Time Series Prediction Model Based On Multi-layer Symbolic Pattern Network

Time series are prevalent in many fields of production life,and it is of positive significance to mine the potentially valuable information in them.Using data mining techniques to predict time series is a hot research topic in time series data analysis.The common time series forecasting methods are mainly traditional econometric models and artificial intelligence-based machine learning methods.Due to the highdimensional nature of the original time series and the large amount of noise they contain,traditional statistical methods cannot capture the dynamics of nonlinear data,and the prediction models based on neural networks and support vector machines suffer from poor generalization ability,susceptibility to interference and weak interpretation.In this paper,we not only consider the volatility characteristics of the time series itself,but also considers the mutual influence between related sequences with time lag,constructs a Multi-layer Symbolic Pattern Network(MSPN)based on causal entropy,and combines it with a Long Short-Term Memory(LSTM)neural network prediction method In combination,a novel time series hybrid forecasting model is proposed.The model predicts the fluctuation trend of the time series by extracting the effective information of the multi-layer network topology,and then uses the predicted fluctuation trend to optimize the selection process of the neural network training set,reduce the training complexity,and effectively improve the time series prediction accuracy.The main work and contributions of this paper are as follows:(1)A multilayer symbolic pattern network based on causal entropy is proposed.The model mainly solves the following two problems: the noise reduction problem of time series and the problem of high-dimensional characteristics of time series.According to the characteristics of the time series,the time series is mapped to the symbol space,and the maximum entropy principle is used to optimize the threshold selection in the symbolization process,and the discretization reduces the influence of the time series noise on the prediction model.By calculating the causal entropy between different symbol pattern sequences,the causal structure between the sequences is obtained,and a multilayer symbol pattern network is established.(2)A time series fluctuation trend prediction model based on multi-layer symbolic pattern network is established.Based on node degree and connection weight,the‘transfer rate' and ‘acceptance rate' are proposed to describe the conversion relationship between nodes in the network.A rolling framework is used to configure the forecasting model so that the model accepts new data and closes older historical data.Taking the WTI international crude oil futures price volatility trend as an example and considering the impact of supply and demand and the three major U.S.stock indexes on oil price volatility,experimental data show that the multilayer symbolic pattern network forecasting model has higher directional forecasting accuracy and value-added forecasting accuracy than the single-layer network.(3)An improved LSTM-based time series forecasting model is proposed.The MSPN is combined with LSTM to predict volatility further with volatility trend prediction.By optimizing the selection process of the training set in the neural network prediction model,only the volatility fragments with consistent fluctuation trends in the historical data are input into the neural network training.The results prove that the prediction accuracy of the MSPN-LSTM hybrid prediction model is better than that of the single LSTM neural network prediction.In summary,this paper establishes a novel hybrid model for time-series prediction under the perspective of multilayer networks and considering the correlation of time series,which provides a new idea for time-series data mining.