Economic Data Forecasting Using Hybrid Vector Autoregressive-Long Short Term Memory Model

Abstract

Fluctuations in stock prices and the Rupiah exchange rate create uncertainty for investors in their investment decision-making. One approach to minimizing investment risk is through forecasting utilizing a reliable method. Traditional forecasting models, such as Vector Autoregressive (VAR), are effective in capturing linear patterns but struggle to accommodate more complex patterns. On the other hand, modern deep learning models like Long Short Term Memory (LSTM) can handle dynamic patterns (both linear and nonlinear) but have limitations in consistently processing simultaneous relationships among variables. This research aims to develop a Hybrid forecasting model by integrating VAR and LSTM approaches to predict the Composite Stock Price Index (IHSG) and the Rupiah exchange rate against the US Dollar. The Hybrid VAR-LSTM model leverages the strengths of VAR for linear patterns and LSTM for nonlinear patterns in multivariate time series data. Using the OSEMN framework (Obtain, Scrub, Explore, Model, iNterpret), this study ensures a systematic and comprehensive analysis process. Data from January 2004 to December 2023 is used to build the model, while data from January to July 2024 is used for validation. The model's performance is evaluated using Mean Absolute Error (MAE) to measure the prediction error. The results indicate that the Hybrid VAR-LSTM model significantly improves prediction accuracy compared to the VAR model used independently, as evidenced by a reduction of 42.72 points in MAE for IHSG predictions and 55.82 points for Rupiah predictions.
 
Keywords — Composite Stock Price Index; Hybrid VAR-LSTM; OSEMN Framework; Rupiah Exchange Rate; Time Series Forecasting.