EcoSta 2022: Start Registration
View Submission - EcoSta2022
A0970
Title: On the construction of neural networks for value-at-risk forecasting Authors:  Yi Jiang - Macquarie University (Australia)
Wilson Chen - The University of Sydney (Australia) [presenting]
Richard Gerlach - University of Sydney (Australia)
Abstract: In the context of forecasting Value-at-Risk (VaR) using quantile regression models, much research effort has been devoted to specifying the functional forms of quantile dynamics that relate the present period VaR to a set of explanatory variables available at the previous period. Typical choices of predictors include past returns and summaries of past returns observed at a higher frequency. Neural networks aimed at minimising the quantile loss have been proposed as generalisations of quantile regression models and applied to produce VaR forecasts. However, little attention has been paid to analysing the architecture of quantile regression neural networks and their impact on VaR forecasts, especially when higher-frequency returns are provided as inputs. We empirically assess performance in forecasting daily VaR for various designs of feedforward and recurrent networks of different input types. We also compare the performance of neural network models with that of the more traditional CAViaR- and GARCH-type models.