Research On The Prediction Of Bulk Futures Price Trend Based On Time Series Classificatio

At present,it is a highly information-based era,and data has made a leap of magnitude in terms of quantity and dimension compared with the past.In a large number of data,a considerable part of them belong to time series data in a broad sense,that is,there is a time relationship before and after the data.The research on time series has always been an important and challenging task.With the development of the times,time series analysis has also derived from relatively traditional research,such as meteorological data prediction,ECG classification,and other more challenging branch fields,such as text translation,emotion classification.However,with the development of technology,more effective research methods have emerged.In order to solve the challenges in time series classification,many methods have been proposed in recent years.Most of these methods are based on traditional machine learning technologies,such as decision trees,support vector machines,random forests,etc.However,these methods often ignore the time correlation and trend of time series data,resulting in poor performance of the model.Recently,with the development of deep learning technology,many time series classification methods based on deep learning have been proposed,but the disadvantages of poor interpretability have become increasingly prominent.This thesis summarizes many methods for time series analysis,and constructs a time series classification model from two aspects of the interpretability and classification accuracy of time series classification.The model uses functional data analysis and Shapelet method to extract time series features,and then combines the extracted features with the original data as the input of bidirectional long short term memory networks(BiLSTM)for time series classification.This thesis uses futures price data to make an empirical analysis of the effect of the model,and compares it with other existing time series classification methods.The main research contents of this thesis are as follows:(1)The forecast of futures price is divided by the amplitude of rise and fall,and the forecast problem is transformed into a classification problem.For the research of financial data prediction,previous research often predicted the specific value of the future trend one by one.This research method is not suitable for some situations.This method is used to make the goal clearer.(2)Extract the characteristics of time series data.In order to consider the interpretability of the model,this thesis uses functional data analysis and Shapelet methods to extract the global features and distinctive local features of the time series itself,and obtains two feature sequences.From the point of view of samples,functional principal component analysis is carried out for each futures price data,and its global characteristics are abstracted into a very short data vector of principal component scores;The Shapelet method analyzes all fragments of data with different labels from the perspective of classification labels,and learns differentiated fragments from them as local features of this label type data.(3)Use BiLSTM to classify data.Data containing global and local features of time series are used for classification.In model construction,Soft max is used as classifier,Dropout layer is introduced in training,and batch normalization,gradient clipping and early stop are adopted to limit the over fitting.After training in training set,the model constructed in this thesis has an accuracy rate of 82%,a weighted average recall rate of 82%,a weighted average precision of83%,and a weighted average F1-score of 83% on 60% testing machines.(4)Comparison with other time series classification models.This thesis also uses random forest,CNN,Shapelet,LSTM and BiLSTM to classify the original futures data.After the classification result is obtained,it is compared with the effect of the model built in this thesis.It is found that the model built in this thesis is superior to other models in effect,and the classification has a certain degree of interpretability.