Please use this identifier to cite or link to this item: https://hdl.handle.net/10419/155530 
Year of Publication: 
2016
Series/Report no.: 
CESifo Working Paper No. 6288
Publisher: 
Center for Economic Studies and ifo Institute (CESifo), Munich
Abstract: 
In this study, we compare the out-of-sample forecasting performance of several modern Value-at- Risk (VaR) estimators derived from extreme value theory (EVT). Specifically, in a multi-asset study covering 30 years of stock, bond, commodity and currency market data, we analyse the accuracy of the classic generalised Pareto peak over threshold approach and three recently proposed methods based on the Box-Cox transformation, L-moment estimation and the Johnson system of distributions. We find that, in their unconditional form, some of the estimators are acceptable under current regulatory assessment rules but none of them can continuously pass more advanced tests of forecasting accuracy. In their conditional forms, forecasting power is significantly increased and the Box-Cox method proves to be the most promising estimator. However, it is also important to stress that the traditional historical simulation approach, which is currently the most frequently used VaR estimator in commercial banks, can not only keep up with the EVT-based methods but occasionally even outperforms them (depending on the setting: unconditional vs. conditional). Thus, recent claims to generally replace this simple method by theoretically more advanced EVT-based methods may be premature.
Subjects: 
value-at-risk
extreme value theory
historical simulation
backtest
financial crisis
JEL: 
G10
G11
G17
Document Type: 
Working Paper
Appears in Collections:

Files in This Item:
File
Size





Items in EconStor are protected by copyright, with all rights reserved, unless otherwise indicated.