Hybrid ML-Econometric Models for Volatility Spillovers

Risk managers and traders must understand the mechanisms of volatility spillovers from one market to another due to the implications on profits and stock prices. The use of econometric models such as the Generalised Autoregressive Conditional Heteroskedasticity (GARCH) and heterogeneous autoregressive (HAR) to predict volatility spillover effects is limited by their requirement for linearity and their inability to detect rapid changes in volatility within the market. Combining econometric models with machine learning algorithms to generate hybrid models is an effective strategy to improve their performance in detecting volatility spillovers between markets. This study aimed to compare how the HAR model performed relative to the hybrid LSTM-HAR model when predicting volatility spillover effects from the WTI crude oil market to the agricultural stock market in the US. Data was obtained from the FRED database in the period 2015-2025, where monthly global prices were sourced for WTI crude oil, wheat, cotton, corn, and soybeans. The generated HAR model showed the volatility spillover existence from the oil to the soybeans market, which was explained to arise from the disruptions from the global COVID-19 pandemic and the Russia-Ukraine war. However, the HAR model outperformed the LSTM-HAR model for all commodities and was explained to arise from the small dataset available, which was dominated by linear data.