Stock trend prediction is a key issue in the field of financial. The traditional algorithms realize the prediction by matching the historical data similar to the current stock trend. However, they do not fully consider the time series characteristics of stock data, and the prediction results are unsatisfactory. With the development of deep learning, the time series model represented by LSTM has effectively improved the accuracy of stock trend prediction, but the profitability is still far from expected. One of the main reasons is that LSTM cannot directly extract useful features from stock data. In addition, stocks are also influenced by similar stocks, and the traditional similarity algorithms are complicated to determine the influences. A more suitable similarity algorithm is necessary to describe the influence of similar stocks. To solve the problems above, we propose a new stock trend prediction framework based on feature embedding and stock similarity named FESS. FESS embodies three creative aspects: (1) An encoding algorithm of stocks is proposed to simplify the complexity and protect the time series characteristics of the price data. (2) Transformer is introduced to build better feature embeddings of stocks. (3) A learnable stock similarity algorithm is proposed to describe the stock similarity more reasonably. The back-testing results on a large number of historical stock data in the Chinese A-share market show that our method is superior in prediction effect and profitability.