We propose EXFormer, a Transformer-based model that integrates multi-scale trend-aware self-attention, multi-branch convolutions and a dynamic variable selector to forecast 1-day-ahead FX returns and provide pre-hoc interpretability; EXFormer outperforms random walk and multiple deep-learning/econometric baselines in MSFE and directional accuracy, and its signals produce economically meaningful backtest profits (e.g., cumulative returns ~18–26% and Sharpe >1.8 in frictionless tests; positive net returns remain after conservative transaction costs) while the dynamic selector reveals time-varying drivers such as commodity indices and long-term yields .
| Pair | Mean Daily Return (%) | Cumulative Return (%) | Sharpe Ratio | Max Drawdown (%) |
|---|---|---|---|---|
| EUR/USD (EXFormer) | 0.050 | 18.572 | 2.087 | 4.068 |
| USD/JPY (EXFormer) | 0.068 | 25.782 | 1.891 | 6.778 |
| GBP/USD (EXFormer) | 0.050 | 18.373 | 1.903 | 3.526 |
元数据与概览(引言与报告概览)
逐节深度解读
图表与图像逐一解析(关键图/表详解)
估值/性能解读与统计稳健性
批判性视角与局限性提示(基于报告文本与数据)
结论性综合(关键发现与建议)
補充引用(关键表/附录)
總結語(一句话版)
2512.12727.pdf