HYBRID VECTOR AUTOREGRESSIVE AND LONG SHORT TERM MEMORY MODEL FOR PREDICTING ECONOMIC GROWTH INDICATORS IN INDONESIA: A COMPARISON OF ADAM, NADAM, AND RMSPROP OPTIMIZATION METHODS

Abstract

This study aims to compare the performance of three optimization methods—Adam, Nadam, and RMSProp—in forecasting monthly economic indicators of Indonesia, namely the Consumer Price Index (CPI), Inflation, and Gross Domestic Product (GDP), using a hybrid Vector Autoregressive–Long Short-Term Memory (VAR–LSTM) model. The analysis begins with Vector Autoregression (VAR), where VAR(4) is selected as the best model based on the lowest Akaike Information Criterion (AIC) value of 1.075. Significant parameters from the VAR model are then used as input variables for the LSTM to enhance forecasting accuracy. The experimental results show that all three optimization methods generate similar prediction patterns, with forecasted values closely tracking the actual data. Nevertheless, the best optimizer differs across variables: Nadam performs best for CPI with a Root Mean Square Error (RMSE) of 0.4996, Adam yields the best performance for Inflation with an RMSE of 0.676, and RMSProp performs best for GDP with an RMSE of 1.288. Despite these variations, the overall forecasting performance of the three methods is comparable. These findings indicate that the VAR–LSTM approach can effectively capture the dynamic patterns of multiple economic variables and that the choice of optimization method should be aligned with the specific characteristics of the data, considering both accuracy and computational efficiency.