媒体信息、预期冲击与经济周期波动——基于中文财经类报刊数据

Media Information,News Shock and Business Cycle Fluctuations:Based on the Data of Chinese Financial Newspapers

新闻媒体作为公众信息的主要来源,对公众预期具有不可忽视的作用,进而可能对宏观经济波动产生影响。本文首先使用潜在狄利克雷分配模型,将中文财经类报刊文本数据转化为新闻主题关注度,并依据新闻主题关注度构建新闻指数作为公众预期的测度,然后结合消费、产出等关键宏观经济变量,利用结构VAR模型考察预期冲击和噪声冲击对经济周期波动的影响。研究发现,新闻指数对消费、产出等关键宏观经济变量具有明显的领先关系;从新闻指数中识别的预期冲击会对实际经济变量产生永久性的影响,而噪声冲击的影响会逐渐衰减,最终回到冲击前的水平。本文的研究验证了预期管理在宏观经济调控中的重要性,并为文本数据在宏观经济研究中的应用提供了新视角。

The origin of the business cycle is a long-standing question in macroeconomics research.Among many theories,news-driven business cycle theory addresses the importance of public expectation.According to the news-driven business cycle,changes in public expectation can cause co-movement in key macroeconomic variables under the absence of fundamental exogenous shocks.The news-driven business cycle provides theoretical support for policies such as expectation management and expectation guidance.However,how to measure public expectation is still an open question.Previous studies utilized forward-looking variables such as stock market returns or surveybased data to measure public expectation,but these approaches remain controversial.As the main information of the public,news media play a vital role in the formation of public expectation.According to the theory of narrative economics,popular narratives in the news media will cause macroeconomic fluctuations by changing public expectation,which has previously been ignored.The availability of natural language processing techniques enables economists to incorporate textual data in the news and evaluate their role for the business cycle.In this paper,first,the full-text news reports of three most prevalent business-related newspapers in China are collected from 2000 to 2021.By using the latent Dirichlet allocation(LDA)model,unstructured textual data are converted into multivariate time series data.LDA is an unsupervised machine learning method that decomposes the raw word count in the corpus into various topics.Analogous to principal component analysis in macroeconometrics,LDA summarizes the word count via topics.There are two outputs of the LDA model;topic-word distribution and documenttopic distribution.The former can be employed to interpret a topic,and the latter can be viewed as news topic attention as it measures the attention editors allocate to each topic.Second,to combine news topic attention with news-driven business cycle theory,a news index is constructed based on the predicted result of stock index return using news topic attention data As the explained part of the news topic attention data in the stock market return,the news index can be understood as the public expectation regarding future fundamentals.Moreover,the forward filtering backwards sampling method that was used for the estimation can mimic the formation of expectations and the revision of public expectation.As agents observe the current macroeconomic fundamentals and read newspapers,they will form forecasts about next period fundamentals.The results show that the news index exhibits leading properties to several key macroeconomic variables,such as output,consumption,and investment.The leading property further confirmed the utility of the news index as a measure of public expectation.Third,a structural vector autoregressive(SVAR)model is constructed to quantitatively evaluate the role of news shocks in China.By including news index and stock market return in the SVAR model,both news shocks about future fundamentals and noise shocks that are related to animal spirits can be identified.The impulse response function shows that,following a positive news shock,macroeconomic fundamentals such as total factor of productivity,output,and consumption will increase to a higher steady state.While such a positive noise shock will boost the macroeconomic fundamentals on impact,eventually,fundamentals will reverse to their initial level.These results consolidate the news-driven business cycle in China,and corroborate that the news index is a true representation of public expectation.A comparison of the forecast error variance decomposition results shows that the noise shock can explain more forecast error variance in the short-term,but the news shock becomes dominant in the long run.This work combines state-of-the-art textual analysis with classic news-driven business cycle theory.By using textual data in a newspaper,empirical evidence for news-driven business cycle theory and narrative economics is provided.The results in this paper also indicate that news media play an important role in the process of public expectation formation and policy implications are provided for expectation management and forward guidance.