<![CDATA[Accurate inflation forecasting is crucial for economic stability and effective policymaking, particularly in emerging economies such as Indonesia, where monetary policy, global commodity markets, exchange rate fluctuations, and recurring religious seasonal events simultaneously influence price dynamics. This study proposes an inflation forecasting framework using the N-BEATSx (Neural Basis Expansion Analysis for Time Series with Exogenous Variables) deep learning model, incorporating macroeconomic variables, global oil prices, BI Rate, and the USD/IDR exchange rate, alongside Ramadan and Eid al-Fitr calendar dummy variables as exogenous inputs. The dataset comprises 153 monthly observations spanning January 2013 to September 2025, split into training, validation, and test sets, with a forecasting horizon of six months. The N-BEATSx model is benchmarked against SARIMAX, LSTM, and Prophet. Results on the test set show that N-BEATSx achieves competitive performance (RMSE 0.0067, MAE 0.0058, SMAPE 51.77%) outperforming SARIMAX (RMSE 0.0297, MAE 0.0266) and LSTM (RMSE 0.0098, MAE 0.0084). Although Prophet yields marginally lower absolute errors, the MAE gap is minimal (0.0002), while N-BEATSx offers superior interpretability through an explicit decomposition of forecasts into trend, seasonality, and exogenous components. Component decomposition analysis reveals that macroeconomic exogenous variables dominate the forecast output, confirming their theoretical relevance as inflation drivers. Six-month-ahead forecasts project inflation in the range of 2.47% - 3.54% for October 2025 to March 2026, approaching Bank Indonesia’s upper target corridor, suggesting the need for preemptive monetary policy measures.]]>
Copyrights © 2026
