Nasdaq Constituents Price Prediction Based On LSTM Networks And A New Performance Indicator

The use of deep learning methods to study financial markets has been a hot topic in recent years.People are eager to use it to discover the internal connections of data and grasp the laws or trends of the market,because the correct prediction of stock changes can invest and buy stocks play a very strong guiding role.Deep learning research on the stock market involves economics,investment science,operations research,mathematics,computer science,statistics and other disciplines,and the fluctuations of the stock itself may also be affected by company news and performance,industry performance,investor sentiment,and social media sentiment And economic factors.This has created resistance to stock market research.However,the application of traditional performance indicators such as mean square error has some shortcomings in the stock market.This thesis uses the data of one day of NASDAQ 100 Constituents from January 1,2000 to December 31,2018 as a data set,and selects two typical stocks to conduct research using deep learning methods.In this thesis,wavelet transform and stacked autoencoders are used to reduce noise and construct technical indicators and feature extraction of existing data.This thesis analyzes and demonstrates the ability of stock characterization based on the performance measures of traditional regression models,and proposes a new performance measure based on MAPE,and gives definitions and proofs to illustrate its correctness and advantages.Then this thesis uses this performance measure as a loss function for modeling and training.After processing parameters and preventing overfitting,a good model is obtained.Finally,we use the mean square error as our loss function to build the same Long Short-Term Memory network model.We compare it with the network we built with new performance indicators.The analysis experiments show that both the mean square error and the new indicator can complete the regression prediction of the stock market as a loss function,but the performance of the indicator proposed in this thesis on the same network structure is better by the average absolute percentage error,and the indicator itself can Reflecting the degree of the model’s prediction error,it is possible to compare the strength of a model’s ability to characterize different stocks in a longitudinal direction.Compared with traditional performance indicators,it has more advantages for future research.Finally,this article summarizes its own research and deficiencies,and puts forward the future development of the views of this article.