Transitioning from static to temporal machine learning models: Forecasting the EUR/USD exchange rate utilizing lagged factors and feature significance

The prediction of exchange rates remains a critical and challenging task in international finance, given their sensitivity to macroeconomic, energy, and geopolitical factors. This paper focuses on forecasting the EUR/USD exchange rate using daily data between January 2024 and June 2025, with particular attention to the influence of oil prices and the Geopolitical Risk (GPR) index. Unlike most studies that rely on static machine learning (ML) models, we propose a temporal feature engineering framework that transforms Support Vector Regression (SVR), Random Forest (RF), and k-Nearest Neighbors (KNN) into time-aware predictors. Specifically, lagged variables over a seven-day horizon are incorporated, yielding 21 explanatory features. To further enhance model interpretability and efficiency, feature importance is computed via RF, allowing us to identify and retain the most relevant predictors. The proposed methodology is evaluated by comparing six models: SVR, RF, and KNN trained with the complete feature set versus their counterparts trained only on the selected variables. Results demonstrate that embedding temporal dependencies significantly improves forecasting performance and that variable selection provides both accuracy gains and economic insights. These findings highlight the pivotal role of oil price dynamics, geopolitical uncertainty, and short-term memory in driving EUR/USD fluctuations. Overall, our study offers a practical and interpretable ML-based framework for exchange rate forecasting, bridging the gap between econometric intuition and modern data-driven approaches.