Abstract: Aiming at the phenomenon that the predicted value of consumer price index（CPI） lags behind the real value, a CPI prediction model based on the convolutional neural network-long-and short-term memory（CNN-LSTM） deep network is proposed. Compared with the traditional method, the prediction result has smaller root mean square error and average absolute percentage error, and the orientation accuracy and Pearson correlation coefficient of the prediction result are significantly higher than those of the traditional method. Firstly, the convolution neural network-long-term and short-term memory depth network is used to learn the spatial and temporal characteristics of futures data, and dynamically and quantitatively predict the changes of daily CPI. Then, in order to effectively improve the number of samples of in-depth network training, the monthly CPI data are enhanced. Finally, through the dynamic training model of sliding time window, the change of CPI from January 2019 to May 2020 is predicted. The model achieves high accuracy in CPI prediction and has obvious advantages in CPI prediction based on daily level data.

Key words: CPI forecasting, convolutional neural network-long-and short-term memory（CNN-LSTM） deep neural network, panel data, data augmentation, dynamic forecasting

摘要： 针对消费价格指数（CPI）的预测值滞后于真实值的现象，提出一种基于卷积神经网络-长短期记忆（CNN-LSTM）深度网络的CPI预测模型，预测结果相较于传统方法有较小的均方根误差和平均绝对百分比误差，且预测结果的定向精度和Pearson相关系数显著高于传统方法。用卷积神经网络-长短期记忆深度网络学习期货数据的空间特征和时间特征，动态定量预测每日CPI的变化情况。为有效提高深度网络训练的样本数量，对月度CPI数据进行数据增强。通过滑动时间窗口动态训练模型，预测2019年1月至2020年5月CPI变化情况。模型预测CPI取得了较高的准确率，在基于日级别数据进行CPI预测时具有明显优势。

关键词: CPI预测, CNN-LSTM深度网络, 面板数据, 数据增强, 动态预测