基于深度学习的中国宏观经济运行评估  

Evaluation of China's Macroeconomic Operation Based on Deep Learning

作　　者：韩阳 Han Yang

机构地区：[1]哈尔滨工业大学(深圳)经济管理学院

出　　处：《复印报刊资料（统计与精算）》2023年第6期32-50,共19页STATISTICS AND ACTUARIAL SCIENCE

基　　金：国家社会科学基金重大项目(22ZDA029);2020年度上海市哲学社会科学规划青年课题(2020EJB001)的资助。

摘　　要：数字经济时代,数据要素积累和计算机算力迎来跃升,党的二十大报告也适时强调了人工智能等先进科技发展和交叉学科建设的重要性。基于此,本文从扩展后的终端产品(CFP)视角出发,将时空融合转化器算法(TFT)引入经济预测领域,对我国宏观经济系统内的产出指标进行了较为精准的批量式预测,并通过算法内部的变量选择网络和多头注意力机制从空间和时间两个维度综合评估了2019-2021年间的中国宏观经济运行。研究结论表明,在增长动能方面,我国产业演进具有以下三个特征,即工业的清洁生产转型、建筑产业链的外需驱动以及高技术产业的加速发展;在生产结构稳定性方面,我国工业生产体系在经济环境发生重大改变时较服务业等其他产业更具韧性;在政策效能方面,宏观层面上我国货币政策较财政政策整体效能更强、但传导速度稍慢,产业层面上我国工业、农业和建筑业的发展均在更大程度上受货币政策驱动,而对服务业来说,财政政策效能更为显著。本文拓宽了TFT算法的应用范围,印证了前沿深度学习技术在经济学领域存在着较大应用潜力,有助于促进人工智能和经济研究的交叉融合。The traditional economic research paradigm emphasizes transcendental theory and the interpretation of historical phenomena but often fails in forecasting the future,which makes it difficult to make effective contributions to policy practice.Therefore,it is necessary to try the reverse analysis idea from economic prediction to economic interpretation,that is,first ensure the accuracy of the prediction results and then reversely conclude the economic laws with practical explanatory power based on the prediction results.The complexity of the socioeconomic system itself is relatively high.In particular,in the digital era,the increasingly close social ties have further enhanced the nonlinearity of the whole system,bringing greater challenges to the current mainstream macroeconometric prediction.Classical machine learning algorithms,which are good at capturing nonlinear connections between variables,can often achieve considerable prediction results.However,the"black box"nature of the algorithm restricts the guiding role of policy practice because the prediction results are hard to interpret.In recent years,the cutting-edge deep learning technology represented by Temporal Fusion Transformers(TFT)can not only conduct global training for the nonlinear complex relationship between multiple variables in a system but also can interpret the prediction results from temporal-spatial dimensions through the variable selection network and multi-head attention mechanism in its internal structure,which can be applied in economic research.In the era of the digital economy,the accumulation of data and computing power have seen great progress.The report of the 2Oth National Congress of the Communist Party of China also timely emphasized the importance of advanced science,such as artificial intelligence and interdisciplinary development.Therefore,this study introduces a state-of-art TFT algorithm into economic prediction to forecast the monthly growth rates of output indicators in China's macroeconomic system from the perspective of ext

关 键 词：深度学习 宏观经济预测 增长动能 生产结构稳定性 政策效能 

分 类 号：F124[经济管理—世界经济]