Optimise the performance of the transformer model in stock forecasting for technology companies by aggregating information from different stocks

The prediction of stock trends has become an increasingly important research area in recent decades. Recently, investors have been looking for correlations by modelling stocks. Existing research mainly models stock correlations by learning models from individual stocks and then mixing different time features. This modelling method only focuses on the correlation of stocks at the corresponding time point. However, in the actual financial market, stocks may be correlated with instantaneous information or information across time. Moreover, the effectiveness of individual information features for different stocks changes in real time as the financial market changes. In addition, due to the high volatility of stocks, the complexity of the model and the efficiency of real-time calculation are very important. This paper proposes a model called FR-transformer based on the above issues. It models instantaneous and cross-time information of stocks. It also adapts the weights of different features according to the current market situation based on historical data. In addition, in order to cope with the rapidly changing financial market, this paper optimises the complexity and optimises it to O (NSlogSdD2). FR-transformer solves the problem of stock modelling complexity to a certain extent by aggregating information within stocks and across different stocks at different times. Experiments have shown that the FR-transformer has a 13.46% improvement in overall performance compared to other classic models.