基于贝叶斯向量自回归模型的中国CPI预测研究

Study on CPI Forecasting in China Based on Bayesian Vector Autoregression Model

摘要: 宏观经济预测是经济决策的核心前提，CPI作为衡量通胀与物价波动的关键指标，其预测精度直接影响政策效果。针对传统模型与机器学习模型的固有缺陷，以及宏观数据非平稳、小样本等难点，本文选取失业率、货币供应量同比增速、进出口总值同比增速、固定资产投资额累计增速、财政收入累计增速为解释变量，以1990~2024年中国宏观经济数据为样本，构建BVAR模型开展CPI滚动窗口预测，并与贝叶斯线性回归、调优后随机森林、BP神经网络进行对比，同时通过单位根检验、滚动窗口框架等保障研究稳健性与公平性。实证结果表明：BVAR模型滚动预测MAE为0.5654、RMSE为0.6994，精度显著优于其余三类模型；该方法可有效融合先验信息与样本数据、规避过拟合、输出预测置信区间以量化不确定性，脉冲响应分析能够清晰地揭示宏观指标对CPI的动态传导机制与冲击效应，可解释性强且适配宏观经济数据特征。本文为小样本非平稳场景下CPI精准预测提供了可行方法与实证支撑，同时配套MATLAB代码，具有较强的实用性与可移植性。

Abstract: Macroeconomic forecasting serves as the core prerequisite for economic decision-making. As a key indicator measuring inflation and price fluctuations, the prediction accuracy of the Consumer Price Index (CPI) directly affects the effectiveness of economic policies. In response to the inherent defects of traditional models and machine learning models, as well as the challenges of non-stationarity and small sample sizes in macroeconomic data, this paper selects the unemployment rate, year-on-year growth rate of money supply, year-on-year growth rate of total import and export value, cumulative growth rate of fixed asset investment, and cumulative growth rate of fiscal revenue as explanatory variables. Based on China’s macroeconomic data from 1990 to 2024, this paper constructs a Bayesian Vector Autoregressive (BVAR) model to conduct rolling window forecasting of CPI, and compares it with Bayesian linear regression, optimized random forest and BP neural network. Meanwhile, unit root test and rolling window framework are adopted to ensure the robustness and fairness of the research. The empirical results show that the rolling forecast of the BVAR model achieves a Mean Absolute Error (MAE) of 0.5654 and a Root Mean Square Error (RMSE) of 0.6994, with its prediction accuracy significantly outperforming the other three models. This method can effectively integrate prior information and sample data, avoid overfitting, and output prediction confidence intervals to quantify uncertainty. In addition, impulse response analysis can clearly reveal the dynamic transmission mechanism and impact effects of macroeconomic indicators on CPI, featuring strong interpretability and compatibility with the characteristics of macroeconomic data. This paper provides a feasible method and empirical support for the accurate prediction of CPI in scenarios with small-sample and non-stationary data. It is also accompanied by MATLAB codes, thus possessing strong practicability and portability.