A deep recurrent neural network approach to inflation forecasting in Nigeria: Evaluating the impact of feature selection methods

This study developed and compared statistical, machine learning, and deep learning models for forecasting inflation in Nigeria, using monthly macroeconomic and financial indicators from 2006 to 2025. Traditional econometric approaches such as SARIMA and SARIMAX were benchmarked against advanced machine learning methods including Random Forest, Support Vector Regression (SVR), XGBoost, and a Long Short-Term Memory (LSTM) deep neural network. To ensure robust model design, feature engineering methods, block-permutation feature importance, and gradient-based interpretability techniques were employed. Results demonstrate that the LSTM model outperformed traditional time-series models by effectively capturing nonlinearities, persistence, and structural breaks inherent in Nigeria’s inflation dynamics. Among machine learning models, SVR exhibited strong short-term accuracy, but LSTM provided superior stability and generalization. Feature importance analysis revealed that net domestic credit (NDC), other money supply aggregates (MS3 and MS2), currency outside banks (COB), currency in circulation (CiC), and demand deposits (DD) were the most critical drivers of inflation, reflecting the combined impact of credit expansion, liquidity growth, and cash-dominated transactions on Nigeria’s price stability. Policy implications highlight the need for coordinated credit regulation, prudent monetary management, and reforms to reduce the economy’s structural reliance on liquidity growth. The study provides both methodological innovation and practical insights for more reliable inflation forecasting in developing economies.