A nonlinear model-predictive control strategy is developed to maintain the superior-to-steady-state performance of a periodically forced chemical reactor. The performance of the predictive con troller is investigated in the presence of measurement disturbances and parametric uncertainty. It is also shown that statistically inferred input-output models can be a substitute whenever detailed fundamental models are not available. A nonlinear autoregressive polynomial model based on observed plant data is built and incorporated into the control scheme. The catalytic oxidation of ethylene in a periodically-forced, continuous stirred-tank reactor is considered as the test case.