The study examines whether machine learning models outperform traditional econometric models in forecasting EUR/USD exchange rates. The main models evaluated in this study are Random Walk, ARIMA, Support Vector Regression (SVR), and Random Forest, evaluated across two distinct periods across 48 months. The test periods are classified as Period A, a calmer environment (January 2016 to December 2019), and Period B, a more volatile environment (January 2020 to December 2023). Using averaged monthly data from the Federal Reserve Economic Database and three main accuracy metrics (RMSE, MAE, and MAPE), the study concludes that model performance is strongly dependent on the economic environment. In the calmer period, ARIMA performed best, followed closely by Random Forest, with both outperforming the Random Walk benchmark. In the volatile period, however, both machine learning models deteriorated dramatically with RMSE increases of approximately 180–191 percent while ARIMA remained the most accurate model across both periods. In comparison to this study’s specific outcome, SVR underperformed the benchmark Random Walk in both environments. These findings suggest that machine learning is vulnerable to structural breaks and out of trend economic conditions, while simpler traditional models offer greater accuracy. Our results reinforce the relevance of the Meese-Rogoff puzzle and highlight the importance of evaluating different forecasting models across varying economic conditions.