To test the forecast accuracy of local and foreign analysts, they limited their study to firms that were followed by at least one local and one foreign analyst over the sample period.
As a result, they focused on 39,381 reports of about 403 different mainland-listed companies written by 5,192 analysts at 104 brokerages, of which 37,843 reports were by 5,000 analysts at 94 local brokerages, and the remaining 1,538 reports by 192 analysts at 10 local branches of foreign brokerages in major cities such as Beijing, Shanghai and Hong Kong.
The researchers filtered out all the Chinese words that appeared in more than 95 percent of the reports, or in fewer than three articles, and were left with 218,335 words that appeared most frequently.
Researchers Identify Analysts’ Knowledge Advantage
Using a topic modelling technique, Latent Dirichlet Allocation (LDA), they processed the data and highlighted topics, or groups of similar words.
Statistical algorithms, based on the number of instances of certain words, can reveal instructive structures in data and work out, not only what the topics might be, but also how much focus is paid to different topics in each document.
“One important innovation of this paper is that the textual analysis of the analyst report contents allows us to improve our identification of the knowledge that is being used to assist analysts in making earnings forecast,” he says.
The LDA classified the words into 1,000 topics, which were manually grouped into three categories: ‘political’, ‘financial’ and ‘others’.
The researchers were left with 237 political topics — covering government regulations, initiatives, policies, and contracts, which featured words such as ‘department’, ‘food safety’, ‘national defence’, ‘military’ and ‘judicial’.
Comment