Forecasting Crude Oil Volatility with a GARCH–Time-MoE Framework

Accurate volatility forecasting is essential for financial risk management and Value at Risk (VaR) estimation. Traditional econometric models such as GARCH capture volatility clustering in financial time series but may struggle to adapt to changing market regimes. This research proposes a hybrid volatility forecasting framework that combines GARCH type models with a Time based Mixture of Experts (Time MoE) neural network. Forecasts from GARCH, exponential GARCH (eGARCH), and threshold GARCH (tGARCH) models are treated as expert predictions, while a neural gating network dynamically assigns time varying weights to each model. The hybrid framework produces one day ahead volatility forecasts for crude oil log returns, which are used to compute 95% and 99% VaR using Parametric and Filtered Historical Simulation methods. Model performance is evaluated during the 2007 to 2009 Financial Crisis and the 2020 to 2021 COVID recession using coverage tests and loss functions. Results show that the