Asymptotic behaviour of the CUSUM of squares test under stochastic and deterministic time trends

We undertake a generalization of the cumulative sum of squares (CUSQ) test to the case of non-stationary autoregressive distributed lag models with quite general deterministic time trends. The test may be validly implemented with either ordinary least squares residuals or standardized forecast errors. Simulations suggest that there is little at stake in the choice between the two in the unit root case under Gaussian innovations, and that there is only very modest variation in the finite sample distribution across the parameter space.