Revisiting exchange rate predictability: Can machine learning with theoretical filtering outperform canonical models?

We revisit the exchange-rate predictability puzzle by asking whether standard, widely used machine-learning (ML) algorithms convincingly improve exchange rate forecasting once evaluation is disciplined and implementation is made robust. Using monthly data, we study US dollar to British pound as the baseline case (in both levels and monthly percent changes). We compare five ML methods ¬random forests, neural networks, LASSO, gradient boosting, and linear support-vector classification ¬against canonical benchmarks (random walk and ARIMA) in a rolling one-step-ahead out-of-sample forecasting design. To mitigate sensitivity to stochastic estimation, we average forecasts across multiple random seeds and assess performance using RMSE and Diebold-Mariano tests. We find that ML does not improve level forecasts and typically underperforms ARIMA. For exchange-rate changes, ML methods consistently outperform the random-walk benchmark, but only neural networks¬ under a specific design¬ reliably beat ARIMA. A theory-based UIP/PPP filtering approach improves accuracy for both ML and univariate methods, yet does not change the overall ranking. Extensive robustness checks across windows, currencies, frequencies, and tuning choices confirm that ML’s advantages are limited and fragile relative to conventional univariate benchmarks.