COMPARATIVE EVALUATION OF GENERALIZED AUTOREGRESSIVE CONDITIONAL HETEROSKEDASTICITY-TYPE AND MACHINE-LEARNING MODELS FOR EXCHANGE-RATE VOLATILITY FORECASTING IN NIGERIA

Abstract

This study investigates exchange-rate volatility forecasting in Nigeria by comparing conventional econometric models with machine-learning approaches using a common volatility target. Monthly USD/NGN exchange-rate data obtained from the Central Bank of Nigeria covering January 2004 to December 2025 were analysed. Monthly log returns were computed, while squared returns served as a proxy for realized volatility. The empirical framework combines an ARMA mean specification with GARCH(1,1) and EGARCH(1,1) variance models, alongside Support Vector Regression (SVR) and Artificial Neural Networks (ANN). Forecast performance was evaluated using rolling time-series cross-validation and standard error measures, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and the Diebold–Mariano test. Results indicate that returns are stationary, leptokurtic, and exhibit volatility clustering with evidence of structural instability. ARMA–GARCH(1,1) outperforms EGARCH(1,1) in stability, while SVR achieves the best forecasting accuracy followed by ANN. However, the Diebold–Mariano test shows that differences in predictive accuracy are not statistically significant at the 5% level. The findings suggest that machine-learning methods, particularly SVR, complement traditional volatility models in exchange-rate risk management in Nigeria