Perbandingan Performa MSGARCH, LSTM, dan Hybrid MSGARCH-LSTM pada Peramalan Data Deret Waktu yang Mengandung Heteroskedastisitas

Abstract

Volatility forecasting is crucial for estimating potential portfolio losses, particularly in cryptocurrency markets like Bitcoin, which exhibit high and irregular price fluctuations. Models from the Generalized Autoregressive Conditional Heteroscedasticity (GARCH) family, including Markov Switching GARCH (MSGARCH), are widely used to handle heteroscedastic data and capture regime changes. Meanwhile, Long Short-Term Memory (LSTM) is effective for modeling nonlinear and complex patterns in financial time series. This study proposes a hybrid MSGARCH-LSTM model by incorporating MSGARCH predictions as additional input to the LSTM. The model is evaluated using simulated data resembling Bitcoin's characteristics, with Heteroscedasticity Mean Absolute Error (HMAE) as the primary metric, and analyzed using ANOVA and Tukey's post-hoc test. The results identify four superior hybrid configurations, all of which significantly outperform the standalone MSGARCH and LSTM models. Based on the characteristics of Bitcoin data, the MSGARCH (2-regime with sged error distribution)-LSTM model is selected for empirical analysis. This model achieved an HMAE of 0.3197 and an HMSE of 0.2088, with accuracy improvements of 61.20% and 83.50% compared to the standalone MSGARCH model. These findings indicate that the hybrid MSGARCH-LSTM model improves volatility forecasting accuracy in highly volatile cryptocurrency markets.