Deep Neural Network-based Consumer Price Index Forecasting in the Case of Macau

This study systematically compares the forecasting efficacy of SARIMA, LSTM and BP neural networks based on multi-source economic data from 2002-2024 in Macau SAR, with Consumer Price Index (CPI) as the core forecasting target. By integrating quarterly and monthly CPI, money supply M2, gaming revenue and other indicators, and constructing a multi-frequency data alignment framework, it is found that: the LSTM model performs optimally in complex economic fluctuations (MAPE=1.5%), and its ability to capture the nonlinear features of CPI is significantly better than that of the traditional SARIMA model (MAPE=1.8%) and BP neural network (MAPE=2.2%). ); SARIMA demonstrates the advantage of parameter interpretability in policy effect simulation, while BP network is most efficiently trained in small sample scenarios. The empirical evidence shows that money supply has a significant prior effect on CPI forecasting and there is a 3-quarter lagged effect of betting revenue shocks. A hybrid LSTM-SARIMA model is suggested to enhance forecast robustness and provide a dynamic monitoring tool for small open economies.