Backtesting Value-at-Risk Models: A Multivariate Approach

The purpose of this paper is to develop a new and simple backtesting procedure that extends the previous work into the multivariate framework. We propose to use the multivariate Portmanteau statistic of Ljung-Box type to jointly test for the absence of autocorrelations and cross-correlations in the vector of hits sequences for different positions, business lines or financial institutions. Simulation exercises illustrate that this shift to a multivariate hits dimension delivers a test that increases significantly the power of the traditional backtesting methods in capturing systemic risk: the building up of positive and significant hits cross-correlations which translates into simultaneous realization of large losses at several business lines or banks. Our multivariate procedure is addressing also an operational risk issue. The proposed technique provides a simple solution to the Value-at-Risk(VaR) estimates aggregation problem: the institution's global VaR measure being either smaller or larger than the sum of individual trading lines' VaRs leading to the institution either under- or over- risk exposure by maintaining excessively high or low capital levels. An application using Profit and Loss and VaR data collected from two international major banks illustrates how our proposed testing approach performs in a realistic environment. Results from experiments we conducted using banks' data suggest that the proposed multivariate testing procedure is a more powerful tool in detecting systemic risk if it is combined with multivariate risk modeling i.e. if covariances are modeled in the VaR forecasts.