Enhancing Time Series Forecasting with an Ensemble Model of Neural Networks and HGARCH
Abstract
In this paper, we propose a novel ensemble model that integrates a Neural Network (NN) with a GARCH-incorporated loss function and an HGARCH (Hierarchical Generalized Autoregressive Conditional Heteroskedasticity) model to improve forecasting accuracy in time series data with volatility clustering. This approach is particularly useful for financial data, healthcare demand prediction, or call center workload forecasting, where volatility is a key feature of the data.The neural network serves as the first ensemble member and is known for its strength in capturing complex, non-linear relationships. By embedding a GARCH penalty in its loss function, the NN is better equipped to manage volatility patterns that are typically modeled by GARCH. This allows the network to account for dynamic fluctuations in the variance of residuals, enhancing its ability to forecast in volatile environments.The second ensemble member is an HGARCH model, which captures long-term volatility persistence in the data. Combining these models in an ensemble framework allows for improved accuracy, as the NN addresses non-linearities, while the HGARCH model robustly handles volatility. This hybrid approach offers a powerful tool for time series forecasting where volatility and persistence are crucial factors.