Since the parameters of an autoregressive conditional heteroskedasticity (ARCH) process must be non-negative, inference on ARCH parameters can be improved by using inequality constrained estimation. In this paper, we extend this principle to the problem of ARCH lag order selection. We show that...