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

Battle Count: 79

This paper is directly and fundamentally about quantitative trading. It proposes an end-to-end trading policy model that outputs long/short positions, optimizes a trading-aware objective (ROI + Sharpe), incorporates risk-budget constraints, and evaluates on real intraday equity data. The entire framework is designed to replace the traditional forecast-then-trade pipeline with direct position optimization. Results show substantial improvements in both profitability (ROI 0.607 vs 0.225) and risk-adjusted returns (Sharpe 2.157 vs 1.024) over baselines.

High complexity due to: (1) learnable FIR wavelet filter bank with differentiable spectral regularization via rFFT, (2) Low-Guided High-frequency Injection (LGHI) module with gated residual control, (3) multi-component loss function (soft-label CE + Sharpe regularizer + ROI penalty + wavelet spectral loss), (4) risk-budget rescaling with dead-zone and leverage cap, (5) DTW-based stock selection and Granger causality filtering pipeline, (6) two-phase training schedule for wavelet stabilization, (7) ProbSparse attention with distillation for efficiency. Requires careful hyperparameter tuning (lambda_roi, lambda_spec, gamma initialization, k for soft labels).

The paper provides detailed hyperparameters (d_model=512, d_ff=1024, 6 layers, 128 heads, sequence length 96, batch size 256, lr=1e-5, 80 epochs, 10 random seeds), explicit loss formulations, and data pipeline descriptions. However, no code repository is provided. Data sources (polygon.io, StockAnalysis) are identified but polygon.io is commercial. The DTW and Granger causality filtering pipeline is described but implementation details for reproducibility are partially lacking. The risk-budget scaling and dead-zone parameters are specified.