Stock Price Prediction Based On Multilevel Wavelet Decomposition Interaction Network

The stock market is an important component of the national financial market,playing a positive role in providing investment channels,raising funds,optimizing fund allocation,and promoting economic development.However,the stock market is complex and highly risky,which may result in significant losses for market participants.Therefore,the establishment of an effective stock prediction model is of great significance for investors to obtain higher expected returns at a reasonable risk level and for enterprises to develop scientific financing plans.Stock price prediction is a classic problem in the interdisciplinary fields of finance,mathematics,economics,and computer science.Stock data usually exhibit nonlinearity,nonstationarity,and highly complex characteristics,making stock price prediction tasks highly challenging.In recent years,deep neural networks have shown outstanding learning ability and achieved significant results in stock price prediction tasks.However,many existing deep learning solutions only explore the temporal information or lack effective modeling of the frequency domain information,resulting in the inability to effectively utilize both the temporal and frequency domain information of the data.Additionally,existing methods overlook the autocorrelated errors caused by missing effective information data in stock price prediction tasks,which will damage the maximum likelihood estimation assumption and weaken the model’s performance.Therefore,this paper proposes an end-to-end learning framework for stock price prediction called the Multilevel Wavelet Decomposition Interaction Network(MWDINet).The MWDINet first uses multi-scale interaction decomposition modules(MWDIBlock)and HMA-Block to extract the frequency domain information and complete temporal information of the data,respectively.In the MWDI-Block,the maximum overlap wavelet transform(MODWT),a traditional signal processing method,is seamlessly embedded into the deep learning framework,called DMODWT.DMODWT can automatically extract the frequency domain information from the data,and the wavelet coefficients of MODWT can be fine-tuned by deep learning to adapt to data changes.In the HMA-Block,this paper improves the commonly used Hull moving average(HMA)in the industry into a learnable deep learning module to learn the changes in different periods and markets.Finally,inspired by the research on autocorrelated errors correction in linear models in econometrics,this paper designs a deep difference module to correct the model’s autocorrelated errors and improve the model’s prediction performance.This paper uses four evaluation metrics to assess the performance of the model from multiple perspectives and finds that the MWDINet framework performs better compared to current state-of-the-art models in experiments on seven stock index datasets and six individual stock datasets.Meanwhile,this paper demonstrates the effectiveness of each of the proposed modules through a large number of ablation experiments.In conclusion,this paper presents an innovative end-to-end deep learning forecasting framework.MWDINet can make full use of the time and frequency domain information of the data and correct for autocorrelated errors,effectively improving the accuracy and robustness of stock price prediction and providing a new idea of thinking for stock price prediction.