The advances in financial theory and power of computer technology have driven the development of quantitative investing. Stock trend forecasting as an important topic in quantitative investment has attracted more and more researchers. The development of deep learning has also led to the emergence of many interesting models in the field. However, most of the existing models only use daily frequency data to predict the stock price fluctuations, lacking the consideration of high-frequency data and relational data. At the same time, it remains a challenge to incorporate multi-granularity data into model training. Based on this, we propose a contrastive pre-training model fusing high frequency data and daily frequency data(HRC), which construct pre-training tasks with a positive and negative sample design in terms of data granularity and stock relationships. In addition to this, we also constructed a label-based training task using the performance of returns to classify stocks.The experiments on real-world stock dataset show significant improvements of our approach over the baselines.