This paper has evaluated a number of available VaR models, such as, variance-covariance/normal (including Risk-Metric approach), historical simulation and tail-index (Hill's estimator) based method for estimating VaR for a number of selected GOI bonds and representative portfolios of GOI bonds for banks and PDs. Competing VaR methods/strategies are evaluated through backtesting and assessment of two typical loss-functions. Empirical results we present are quite interesting. It is seen that normal methods (including Risk-Metric approach) generally underestimate VaRs.
On the other hand, VaR models based on HS and tail-index (using Hill's estimator) are quite good, though the later produces slightly more conservative VaR estimates. But when we look at the loss-functions, tail-index method appears to give the least magnitude/amount of excess loss (i.e., loss over estimated VaR). These results, however, are tentative. One needs to experiment with alternative sizes of rolling sample to check the robustness of the results. Future research may also investigate on appropriate formulation of loss-function while evaluating VaR models.
|
