Does Machine Learning Beat ARIMA in Foreign Exchange? Out-of- Sample Evidence on Trading Activity, Volatility and Prices for EUR/USD, GBP/USD and USD/CAD, 2010–2021

Machine-learning models are used more and more to predict financial markets, but whether they beat simple time-series models is rarely tested under strict out-of-sample conditions. This study compares a Random Forest and a Gradient Boosting Machine (GBM), each using 26 price-derived indicators, with ARIMA models and simple rules for EUR/USD, GBP/USD and USD/CAD, using daily data from 2010 to 2021 and an 890-day test period from February 2018 to July 2021. For next-day tick-volume activity the comparison depends on information. Without volume history the Random Forest correctly classifies 31.8–41.6% of test days, while ARIMA correctly classifies 61.5–63.8%; with volume history the difference between the two is not significant. The GBM forecasts five-day realised volatility less accurately than ARIMA for EUR/USD and USD/CAD and is not significantly different from it for GBP/USD. In the main design it beats a trailing-volatility baseline only for GBP/USD, where a constant forecast performs just as well. For the price, ARIMA does no better than the random walk. Seven robustness analyses, including GARCH and HAR benchmarks, yearly re-fits, a single data source and other horizons, leave these results unchanged. Of 114 tests, four machine-learning wins remain significant after the Holm correction, and all of them are against naive baselines. The results suggest that a price-based machine-learning forecast should beat a simple time-series model, on the same information and test days, before it is trusted. The code, a pinned data download and file hashes are released for reproduction.