| With the continuous development and growth of China’s stock market,the demand for predicting stock trends and making wise investment decisions by investors is increasing.Therefore,many stock analysts,investment advisors,and research institutions have started to study the stock market to provide predictions about stock trends.In recent years,with the development of artificial intelligence technologies such as machine learning and deep learning,more and more scholars and investment institutions have begun to apply these technologies to stock trend prediction tasks and have achieved remarkable results.However,due to the uncertainty of the market,as well as the impact of global economic,political,and environmental factors on stock trends,predicting stock trends is very challenging.In stock trend prediction tasks,multi-scale information is crucial.Traditional stock trend prediction methods typically only focus on single-scale stock price sequences,ignoring information at multiple time scales.Information at different time scales complements and cooperates with each other,contributing to accurate prediction.Therefore,prediction models that use multi-scale information can have better performance,but existing models still have problems such as information loss,poor ability to extract and use multi-scale information.We propose a stock trend prediction model based on the fusion of time domain and frequency domain information(MTIFN).This model combines traditional machine learning and deep learning methods to extract and fuse multi-scale information in the time and frequency domains,with time domain and frequency domain information complementing each other,playing to their respective strengths,and serving stock trend prediction together.The time domain multiscale information extraction layer adopts Temporal Convolutional Network(TCN)with different kernel sizes to capture the long-and short-term dependencies of the corresponding time scales of the stock price sequence,correctly describing the different degrees of fluctuations in the stock price sequence over time.We then introduce self-attention mechanism to effectively fuse time domain multi-scale features from different receptive fields,and use Gated Recurrent Unit(GRU)to further enhance the features to solve the problem of module instability.For frequency domain multi-scale signals,we use the LightGBM model for feature extraction and fusion.Experimental results on two standard datasets show that our model has advantages in capturing multi-scale information.The proposed MTIFN model in this paper combines the advantages of traditional machine learning and deep learning through multi-scale feature extraction and fusion,enabling the model to have better feature extraction capabilities and prediction performance.The model combines multi-scale information in both frequency and time domains,using moving averages combined with TCN to extract multi-scale time domain features of stock price sequences,and adopting attention mechanisms and GRU to effectively fuse and enhance the multi-scale temporal features.At the same time,MTIFN uses Discrete Wavelet Transform(DWT)combined with LightGBM to extract and fuse the frequency domain multi-scale features of stock price sequences,achieving effective extraction and fusion of multi-scale frequency domain information.These innovative methods provide new ideas for multi-scale informationbased stock trend prediction models.Finally,based on the above research,this paper designs and implements a stock trend prediction system to help users make more informed investment decisions,improve their investment success rate,and achieve wealth preservation and appreciation. |
