Short Horizon GARCH and LSTM Volatility Forecasting : A Comparison of Models Using a Value at Risk Framework

Sammanfattning

Accurate forecasting of financial market volatility is critically dependenton reliable models. This study compares the econometric Generalized Autoregressive Conditional Heteroskedasticity (GARCH) and ExponentialGARCH (E-GARCH) models with a Long Short-Term Memory (LSTM) model for one day ahead volatility forecasting using historical data fromfive large Swedish stocks. The models are evaluated using a rolling window approach, combining point forecast accuracy measures and a Valueat Risk framework to assess their applicability in risk management. Theanalysis shows that while the LSTM model achieves the lowest mean absolute error, it systematically underestimates risk whereas the GARCH andE-GARCH models produce more reliable Value at risk breach frequenciesdespite higher forecast errors.