Bootstrapping tests for conditional heteroskedasticity based on artificial neural network

This paper deals with bootstrapping tests, based on the LM statistic and on a neural statistic, for detecting conditional heteroskedasticity in the context of standard and non-standard ARCH models. Although the tests of the literature are asymptotically valid, they are not exact infinite samples, and suffer from a substantial size distortion, and has to be accounted for. In this paper, we propose to solve this problem using parametric and nonparametric bootstrap methods, based on simulation techniques, making it possible to obtain a better finite-sample estimate of the test statistic distribution than the asymptotic distribution.