Adaptively weighted combinations of tail-risk forecasts
Conference
64th ISI World Statistics Congress - Ottawa, Canada
Format: IPS Paper
Keywords: expected shortfall, forecasting, risk_prediction, value-at-risk
Session: IPS 433 - High-Dimensional Financial Time Series
Tuesday 18 July 10 a.m. - noon (Canada/Eastern)
Abstract
Several methods have been proposed in the literature for forecasting daily Value at
Risk (VaR) and Expected Shortfall (ES), with tail risk models generally classified
into three main categories: parametric, semi-parametric, and non-parametric.
However, given the various sources of uncertainty associated with the data, market
conditions, estimation methods, and exogenous variables that can have a significant
impact on the dynamics of tail risk, there is no single model that consistently outperforms
all the others. To mitigate the impact of model misspecification and improve
the predictive accuracy of individual models, we investigate the use of two forecast
combination approaches. In the proposed approaches, the weight of the most accurate
set of predictors is determined adaptively according to strictly consistent loss
functions for VaR and ES employed in the Model Confidence Set procedures. Our
findings reveal that combinations of VaR and ES forecasts result in higher predictive
accuracy over a wide range of individual competitors.