A quantile-data mapping model for value-at-risk based on BP and support vector regression

Abstract

A novel "Quantile predicting Return". method (Q→R) is presented to predict VaR (Value-at-Risk). In the paper, a nonlinear mapping between quintile and VaR is constructed by using neural networks instead of traditional statistical methods. Two related models are proposed. One is QDMN (Quantile-Data Mapping Network) based on BP and the other is SVR-QD (Support Vector Regression Quantile- Data mapping) based on SVM. There is no assumption for distribution in both proposed models. The operation mode and the reasonableness of measuring VaR using the two models are analyzed. QDMN and SVR-QD are applied to predict VaR in Shanghai Stock Exchange. Comparisons of the proposed methods with Monte Carlo approach are performed. Numerical experiments based on Basle traffic lights, Proportion of Failures and exception computing show that the two new models are superior to Monte Carlo approach based on a certain assumptions in accuracy. © Springer-Verlag Berlin Heidelberg 2005.