Forecasting Value-at-Risk using GARCH(1,1) and Neural Networks as Volatility Estimation Methods – A Comparative Study

Abstract: Northvolt was founded in 2015 with the goal to create the world's greenest battery. Today, Northvolt is mainly funded by investors and have suppliers all over the world, which does not come risk free. Northvolt's main risks are prices of commodities, foreign exchange rates and interest rates and they are interested in evaluating the risks standalone and as a portfolio. In this project, the risks are estimated using a Value-at-Risk (VaR) measure called Volatility Weighted Historical Simulation (VWHS) with GARCH(1,1) as volatility estimator. The method is backtested for each risk exposure and compared to the yearly change to see how often the VaR estimate is violated. Northvolt was also interested in possibilities in including machine learning tools in their risk assessment. Therefore, the same VaR calculations was conducted using Neural Networks as volatility estimator for each risk exposure, where the networks were trained to predict realized volatility. The VWHS was found to be effective for the foreign exchange- and interest rates, while the method significantly underestimated the risk for commodities and all-risk portfolio. The Neural Networks performed well during the training, but performed worse overall compared to the VWHS calculated with GARCH(1,1). This is suspected to depend on the volatility estimate used to train the networks. The poor results for the commodities is suspected to depend on inadequate data and an abnormal increase in price. Therefore, we propose evaluating other versions of GARCH to estimate the volatility, both for commodities and the all-risk portfolio. Our final recommendation for Northvolt is to use VWHS using GARCH(1,1) as volatility estimate for the exchange- and interest rates, while investigating other methods to calculate VaR for the commodities.