![Python 3.10](https://img.shields.io/badge/python-3.10-green.svg?style=plastic)
![PyTorch 1.13](https://img.shields.io/badge/pytorch-1.13-green.svg?style=plastic)
![PyTorch Lightning 1.8](https://img.shields.io/badge/pytorch%20lightning-1.8-green.svg?style=plastic)

# WaveLSFromer

WaveLSFromer is a research codebase for long-sequence financial time-series
forecasting. It extends the Informer/Stockformer style transformer stack with
stock-specific training objectives, PyTorch Lightning experiment loops,
config-driven model runs, and learnable wavelet front-end components for
low/high frequency feature extraction.

Paper: [A Learnable Wavelet Transformer for Long-Short Equity Trading and Risk-Adjusted Return Optimization](https://arxiv.org/abs/2601.13435).

The repository includes:

- transformer, Informer, DLinear, LSTM, and MLP model baselines;
- learnable 1D wavelet filters with frequency-domain regularization;
- PyTorch Lightning training, validation, prediction, and checkpoint workflows;
- stock-return metrics and differentiable trading-oriented loss functions;
- YAML experiment configs for financial and benchmark time-series datasets;
- notebooks and scripts for data collection, preparation, and result analysis.

Thanks to [polygon.io](http://polygon.io/) for being our financial data provider.