Using temporal knowledge graph to predict future events has become a hot topic in recent years. Generally, historical information has important impact on the prediction of future events. How to use the historical information has drawn significant attention. In this paper, a new model HSRMNet is proposed for improving the prediction performance, and this model is based on a historical statistical reward mechanism. Specially, two modes are designed to capture history information effectively. The Global Mode use the frequency of historical events to simulate the evolution of events by statistical methods. The Standard Mode divides historical events into two categories according to whether the events have occurred or not. After obtaining the frequency and category information about the events, the two modes jointly learn the probability distribution of entities. The proposed model is evaluated on two datasets and it’s performance gain is demonstrated by the experimental results. On YAGO, the hit@10 raw metric reaches 77.65%.