Empirical Analysis of the Performance of Time Series Forecasting Models for Naira Exchange Rate Prediction

Accurate exchange rate forecasting has significant economic consequences for Nigerian businesses, policymakers, and investors, and empirically comparing the performance of classical and deep learning time series models for naira exchange rate prediction fills an important applied data science gap. This study empirically compared five forecasting approaches (ARIMA, SARIMA, Exponential Smoothing, LSTM, and a Transformer-based model) for predicting the official and parallel market naira-to-dollar exchange rates using daily data from January 2015 to December 2023. Data were sourced from the CBN and an aggregated parallel market price tracker maintained by a licensed financial data provider. Models were evaluated using RMSE, MAE, and directional accuracy on a six-month rolling holdout window. The LSTM model achieved the lowest RMSE of 12.8 naira per dollar for the official rate but showed higher instability at structural break points including the 2016 devaluation and the 2023 unification. The Transformer model showed superior directional accuracy at 71.3 percent. ARIMA performed comparably to LSTM during stable periods. Parallel market rates showed significantly higher volatility and were less predictable across all models (mean RMSE increase of 34.7 percent). Ensemble averaging of LSTM and ARIMA predictions reduced RMSE by 8.2 percent versus the best individual model. The study provides an original empirical comparison unavailable in Nigerian-specific financial forecasting literature and recommends ensemble LSTM-ARIMA as the baseline for naira exchange rate forecasting in both corporate treasury and policy contexts.

Keywords: exchange rate forecasting, LSTM, ARIMA, naira, time series models