An exact test for serial correlation in regression models is proposed based on the fact that under classical assumptions and including normality for disturbances, the successive quantities are independent N(0, σ2). For various types of data the power of the proposed test compares favorably with that of the BLUS test, but neither test is as powerful as the full Durbin-Watson procedure in which an approximate significance point is calculated when the bounds test is inconclusive. However, simplicity of the proposed test makes it attractive as a practical procedure.