EXFormer: A Multi-Scale Trend-Aware Transformer with Dynamic Variable Selection for Foreign Exchange Returns Prediction

Battle Count: 255

Highly relevant to quantitative trading. The paper directly addresses daily FX return prediction with rigorous out-of-sample evaluation, trading backtests with transaction costs, Sharpe/Sortino ratios, and regime analysis. EXFormer achieves 18-26% cumulative returns (7-20% after costs) with Sharpe ratios >1.8, outperforming random walk and all baselines. The dynamic variable selector provides actionable interpretability for traders. The 15-day window finding and regime robustness analysis are directly applicable to systematic FX strategies.

High complexity due to multiple custom components: multi-scale trend-aware attention with parallel 2D convolutions, dynamic variable selector with per-timestep softmax weighting, squeeze-and-excitation blocks, multi-scale 1D convolutions, and GRU decoder. Requires careful hyperparameter tuning (Optuna), handling of 28 covariates with mixed frequencies, and proper sliding window implementation. The architecture is not a standard off-the-shelf model and requires custom PyTorch implementation.

The paper provides detailed architecture specifications, hyperparameter search spaces (Appendix Tables A.16-A.19), optimal hyperparameters per currency pair and window, and evaluation protocols. However, no GitHub repository or code link is provided. Data is sourced from Wind Database (proprietary). The model is implemented in PyTorch 2.5.1 on NVIDIA RTX 3090. Reproduction would require access to Wind Database and careful hyperparameter tuning via Optuna.