CPI Prediction Research Based On Machine Learning Theory

Since the reform and opening up,thanks to a relatively stable internal and external environment,our country’s economy has made great achievements,laying a solid economic foundation for high-quality development and common prosperity.However,the economic and political situations at home and abroad are becoming increasingly complex and changeable,and uncertain factors are increasing,especially the long-term severe challenges of the global pandemic of the new coronavirus,population aging,anti-globalization,and key technology bottlenecks,which have greatly improved.risk of my country’s economic operation.Therefore,how to stabilize growth and prevent risks has become an urgent problem to be solved at present.The prediction of macroeconomic indicators is of great significance in predicting and responding to economic risks,guiding the adjustment of industrial structure,and promoting high-quality economic development.It is more critical to achieve"stabilizing growth,preventing risks,and promoting reforms".Therefore,the prediction of macroeconomic indicators has also become a hot topic of current academic research.However,with the increasing complexity of economic operation,the continuous emergence of new economic formats,changes in household consumption patterns,the low frequency of current indicators,and the limitations of forecasting methods,the predictability of macroeconomics has been significantly reduced.Efficiency and synergy become more difficult.Therefore,exploring more efficient,accurate and lowcost prediction methods has become a difficult point in academic research.The rise of artificial intelligence and the development of big data technology have provided the possibility to efficiently explore the prediction of macroeconomic indicators,and provided ideas for solving the problems of timeliness,accuracy and complexity of prediction of macroeconomic indicators.Among them,CPI,as an important macroeconomic indicator,has important value in "stabilizing growth,preventing risks and promoting reform",so its prediction is the focus of modern economic theoretical research and practice.Accordingly,based on machine learning theory,this paper conducts an exploratory study on the prediction method of macroeconomic indicators represented by CPI.In order to solve the problem of CPI prediction,this paper first designs a text mining technology based on the human-machine interactive TF-IDF algorithm and the BERT model,extracts CPI prediction keywords from two dimensions of breadth and depth,and performs feature expansion on the original CPI prediction seed keywords..Then,considering the advantages and disadvantages of LSTM for regression prediction,the LSTM neural network structure was improved to encapsulate MultiRepresentational Attention and Soft-Attention,and a two-layer Attention LSTM neural network prediction model was constructed.Further empirical research is carried out on CPI prediction in different periods of long,medium and short periods,and the feasibility of text mining technology and improved models for predicting CPI in different periods is explored.By reviewing and sorting out the literature,this paper conducts a series of theoretical derivation and empirical research.The main innovations are reflected in the following four aspects.First,it enriches the research on the construction of the CPI prediction indicator system.According to the progress made by the existing literature in this field,the definition of indicator categories and boundaries directly determines the scope of forecast variable statistics,and further determines the accuracy of macroeconomic indicator forecasts.However,there is no uniform standard for the acquisition of traditional CPI prediction indicators,which are usually based on experience or reference to existing literature.Experience acquisition indicators are easily affected by the limitations of personal thinking,resulting in incomplete prediction indicators.Therefore,based on the existing research experience and classification,this paper obtains CPI prediction seed words,adopts natural language processing technologyinteractive TF-IDF keyword extraction algorithm and BERT language representation model,and automatically matches the full Chinese corpus to expand CPI seed keywords,so as to build a keyword index system,to achieve a breakthrough and innovation in the construction of the CPI prediction index system.Theoretical research shows that the research method of this paper has better comprehensiveness and obj ectivity in exploring the influencing factors of macroeconomic indicators,which also provides a new idea for theoretical research and practical application of other macroeconomic indicators prediction.Second,it enriches the theoretical research of LSTM model.It is different from the existing studies that use different machine learning combinations to improve the model,and only consider the classification prediction of the LSTM hidden layer vector to allocate the attention weight.In view of the CPI regression prediction problem,this paper fully considers the important features of the LSTM model and the different effects of time series on the prediction target,and encapsulates the Multi-Representational Attention and Soft-Attention in the encoding layer and decoding layer of the LSTM network structure respectively.The input dimension features and the hidden layer vector time series are allocated with different attention weights,and the LSTM neural network regression prediction model(ATT-LSTM-ATT)based on the double-layer Attention mechanism is constructed,and the theoretical derivation of the model is given,which enriches the LSTM model’s theoretical research.Third,the application research of machine learning prediction model is extended.Few studies have explored the heterogeneity and robustness of machine learning prediction models in different periods of prediction.Based on improved model’s theoretical derivation and its application in the CPI forecast empirical,with different ratio of the training set and prediction set,under different prediction set to improve the model and the reference model of seven kinds of machine learning model prediction experiment,researchs and probes into the benchmark model in the heterogeneity of model prediction in different period and improve robustness of the model.The research of this paper shows that the improved model(ATT-LSTM-ATT)achieves the effect of timely and accurate in CPI prediction,and expands the application of machine learning model.Fourth,the method system for predicting macroeconomic indicators has been improved.The existing literature introduces artificial intelligence machine learning into the research of macroeconomic indicator prediction,and most of them focus on discussing the prediction of indicators from the perspective of selecting the optimal prediction model.On the basis of the existing research,this paper discusses the prediction suitable for CPI from multiple perspectives,from establishing the CPI forecasting seed vocabulary,constructing the forecasting variable index system,constructing the forecasting variable input set,and then improving and optimizing the forecasting model and application examples.method.The research in this paper not only achieves a breakthrough in the accuracy and timeliness of CPI prediction,but also improves the methodological system for predicting macroeconomic indicators.The main conclusions drawn from this paper are as follows:First,aiming at the problem of CPI prediction,this paper designs a text mining technology to expand the seed keywords of CPI prediction,and builds a CPI prediction index system.The technology uses the human-computer interactive TF-IDF algorithm and the BERT model to extract the CPI prediction keywords in the CPI related news from the two dimensions of breadth and depth,and realizes the feature expansion of the key words of the original CPI prediction seeds.By comparing the regression prediction effects of text mining technology before and after feature expansion and under different variable sets fused before and after,it is concluded that the CPI prediction index system constructed in this paper can more comprehensively capture the elements related to the target predictor variables,and improve the interpretation ability of CPI to achieve a more timely and accurate prediction effect.Second,this paper proposes an LSTM neural network model(ATT-LSTM-ATT)based on a two-layer attention mechanism,which improves the prediction ability of the LSTM model.The model proposed in this paper can adaptively assign weights,strengthen the influence of key features and time series information,and improve the model according to the difference in the influence of different importance features on the target predictor and the influence of the hidden layer state on the current output result at historical moments information at a particular moment.This paper sorts out and deduces the back-propagation process of the improved model,and the empirical results show that the proposed model improves the problems of LSTM processing input features and time series features,and improves the accuracy of CPI prediction,achieving timely and accurate CPI prediction effect.Third,there is heterogeneity in the prediction effects of the five machine learning models used as benchmark models in different periods.The same CPI data set is predicted on five machine learning models respectively.The experimental results show that the LSTM model has better prediction effect in the long-term and short-term than in the medium-term,and there is instability in the problem of CPI prediction.The need to improve the model,and the improved model also improves the instability of the LSTM.In the long-term and medium-term,the SVR model is the model with the best prediction performance among the five benchmark models.In the short-term,the prediction performance is not as good as that of RF and LSTM,indicating that SVR is more suitable for long-term and medium-term prediction and has advantages in small sample prediction.Among the three decision tree-based machine learning prediction models,XGBoost is more suitable for long-term and medium-term CPI forecasting,RF is more suitable for short-term CPI forecasting,and LGBM has poor performance in different periods of forecasting and is not suitable for CPI forecasting problems.Fourth,the improved model proposed in this paper has strong self-learning ability and generalization ability.In this paper,through the prediction experiments on long,medium and short CPI data sets of different periods,it is verified that the ATT-LSTMATT neural network model proposed in this paper is stable in CPI prediction.At the same time,this paper also uses the stock daily data set with high frequency to test the robustness of the improved ATT-LSTM-ATT model.The experiment found that the prediction accuracy of the improved model is higher than that of the original LSTM neural network model,which further illustrates the proposed method in this paper.The improved model improves the learning ability and generalization ability of the original model.Fifth,this paper takes CPI forecast as an example,and constructs a set of methods suitable for forecasting macroeconomic indicators.The timeliness of CPI forecasting and the accuracy of CPI forecasting are improved by constructing the CPI forecasting index system and improving the LSTM model.This research conclusion shows that the prediction model set in this study has high value,and it can predict the monthly value of CPI in the second half of the month.About 3 weeks ahead of time.