Rolling-Origin Evaluation of Lag-Based Regularized Regression Models for Indonesian Inflation Forecasting

Abstract

Inflation forecasting studies often rely on external macroeconomic predictors or complex machine learning models, while the predictive value and limitations of the internal lag structure of Indonesian inflation remain less explicitly examined. This study evaluates Indonesian inflation forecasting within a univariate lag-based framework using monthly inflation data from January 2010 to December 2025. Lagged inflation values are used to construct regression predictors, and OLS, Ridge, LASSO, and Elastic Net are evaluated through rolling-origin forecasting with an expanding window. To strengthen the time-series basis of the analysis, this study also conducts stationarity diagnostics, ACF and PACF analysis, seasonal diagnostics, lag-specification robustness checks, and comparisons with standard forecasting benchmarks, including Naive, Seasonal Naive, AR selected by AIC, and ARIMA selected by AIC. The ADF test produces a p-value of 0.608360, indicating weak evidence of stationarity in level form. Among the regularised regression models, Ridge produces the lowest descriptive forecast errors, with RMSE of 0.405018 and MAE of 0.308967. However, Diebold–Mariano tests indicate that the differences among OLS, Ridge, LASSO, and Elastic Net are not statistically significant. Benchmark comparisons show that the Naive forecast achieves the lowest RMSE of 0.373819, while ARIMA selected by AIC achieves the lowest MAE of 0.279641 and MAPE of 16.560958. Robustness checks also show that the twelve-lag specification is competitive for OLS and Ridge, but it is not uniformly optimal across all models. These findings suggest that the main value of lag-based regularised regression lies in clarifying the limited but useful short-run predictive information contained in the internal temporal structure of Indonesian inflation, rather than in providing a statistically dominant or complete inflation forecasting model.