Cold rolling mill process in steel works uses stands of rolls to flatten a strip to a desired thickness. Most of rolling processes use mathematical models to predict rolling force which is very important to decide the resultant thickness of a coil. In general, these mathematical models are not flexible for variant coil types and cannot handle various elements which is practically important to decide accurate rolling force. A corrective neural network is proposed to improve the accuracy of rolling force prediction. Additional variables-composition of the coil, coiling temperature and working roll parameters-are fed to the network. The model uses an MLP with BP to predict a corrective coefficient. The test results using 1,586 process data collected at POSCO in early 1995 show that the proposed model reduced the prediction error by 30% on average.