A Time-Varying Interpretable Multi-Scale Transformer for Exchange Rate Returns Forecasting and Investment Decision Support

Abstract

Accurately forecasting daily exchange rate returns represents a longstanding challenge in international finance, as the exchange rate returns are driven by a multitude of correlated market factors and exhibit high-frequency fluctuations. This paper proposes EXFormer, a novel Transformer-based architecture specifically designed for forecasting the daily exchange rate returns. We introduce a multi-scale trend-aware self-attention mechanism that employs parallel convolutional branches with differing receptive fields to align observations on the basis of local slopes, preserving long-range dependencies while remaining sensitive to regime shifts. A dynamic variable selector assigns time-varying importance weights to 28 exogenous covariates related to exchange rate returns, providing pre-hoc interpretability. An embedded squeeze-and-excitation block recalibrates channel responses to emphasize informative features and depress noise in the forecasting. Using daily EUR/USD, USD/JPY, and GBP/USD returns, we conduct out-of-sample evaluations across five different sliding windows. The model consistently outperforms the random walk benchmark, improving directional accuracy by a statistically significant margin of 8.5-22.8%. Nearly one year of trading backtests shows that these statistical gains translate into economically meaningful performance, obtaining cumulative net returns of 7%, 19%, and 9% with Sharpe ratios exceeding 1.8 after transaction costs. The robustness checks further confirm the model’s superiority under high-volatility and bear-market regimes. EXFormer offers a practical, interpretable, and knowledge-driven tool into exchange rate dynamics for international investors, multinational firms, and policy institutions.