Research On Stock Index Prediction In Combination With Corporate Financial Statement Data

With the continued prosperity and expansion of the financial market,the investor group is increasingly expanding.No matter individual investors or institutional investors,they are more and more eager to obtain the support of financial information,so as to understand the market changes and avoid financial risks to obtain profits.Yet financial market volatility is complex.Want to find regular and accurate prediction has a certain difficulty.At present,researchers have studied a large number of forecasting methods in the field of financial time series prediction,among which machine learning is the most widely used.However,only using the historical data of the stock market itself is not enough to explore the complex rule of stock market change,but also need to use other information related to the change of the stock market,among which paying attention to the related activities of stock market participants is very effective to understand the rule of stock market change.In recent years,researchers have applied information related to the stock financial market,such as stock market comments,stock market news,and corporate earnings reports,to their studies.The emotional characteristics of market participants can be mined from this information to provide more information support for time series prediction.Considering that the company's disclosure information is of great help in the financial decision-making process in the fundamental analysis of stock prediction,this paper combines the company's financial statement data with the historical data of stock index to further improve the accuracy of time series prediction.The stock index prediction method is increasing at the same time,many prediction models are using a single machine learning algorithm to predict the stock index,and weak generalization ability,is not easy to popularize on a large scale.In this paper,ensemble learning is used to improve the generalization ability and prediction performance of prediction models.Major efforts to address these issues include:First,a method for stock index prediction is proposed based on the ratio data of financial statements regularly disclosed by the companies.This method firstly selects the most representative feature from the ratios of numerous financial statements according to the ratio analysis method,which respectively reflect the ratio of the company's debt paying ability,operating ability,growth ability and profitability.Then,the high-dimensional matrix composed of the historical data of a stock index and the data of corporate financial statements is reduced to reduce the influence of data noise and retain the favorable data information for stock index prediction.Finally,considering the different time span of the two kinds of data,a two-channel machine learning model is used for prediction research.The experimental results show that TCCFR model has some advantages over the prediction method using only historical stock index data.Compared with the best results in the comparison model,the MAPE of SSE 50 index is reduced by 0.032 percentage points,and the MAPE of CSI 300 index is reduced by 0.23 percentage points.At the same time,compared with the optimal case integrating other data,TCCFR model has a better prediction effect.MAPE of TCCFR model on SSE 50 index is reduced by 0.076 percentage points,and MAPE of TCCFR model on CSI 300 index is reduced by 0.138 percentage points.Second,a stock index prediction method based on ensemble learning is proposed.Firstly,the improved Ada Boost.R2 algorithm is used to train several TCCFR models iteratively.Then,these TCCFR models are integrated according to the parameters obtained from the iterative training.Finally,the integration model is used to predict the historical data of the stock index.SSE 50 index and CSI 300 index were used as experimental data sets,and LSTM model,CNN-LSTM model,LSTM-Attention model,VMD-LSTM model,TCN model,Bi LSTM model and TCCFR model were used as comparative models for experimental analysis.The experimental results show that the ensemble learning model has better performance in the research of stock index time series prediction.Compare with the TCCFR model.On the SSE 50 index,the RMSE of the Improved-Ada Boost.R2-TCCFR model decreased by 0.311,MAE decreased by 0.557 and MAPE decreased by 0.019 percentage points.On the CSI 300 index,the RMSE of the Improved-Ada Boost.R2-TCCFR model decreased by 2.837,MAE decreased by 2.446 and MAPE decreased by 0.064 percentage points.