Published Paper Details:

This study constructs and tests hybrid neural-statistical models based on deep learning models and GARCH-family models to model asymmetric business cycle dynamics models on fat-tailed distributional assumptions. Having performed large Monte Carlo experiments that produce synthetic business cycle data with different levels of asymmetry, volatility clustering and heavy-tailed innovations. Our hybrid models involving Long Short-Term Memory (LSTM) networks with GJR-GARCH and EGARCH specifications are compared to pure statistical and pure neural work in several performance measures. Findings indicate that hybrid structures have better forecasting and volatility prediction abilities especially when the regime is changing and in tail events. In conditions with asymmetric parameters greater than 0.15, mean absolute percentage error is minimized by 23-31% in comparison to standalone models. Determining that hybrid methods are best performed when neural components are used to learn nonlinear conditional mean dynamics, and GARCH specifications are used to learn heteroskedastic variance dynamics. Finding the implications of our results on business cycle analysis, macroeconomic forecasting and risk management in asymmetric shock and extreme event dominated environments.