Improving expected tail loss estimates with neural networks

Expected tail loss (ETL) and other ‘coherent’ risk measures are rapidly gaining acceptance amongst risk managers due to the limitations of value‐at‐risk (VaR) as a risk measure. In this article we explore the use of multilayer perceptron supervised neural networks to improve our estimates of ETL numbers using information from both tails of the distribution. We compare the results with the historical simulation approach to the estimation of VaR and ETL. The evaluation results indicate that the ETL estimates using neural networks are superior to historical simulation ETL estimates in all periods except for one, and in that case the historical ETL is slightly superior. Overall, therefore, when the whole period is considered, our results indicate that the network estimates of ETL are superior to the historical ones. Finally, one of the most interesting results of the study is the fact that the neural networks seem to indicate that VaR and ETL (as a function of VaR itself) are dependent not only on the negative returns observed, but also on large positive returns, which indicates that too much emphasis on losses could lead us to overlook important risk information arising from large positive returns. Copyright © 2005 John Wiley & Sons, Ltd.