This paper compares the power in small samples of different tests for conditional heteroscedasticity. Two new tests, based on neural networks, are proposed: the main interest in them arises from the fact that they do not require the exact specification of the conditional variance under the alternative.