Comparative analysis of machine learning models in forecasting exchange rate volatility and tail-risk

Attempts have been made in this research to forecast returns of exchange rates of foreign countries in relation to Naira using the SVM, Neural Network (NN), and Random Forest (RF) forecasting models. The Value-at-Risk and Expected Shortfall results demonstrate that exchange rate risks intensified significantly during the post-crisis period. USD/NGN exhibited the highest post-crisis tail risk under the Neural Network model, with VaR₉₉ and ES₉₉ values reaching 1.2211 and 1.3003 respectively, indicating extreme downside exposure and elevated currency market fragility. Similarly, EUR/NGN and CAD/NGN recorded heightened post-crisis risk levels, reflecting increased investor uncertainty and inflationary exchange rate pressures. By contrast, the RF model generated more moderate and economically plausible risk estimates, suggesting stronger robustness and stability in volatile emerging market environments. Graphical analyses corroborate these findings, showing that Neural Network forecasts produced explosive and exponential depreciation trajectories in the post-pandemic era, while RF forecasts exhibited smoother and more gradual adjustment paths consistent with managed exchange rate dynamics. The study found consistently higher post-crisis VaR and ES values across models signal rising tail risks, which imply potential for large currency swings. Such volatility could exacerbate macroeconomic fragility, increase the cost of external debt servicing, and drive inflationary pressures through more expensive imports. Across all models, the RF system consistently delivered superior predictive accuracy, forecast stability, and tail-risk moderation particularly during the crisis and post-crisis periods. In contrast, the NN model produced exponential post-crisis forecast trajectories especially for USD/NGN and EUR/NGN; highly sensitive to structural breaks and may exaggerate persistent volatility in crisis-prone economies. Though informative, such outputs tended to overshoot plausible devaluation trends for the Naira, likely due to exaggerated extrapolation of recent market behaviors. SVM forecasts showed modest error levels and smoother, more plausible trends, offering less extreme but significant signals of future currency devaluation. This reinforces the importance of hybrid modeling approaches and the integration of non-linear machine learning tools in forecasting into central banking operations, exchange rate surveillance frameworks, and investor risk assessment strategies under turbulent economic conditions to enhance resilience against future external shocks and currency market disruptions.