Hybrid Machine Learning-Bayesian Vector Error Correction Model for Forecasting of Exchange Rate and Inflation Volatility Dynamics in Nigeria

Abstract

This study develops and evaluates a hybrid Machine Learning-Bayesian Vector Error Correction Model (ML-BVECM) for forecasting exchange rate and inflation volatility dynamics in Nigeria using quarterly data from 1990 to 2025. The Unit root tests (ADF, PP, and KPSS) reveal that both inflation and the nominal exchange rate are integrated of order one (I(1)), while Johansen cointegration analysis confirms a stable long-run equilibrium relationship between the two variables. The structural component of the model, the BVECM, reveals a strong exchange rate pass-through to domestic prices, where a 1% increase in the Naira-Dollar exchange rate results in a 0.62% rise in inflation. The error correction coefficients highlight a rapid quarterly adjustment back to equilibrium: 31% for inflation and 28% for the exchange rate while the use of Bayesian priors stabilizes the parameters against extreme structural shocks such as the 2016 foreign exchange crisis and the 2023 currency float. To capture remaining short-term nonlinearities and volatility clustering, the BVECM residuals were modeled using Random Forest (RF) and Gradient Boosting (GB) algorithms. Specifically, the Random Forest hybrid framework yields the highest accuracy for forecasting inflation, whereas the Gradient Boosting hybrid variant is superior for exchange rate predictions, driving overall error reductions  by over 20%. Diagnostic and stability tests  confirm the model's structural integrity and robustness. The findings offer a potent toolkit for the Central Bank of Nigeria and other policy makers to proactively target inflation and navigate market volatility.