Explainable Artificial Intelligence for Inflation Forecasting in Nigeria: Interpreting A Multilayer Perceptron Model Using Shap Values

Abstract

Inflation forecasting is important for monetary policy, economic planning and decision-making, yet neural-network forecasts can be difficult to interpret because their internal relationships are not directly observable. This study applies SHapley Additive exPlanations (SHAP) to interpret a Multilayer Perceptron (MLP) model developed for forecasting inflation in Nigeria. Monthly macroeconomic data covering 2007–2024 were used, comprising inflation rate, interest rate, exchange rate, oil price and Consumer Price Index (CPI), together with lagged and temporal features used in the MLP. The SHAP analysis was used to explain an individual inflation prediction and identify the direction and relative magnitude of feature contributions. The model prediction examined increased from a baseline of 10.20 to 13.49. The largest positive contributions were associated with the third and first lags of oil price (OP_lag3 and OP_lag1), followed by seasonal and lagged inflation features. CPI variables also contributed positively, whereas exchange-rate and interest-rate variables showed smaller and mixed effects for the prediction examined. The result is consistent with Nigerian evidence that oil-price movements can transmit into domestic inflation, while the importance of lagged inflation supports inflation persistence. The study also demonstrates that SHAP can make an MLP more transparent without treating feature importance as causal evidence. Although the broader thesis found that the MLP did not outperform SARIMA and VAR in out-of-sample accuracy, the SHAP analysis shows that the model learned economically meaningful temporal and macroeconomic information. The study concludes that explainable machine learning can complement conventional inflation forecasting by providing interpretable information about the drivers of individual predictions.