Abstract: It is quite well known that Engle (1982)’s ARCH test is not robust to misspecification of the conditional mean model. Especially, when the conditional mean is nonlinear process and we model it by a linear model, it rejects the true null hypothesis too often and thus causes the serious size distortion. This paper proposes a new robust ARCH test using the Yeo-Johnson transformation in which the dependent variable is transformed to deal with nonlinearity. This alternative model specification yields a new GARCH model, i.e. Yeo-Johnson (YJ) GARCH by product. The Monte Carlo simulation results show that the robust ARCH test is significantly superior to Engle (1982)’s ARCH test in terms of the size and power. Empirical application to the returns of Shanghai Composite Index illustrates that the goodness-of-fit for YJ-GARCH model is clearly better than the linear GARCH model.