A Persistence-Driven Naïve-LSTM Residual Hybrid Model for Forecasting the United States Dollar-Nigerian Naira Exchange Rate

Abstract

Accurate exchange-rate forecasting is important for financial planning, investment decision-making, and economic risk management, particularly in emerging economies such as Nigeria. This study evaluates a residual-based Hybrid Naïve-LSTM model for forecasting the Nigerian Naira against the United States Dollar using daily Central Bank of Nigeria (CBN) exchange-rate data from October 2009 to August 2026. Brent crude oil prices were incorporated as an exogenous variable, while feature engineering produced 49 predictors comprising exchange-rate lags, returns, moving averages, volatility measures, calendar variables, and crude-oil features. Exploratory analysis indicated strong persistence, non-stationarity, and substantial volatility in the exchange-rate series. The proposed Hybrid Naïve-LSTM model was evaluated against a standalone Naïve persistence model and an ARIMA-LSTM state-of-the-art benchmark under comparable experimental conditions. The results show that the Naïve model achieved the best overall performance, recording an MAE of 15.3367, RMSE of 34.9274, MAPE of 1.3733%, and R² of 0.9947. The proposed Hybrid Naïve-LSTM model achieved an MAE of 18.7539, RMSE of 36.2391, MAPE of 1.8465%, and R² of 0.9943. These findings demonstrate that, despite the ability of LSTM-based hybrid models to capture nonlinear residual patterns, the strong persistence of the Nigerian Naira–US Dollar exchange-rate series enabled the simple Naïve model to outperform the more complex alternatives. The study therefore highlights the importance of benchmarking complex deep-learning models against strong persistence-based baselines when forecasting highly persistent financial time series. Future research should investigate alternative deep-learning architectures with explicit regime-switching mechanisms, multi-horizon forecasting, and a broader set of macroeconomic variables.