dependencies.

• pandas == 1.5.3
• torch == 1.13.0

2. Install Qlib. We recommend you to install Qlib from SJTU-Qlib.

3. Get the CSI300 and CSI800 stock instruments from Qlib dataset update and put it in your qlib data directory.

4. Use the dataset we have provided. You can follow the link to get our preprocessed dataset 2020_2023. If you fail to download from our link, you may try to use datasets from MASTER.

5. If you want to design and implement your own data preprocessing methods, please refer to util/DropExtremeLabel.py. If you want to change the period of the train/validation/test dataset, please modify util/2023.yaml. To get CSI300 or CSI800, just change the market in util/2023.yaml and universe in util/generate_dataset.py. Run the following commmand:

python util/generate_dataset.py --universe csi300

Preprocessing

The published data went through the following necessary preprocessing.

1. Drop NA features, and perform robust daily Z-score normalization on each feature dimension.
2. Drop NA labels and 5% of the most extreme labels, and perform daily Z-score normalization on labels.

Training

To train the model in the paper, run this command:

python train_model_MATCC.py

Evaluation

To evaluate our model on CSI300 or CSI800 dataset, you can change the universe and model_param_path in test_model_MATCC.py and run:

python test_model_MATCC.py

Then you will get xxxxx_label.pkl and xxxxx_pred.pkl, which are the prediction results and labels.

Backtest

To get the AR, IR, and more portfolio-based metrics, you can run this command:

python backtest.py

Results

Overall performance

Our model achieves the following performance on CSI300, CSI800, S&P500 (2020.07 - 2023.12.31) and NK225 (2022.07 - 2024.07):

Portfolio Results

Extreme Market environment

Normal Market environment

Acknowledgement and Reference Repositories

We appreciate the following github repos a lot for their valuable code base or datasets:

• MASTER: https://github.com/SJTU-Quant/MASTER
• AutoFormer: https://github.com/thuml/Autoformer
• DLinear: https://github.com/cure-lab/LTSF-Linear
• RWKV: https://github.com/BlinkDL/RWKV-LM
• Pytorch_linear_WarmUp_CosineAnnealing: https://github.com/saadnaeem-dev/pytorch-linear-warmup-cosine-annealing-warm-restarts-weight-decay