<![CDATA[Accurate forecasting of regional export values is critical for effective macroeconomic planning. However, these indicators often exhibit complex volatility and structural shocks that challenge traditional frameworks. This study compares the forecasting performance of Autoregressive Integrated Moving Average (ARIMA) method against the machine learning architecture Long Short-Term Memory (LSTM), utilizing the monthly export values of Lampung Province. Data from January 2015 to December 2024 were partitioned into a training set (2015-2022) and testing set (2023-2024). For the linear approach, following Box-Cox transformation and first-order differencing, an ARIMA(1,1,0)(1,0,1)[12] model was fitted to the data based on Akaike Information Criterion (AIC) with comparison to other models of ARIMA. Simultaneously, an LSTM network was constructed using a 12-month lookback window and Min-Max scaling. The results indicate that the optimized ARIMA model achieved a lower Root Mean Squared Error (RMSE) of 94,030,344 compared to the LSTM network of 395,566,847 during the 24 months testing window. The ARIMA model effectively captured the underlying linear trends and stable annual seasonality without overfitting the training data. The study concludes that for moderately sized time series, ARIMA remains highly robust and can outperform complex machine learning architectures. Consequently, while neural networks offer advanced capabilities, classical frameworks should remain a primary tool for establishing baseline indicators in regional forecasting. ]]>
Copyrights © 2026
