Research And Implementation Of Stock Market Prediction Based On Multi-source Data Fusion

Traditional stock prediction methods mainly rely on quantitative data.With the advent of the era of big data,researchers begin to extract effective features(such as events and emotions)from the network to improve prediction performance.Fusing heterogeneous multi-sourced data to achieve superior performance is a challenge.Most of the existing solutions adopt conventional machine learning models.Limited by model capacity,complex interactions between multi-sourced data may not be effectively modeled.For this reason,the paper proposes an effective fusion model for the multi-sourced data of the stock market to improve the accuracy of stock prediction.We propose a stock prediction method based on feature extraction and feature fusion of multi-sourced data.Firstly,we use wavelet transform to capture the high-frequency and low-frequency fluctuation trends of quantitative time-series data.We use discrete wavelet transform decomposition to obtain high-and low-frequency subsequences,and then use wavelet neural networks to fit the subsequences separately,which can fully mine the time series information.Secondly,a multi-tasking model is modeled for stock-related text data to extract high-quality text features.Based on BERT,a single-tasking model for news representation and sentiment analysis of social network posts is trained.Then the output of the single-tasking model is combined with real labels to train a multi-tasking model.Compared with knowledge distillation and standard multi-task training methods,this method can improve the quality of text features,thereby improving the prediction ability of the model.Next,the quantitative and textual features are fused to predict stock movements.Then,the paper proposes a multi-sourced data fusion model based on feature interactions.The low-and high-order feature interactions are modeled with the factorization machine and the deep neural network respectively,and stock correlations are incorporated into the model through the attention mechanism.Finally,evaluations on the China's stock market data from the year 2015 to 2017 show the effectiveness of our model.