Forecasting China’s Consumer Price Index with a Double-Layer Attention-Enhanced LSTM Model

Abstract

Against the backdrop of an increasingly complex and volatile domestic and international economy, timely and accurate forecasting of the consumer price index (CPI) is important for strengthening consumer confidence, implementing the strategy of expanding domestic demand, and supporting macroeconomic management. To address the multidimensional dynamics of CPI movements and the lag in official data release, this study constructs a CPI forecasting dataset by combining official statistics with text-mined online search indicators. A double-layer attention-enhanced long short-term memory model, denoted ATT-LSTM-ATT, is then developed by integrating Multi-Representational Attention and Soft Attention into the LSTM architecture. The first attention layer adaptively weights input features, while the second emphasizes critical temporal states. The proposed model is compared with ATT-LSTM, standard LSTM, support vector regression (SVR), random forest (RF), XGBoost, and LightGBM (LGBM). The empirical results show that: (1) the double-layer attention mechanism substantially improves the ability of the LSTM model to capture key features and critical time points, especially signals associated with real estate policy, shopping festivals, and holidays; (2) compared with six benchmark models, ATT LSTM-ATT delivers the best overall forecasting accuracy and exhibits stronger stability across long-, medium-, and short horizon prediction tasks; and (3) the text-mining-based forecasting framework can generate monthly CPI estimates about three weeks earlier than the official release. These findings suggest that combining big-data text signals with an attention enhanced deep learning architecture provides a useful approach for timely CPI nowcasting and macroeconomic decision support.