In the chemical process where the molecular weight of a polymer is changed by thermal decomposition in an extruder, the need for frequent sampling to measure the melt-flow-rate (MFR) places a large burden on the operator. To resolve this problem, we built a soft sensor for MFR prediction based on Just-In-Time modeling technique. For practical use, this work improves the accuracy of prediction by two proposed methods. The first is to consider a statistical model that combines a physical model. The second is to make effective use of variables that contain missing values and outliers. The system developed was confirmed to be operable without the degradation of model accuracy.