Stock prices forecast is a hot and challenging topic in financial time series research. It is of great significance for the investors in theirs stock trading, to maximize revenue and to avoid risks by adopting a reasonable and effective forecasting method. A stock prices forecast method based on proximal reinforcement learning which combines proximal policy optimization(PPO) and reinforcement learning(RL), namely PPORL, is proposed, and the forecasting process is regarded as a time series prediction problem. Furthermore, the relative strength index(RSI) and move average of five days(MA5) are also introduced working as a trading strategy, which can automatically capture potential trading points, and avoid trading risks. By comparing the prediction performance and trading decision performance with long short-term memory(LSTM) and recurrent neural network(RNN) models on the SSE composite index(SZZS), the SZSE component index(SZCZ) and the CSI300 index(HS300), and a variety of error evaluation methods are used for quantitative analysis of the prediction results, which shows the effectiveness and robustness of the PPORL in forecasting performance and trading decision.