A Learnable Wavelet Transformer for Long-Short Equity Trading and Risk-Adjusted Return Optimization

Shuozhe Li, Du Cheng, Amy Zhang, Liu Leqi

Presentation Attendance: No, we cannot present in-person

Keywords: Neural wavelet regularization, wavelet-transformer network, low-guided high-frequency injection, return optimization.

Abstract: Learning profitable intraday trading policies from financial time series is challenging due to heavy noise, non-stationarity, and cross-sectional dependence. We propose WaveLSFormer, a learnable wavelet-based long--short Transformer that jointly performs multi-scale decomposition and return-oriented decision learning. A neural wavelet front-end learns an end-to-end filter bank to produce low-/high-frequency components, and a low-guided high-frequency injection (LGHI) module enables stable cross-frequency fusion. The model outputs continuous long/short positions under a fixed risk budget and is trained with a trading objective with risk-aware regularization. Experiments on five years of hourly U.S equity data across six industry groups and ten random seeds show consistent gains over matched MLP/LSTM/Transformer baselines with fixed or learnable wavelet front-ends. WaveLSFormer achieves $0.607 \pm 0.045$ ROI and $2.157 \pm 0.166$ Sharpe on average, substantially improving both profitability and risk-adjusted returns over the strongest baselines.

Track: Research Track (max 4 pages)

Submission Number: 3