This paper presents a statistical approach to unknown word type prediction for a deep HPSG grammar. Our motivation is to enhance robustness in deep processing. With a predictor which predicts lexical types for unknown words according to the context, new lexical entries can be generated on the fly. The predictor is a maximum entropy based classifier trained on a HPSG treebank. By exploring various feature templates and the feedback from parse disambiguation results, the predictor achieves precision over 60%. The models are general enough to be applied to other constraint-based grammar formalisms. 1