A Comparative Study of Machine Learning and ARIMA Models for Inflation Forecasting in Nigeria

Reliable inflation forecasting is essential for informed economic planning and policy implementation, particularly in developing nations like Nigeria. This study adopts a hybrid modeling framework that integrates classical time series techniques with machine learning approaches to predict monthly inflation rates in Nigeria. Using data from January 2003 to February 2025 sourced from Kaggle, the study applies preprocessing techniques including stationarity checks and first-order differencing to ensure model readiness. The ARIMA (2,1,2) model was selected as the optimal time series model based on Akaike and Bayesian Information Criteria. Concurrently, Artificial Neural Networks (ANNs) employing various activation functions-ReLU, Sigmoid, LogSigmoid, and TanSigmoid were evaluated. Among these, the ANN with ReLU activation achieved the best performance, recording the lowest error metrics (MSE: 0.6091, RMSE: 0.7805, MAE: 0.6682, MAPE: 0.02008). While a hybrid ARIMA-ANN model was also explored, the standalone ANN with ReLU activation demonstrated superior predictive accuracy. The findings suggest that machine learning models, particularly deep learning architectures with well-chosen activation functions, offer significant advantages over traditional econometric methods in forecasting inflation. This study gives useful ideas for policymakers and economic analysts, and suggests that future research should look into combining different models and adding more related factors to make predictions even better.