Research On Feature Generation Method For Forecasting Securities Market

Quantitative transaction has significant advantages such as high degree of automation and consistency,high overall return and stability.It is the focus of both investment researchers and traders.Quantitative transaction mainly consists of market forecasting and timing strategy.Market forecasting is an important guarantee for quantifying the effective operation of transactions.Market forecasting is divided into data and models,which can be divided into feature generation and forecasting model construction.The existing related research work mainly focuses on the construction and improvement of predictive model,and which has made great progress;however,there is less attention to feature generation which the progress of feature generation is limited.Feature generation is the basis and premise of the prediction model construction,and its performance directly affects the validity of the prediction model.In view of this situation,this paper focuses on the research of feature generation methods.The specific content is the input information design and feature extraction methods in feature generation,and the construction of LSTM prediction model for verification.The input information directly affects the performance of the generated features.The existing research work mainly uses the high,open,low,and closing prices and trading volume in the securities K line as input data.In addition,some research work has added more statistical information based on the above basic information and achieved more results.However,these research work not only lacks valid basis for information selection,but also the effectiveness of the selected information needs to be further improved.This paper designs the input information through theoretical analysis and experimental verification.First analyze the redundancy of the basic information to perform information redundancy.Secondly,information enhancement is made for lack of information and deficiency.The basic data is supplemented with short-term impulse information and directional trend information.Finally,the rationality of feature combinations is verified through experiments.Feature extraction as an important step in feature generation determines the performance of features.Among them,PCA(principal component analysis)and AE(automatic encoder)are the most widely used,but these methods will lose a lot of useful information for prediction.In order to improve the characteristic performance,this article will take the highly promising automatic encoder as the research object.For the problem that label information is not used effectively,this paper proposes to use paradigm constraints and clustering ideas,and uses label information to constrain the objective function.The encoded feature vector not only effectively express the original input information,but also reduces the dimensions,improves the classification ability,and helps to improve the prediction model.