关键词: CPI, 居民消费, LSTM模型, Attention机制, 机器学习模型

Abstract:

Against the backdrop of an increasingly complex and volatile domestic and international economy， timely and accurate prediction of the consumer price index （CPI） plays an important role in boosting consumer confidence， promoting consumption upgrading and implementing the strategy of expanding domestic demand. However， as the complexity of economic operations increases， new industries and new modes of business continue to emerge and resident consumption patterns change， traditional statistical survey data are not conducive to accurate economic expectations and a timely grasp of changes in consumer prices due to their time lag and low frequency. Especially under the impact of big data， the traditional predictive methods and timeliness can no longer well meet the needs of economic development and policy formulation， thus exacerbating the lag in the formulation of the relevant policies， which may lead to bias in the implementation of the corresponding policies. The development of big data technology and the rise of machine learning provide ideas for solving the problems of timeliness， accuracy and complexity of CPI prediction. The purpose of this paper is to construct a predictive model of CPI using big data technology and machine learning methods， with a view to realizing a timely and accurate prediction of CPI.Aiming at the problem of CPI prediction， the natural language processing technology based on TF-IDF algorithm and BERT model are adopted to construct the CPI predictive dataset. Secondly， Multi-Representational Attention and Soft-Attention are introduced into the LSTM neural network structure respectively， which enables the model to dynamically deploy the attention to the features and temporal sequences of dataset， and ATT-LSTM-ATT model is constructed and applied to the CPI prediction problem. Thirdly， several machine learning models （including ATT-LSTM， LSTM， SVR， RF， XGBoost， and LGBM） are introduced for comparison and cross-validation analysis， respectively. Finally， the effect of introducing Attention mechanism on the predictive ability of LSTM model is explored， the accuracy and robustness of ATT-LSTM-ATT model for CPI prediction is tested， and explore the heterogeneity of multiple machine learning models for CPI prediction of different prediction sets is explored.The results of this paper show that　（1） The introduction of Multi-Representational Attention mechanism and Soft-Attention mechanism effectively improves the prediction effect of LSTM model on CPI. The two-layer Attention mechanism can strengthen the LSTM model's attention allocation to real estate policies， double eleven and holidays， etc.， and highlight the impact of important features and important points in time on the trend of CPI changes， which can effectively improve the prediction accuracy of the LSTM model for CPI. Therefore， the ATT-LSTM-ATT model has better features and time series attention allocation， time series memory and prediction functions， and has effectiveness and stability in the prediction of CPI. （2） Through the research of different machine learning models on the prediction of CPI in different periods， it is found that the ATT-LSTM-ATT model has strong stability， and different machine learning model shows heterogeneity in different prediction sets. The heterogeneity is characterized by the fact that RF， XGBoost and LGBM models are more suitable for short-term prediction of CPI， SVR is more suitable for long-term and medium-term prediction of CPI， and LSTM model is more suitable for long-term and short-term prediction， and the heterogeneity characteristics of the predictions of each model are related to its internal structure. （3） Text mining data can grasp the dynamics of resident consumption in advance， and by analyzing the number of lags in the predictor dataset， the CPI value predicted by the text mining constructed dataset combined with the ATT-LSTM-ATT model can is about 3 weeks earlier than the official release time.

Key words: CPI, resident consumption, LSTM model, attention mechanism, machine learning model