Dynamic value at risk under optimal and suboptimal portfolio policies

At present, all value at risk (VaR) implementations – i.e., all risk measures of the ‘maximum loss at a given level of confidence’ type – are based on the assumption that the portfolio mix will not change before the VaR horizon. This hypothesis may be unrealistic, especially when the VaR horizon is established by the regulators (BIS). At the opposite, we measure VaR dynamically, i.e., taking into consideration portfolio mix adjustments over time: adjustments do not occur continuously, since they are costly. We allow both optimal rebalancing policies, which entail changing the portfolio mix whenever it is too far from the optimal one, and suboptimal policies, which mean adjusting at pre-fixed dates. We show that in both cases usual VaR measures underestimate portfolio losses, even if the underlying returns are normal. We study the dependence of the misestimate on the VaR horizon, the initial portfolio mix and the risk aversion of the portfolio manager, which in turn determines the frequency of interventions. The bias can be more relevant over one day than over longer horizons and even if the initial portfolio is nearly optimal. We also perform backtesting and estimate a ‘coherent’ risk measure, namely conditional VaR, which confirms the inappropriateness of the usual, static VaR.