基于多任务学习的金融时间序列预测研究OA

In the domain of quantitative investment,achieving maximized returns and minimized risks during continuous trading periods is a pivotal task.Accurate forecasting of financial time series plays a crucial role in this process.The concept of multi-task learning is introduced innovatively and an information feedback mechanism is incorporated,significantly enhancing the model's capabilities.The advancement allows the model to not only perform multi-dimensional predictions but also simultaneously increase quantitative trading profits while reducing investment risks.The research initially utilizes actual stock data to predict the price changes a week ahead,incorporating these predictions as a critical factor in the stock backtesting scoring system.The method can improve the accuracy of predicting stock trends to enhance investment returns and effectively reduce risks.Further,by replacing historical data with model-predicted future price changes for a week ahead,the stock scoring system is optimized,thereby improving the accuracy of stock trend predictions.Experimental results demonstrate that compared to predictions based solely on daily price changes,the application of Transformer,LSTM,and IGMTF models significantly enhances prediction accuracy and achieves a noticeable increase in returns per unit of risk.