Exchange rate forecasting in post-subsidy Nigeria: A hybrid approach using asymmetric jump diffusion and deep learning

Abstract

The rapid, nonlinear-asymmetric jumps, long-term dependencies, and chaotic price fluctuations in the post-subsidy Nigerian exchange rates’ era have posed increased challenges on the existing traditional univariate ARIMA and GARCH models’ forecasting power. To fill the existing gap, the study has fused an Asymmetric Jump Diffusion (AJD) that models asymmetric jumps with deep learning based on Long Short-Term Memory (LSTM) and Recurrent Neural Network (RNN) to capture hidden nonlinear patterns in the pre and post subsidy era in the Nigerian exchange rates. The data used are daily values of exchange rates of the Nigerian Naira for various dominant currencies (USD, GBP, and EUR) over a period July 11, 1995, up to May 9, 2025. The pre and post subsidy periods are respectively, 11 July 1995 up to 29 May 2023 and 30 May 2023 up to 9 May 2025. Based on the results obtained, the hybrid Asymmetric Jump Diffusion Deep Learning model performed better compared to traditional forecasting models on accuracy and robustness. While market jumps were captured by the AJD component, the long-term patterns were captured by the LSTM model. The Long Short-Term Memory and Recurrent Neural Network models did extraordinarily well in the pre-subsidy period and showed significant degradation in the post-subsidy period. The error measures increased significantly to reflect how much they struggled with the instability introduced by the subsidy policy change. Only two hybrid models were able to show acceptable adaptability: AJD + RNN and LSTM + RNN, which gave relatively low mean square errors with values of 0.0234 and 0.0153, respectively, post-29 May 2023. The findings from this analysis provide the policymakers, businesses, and financial analysts with a more reliable tool for forecasting exchange rates.