ML Models Struggle to Beat Random Walk in CAD/USD Exchange Rate Forecasting

This research investigates the effectiveness of various machine learning (ML) models in forecasting the monthly USD/CAD exchange rate, comparing them against a simple random walk benchmark. Using daily data from the Bank of Canada, resampled into monthly observations, five ML models—linear regression, random forest, gradient boosting, XGBoost, and AdaBoost—were evaluated. The models were assessed using an expanding-window framework to ensure strict out-of-sample integrity, and forecast accuracy differences were statistically tested. The study identified several structural breaks in the exchange rate series, corresponding to significant economic events. The findings indicate that the naive random walk model remains a very strong benchmark. Only linear regression demonstrated a statistically significant improvement over the random walk. While Random Forest achieved the lowest MAPE among ML models, the overall performance of complex ML ensemble models showed only marginal differences, reinforcing the difficulty of beating the random walk in exchange rate prediction. SHAP analysis confirmed that short-term lags were the primary drivers of predictions.