基于回归树集成学习方法的工业增长预测和分析

Forecasting and Analysis of Industrial Growth Based on Regression Tree Integrated Learning Methods

本文从众多变量中筛选出59个相关经济指标,分别考查疫情前后传统时间序列模型和几种回归树集成学习模型对中国工业增加值增速的预测效果,并结合Shapley additive explanations(SHAP)方法对相关预测变量的作用进行解释分析.研究发现,随着预测步长的增加和新冠疫情的暴发,传统时间序列模型的预测性能明显减弱,而集成学习模型的预测表现则相对较好,其中梯度提升树模型在较长预测步长中更加稳健和准确.基于SHAP方法的分析发现,作为预测变量的经济指标在不同时期的重要性有所不同,除生产、投资等指标外,金融类变量在高风险时期也具有一定的预测作用,需结合具体时间和预期目标来选择合适的经济指标进行工业增长预测.基于预测的视角可在一定程度上说明新冠疫情冲击可能不会改变工业增长未来走势的基本面.

The paper selects 59 related economic indicators from numerous variables,and compares the forecast effects of traditional time series models with several regression tree integrated learning models on the growth rate of China’s industrial added value under different scenarios,and the Shapley additive explanations(SHAP)method is combined for interpretation.Our results show that,with the increase of the forecast step and the outbreak of the COVID-19 epidemic,the forecast performance of the traditional time series model is significantly weakened,while the integrated learning model is relatively better,among which the gradient boosting decision tree model is more robust and accurate in the longer forecast step.Based on the analysis of SHAP method,we find that the importance of economic indicators as predictors in different periods is different.In addition to indicators such as production and investment,financial variables also play a certain early warning role in high-risk periods,and appropriate economic indicators should be selected according to specific time and expected goals for industrial growth forecasting and analysis.From the perspective of forecasting,the impact of the COVID-19 pandemic may not change the fundamentals of the future trend of industrial growth to some extent.