A COMPARATIVE STUDY OF ECONOMETRIC AND DEEP LEARNING MODELS FOR FORECASTING EXCHANGE-TRADE FUND PRICES IN THAI MARKET

Abstract

This study evaluates and compares the short-term forecasting performance of ten Exchange-Traded Funds (ETFs) in the Thai market using a comprehensive suite of econometric and deep learning models. The Mean Absolute Error (MAE) served as the primary performance metric. The results indicate that the Univariate LSTM model consistently outperformed all competing specifications in terms of predictive accuracy across nearly the entire ETF sample. In particular, the Univariate LSTM with a 30-day lookback window, relying solely on each ETF’s own recent real prices proved to be the most effective and parsimonious forecasting structure. Beyond forecasting, the study integrates the superior model outputs into a portfolio allocation and hedging framework, yielding valuable practical insights. The optimal portfolio was concentrated in CHINA and DIV, reflecting their high risk-adjusted returns driven by strong predicted performance and low correlation. This portfolio achieved an expected daily return of 0.218% with a daily volatility of 0.345%. The subsequent hedging analysis employed ABFTH as the primary hedging instrument, demonstrating the highest Hedging Effectiveness (HE) for BMSCITH (28.1%), while CHINA and UBOT exhibited the lowest HE values. These findings highlight the potential of LSTM-based ETF forecasts for enhancing portfolio optimization and hedging strategies within the Thai market.