Value-at-risk - models in extreme market conditions

Abstract

Value-at-Risk has widely been accepted as the standard measure of market risk in the past twenty years. Nonetheless, VaR models are useful insofar they forecast market risk with sufficient accuracy. The excessive number of losses over VaR limits observed during the recent financial crisis of 2008 revealed that VaR might not necessarily be an accurate measure of risk during times of market uncertainty.

The objective of this thesis is to evaluate the performance of various VaR models in high volatility market conditions. The research question could be formed as: are VaR models sufficiently accurate in high volatility market conditions to justify their use as the standard market risk metric? The research question should be given extra attention as VaR is part of the current financial regulation.

The theoretical part of the thesis presents the basic VaR models, as well as some of the formal backtests which are used to evaluate the accuracy of the computed VaR estimates in the empirical part of the thesis. In addition, some of the main critique toward VaR will be reviewed. The empirical part of the thesis concentrates on evaluating the accuracy of the presented VaR models during the years 2007-2009 with data of two stock indices. Evaluation of model performance is based on backtesting the frequency as well as the independence of the VaR exceptions.

The results show that most VaR models in this thesis underestimated market risk during the backtesting period of 2007-2009. In particular, poor performance was observed with parametric VaR models, and with the basic Historical Simulation. The results indicate that the parametric models are highly sensitive regarding the assumptions behind the model, which can be considered as the main drawback of the method. On the basis of the results, combining a sophisticated volatility estimation technique with the Historical Simulation provides an accurate model for risk forecasting during times of high market volatility. Therefore, the future reliance on VaR should be based on these type of models.