EXCHANGE RATE DYNAMICS AND FORECASTING ACCURACY IN EMERGINGECONOMIES: INTEGRATING ENSEMBLE LEARNING MODELS WITH STRUCTURAL TIME SERIES DECOMPOSITION FOR THE USD/ZAR RATE

Abstract

Thisstudyinvestigatesthepredictiveperformanceofmachinelearningmodelsinforecastingthedailyexchange rate of the South African rand (USD/ZAR) using data from March 5, 2020, to July 19, 2024. Employing statistical diagnostics such as ADF and KPSS tests, the exchange rate series is found to be non-stationary in levels but stationary after first differencing. Descriptive statistics and seasonal-trend decomposition reveal notable structural features, including skewness, excess kurtosis, and strong seasonal effects. Among the models tested, AdaBoost outperformed Random Forest and K-Nearest Neighbors in terms of forecast accuracy, as measured by RMSE and MAE, despite lowexplanatorypower across all models. These findings underscore the potential of ensemble learning methods for improving short-termcurrency forecasts in emerging markets, while alsohighlightingthelimitationsofdata-drivenmodels incapturingtheinherent volatilityandunpredictabilityof exchange rate movements. The study offers relevant policy insights for monetary authorities and financial market participants, and recommends the integration of hybrid modeling frameworks that combine machine learning with structural macroeconomic variables.