A comparative analysis of multivariate GARCH and CNN-BiLSTM models for forecasting conditional volatility in financial markets

Abstract

In recent years, there has been significant interest in forecasting volatilities within multivariate frameworks for financial assets. Previous research has utilized the VAR-DCC-GARCH model to explore these relationships, offering valuable insights into market dynamics. This paper presents a novel VAR-CNN-BiLSTM model to forecast the conditional correlation between BTC-USD exchange rates and gold prices. The study aimed to improve the accuracy of volatility forecasting for financial assets by introducing this hybrid approach. The hybrid VAR-CNN-BiLSTM model employed the VAR model to capture the linear features and the deep learning network structure that combines the CNN, to capture the hierarchical data structure and BiLSTM layers to capture the long-term dependencies in the data. Results have confirmed that the VAR-CNN-BiLSTM model can achieve better prediction accuracy than the hybrid VAR-DCC GARCH model, in terms of Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) performance measures. The results of this study further indicate a unidirectional causality from the BTC-USD exchange rate to Gold prices. The findings provide valuable insights for traders, financial analysts, and policymakers aiming to understand and anticipate market behaviors involving cryptocurrencies and traditional assets.