ChatGPT: Bradesco uses robot to evaluate BC communication – 05/17/2023 – Market

The minutes of the last Copom (Monetary Policy Committee) meeting of the BC (Central Bank) showed a still rigid posture by the monetary authority in combating inflation, with no signs of a drop in interest rates.

The assessment could have been made by market analysts, but the interpretation of the minutes, in this case, was made by the artificial intelligence program ChatGPT, based on a “provocation” made by the Department of Research and Economic Studies of Bradesco.

“The ChatGPT understood that the last minutes were a little more Hawkish” (hawkish in English, a term used to define a pro-monetary tightening posture), says the bank’s report published this Wednesday (17) signed by Myriã Bast, Marcelo Gazzano and Hoechst Cornélio da Silva.

They recall that, in the document, the BC reinforced several times that there is no mechanical relationship between the fiscal framework and monetary policy, in an apparent attempt to dissociate the presentation of new fiscal rules by the government from a consequent fall in the Selic rate.

Bradesco specialists say they have developed a methodology for labeling written communications from the Central Bank, in which they use a generative AI to provide a score for each paragraph of a minute.

The AI ​​provides values ​​close to -100 to indicate a tougher stance by the monetary authority and values ​​close to +100 to indicate a stance that is less concerned with inflation. Later, they use the sum of the paragraph scores to indicate the score for the minutes.

This process was carried out in all 55 minutes and 55 communications since July 2016 and, based on it, a historical series was constructed by the bank to indicate whether the document brought a more hawkish or dovish tone (more favorable to a reduction in interest rates). .

Bradesco analysts say that reading with less bias from AI compared to humans should not completely replace the role of market professionals, not least because it does not have the predictive power of the evolution of the scenario, which is a fundamental part of interest expectations futures.

“Even so, this type of indicator is an important ally in the evaluation of texts, either because of the speed with which the interpretation can be carried out, or because of the reduction of reading bias.”

The bank’s professionals also say that, among the next steps, they want to test whether the indicator developed based on AI is capable of anticipating market movements, specifically in the pricing of short-term interest rates immediately after the disclosure of announcements and, eventually, in the bets on movements of the next Copom decisions.

Quantitative manager uses ChatGPT to interpret texts and presentations

At Kadima Asset Management, an asset manager that works with quantitative strategies (which use data analysis to guide resource allocation), ChatGPT has been used for various tasks, such as interpretation of texts and presentations.

The managers say in a report that they asked the program to interpret an academic paper that proposes an alternative allocation approach to the traditional Markowitz approach. “In this example, ChatGPT interpreted the article correctly.”

Among the activities in which it uses AI, Kadima also uses the program to improve some texts used in commercial presentations. “The tool was able to understand the context and the idea we wanted to convey, creating a clearer text than the original we had written.”

The creation of content for posts on social networks, with “creative and coherent” suggestions from ChatGPT, and the revision of texts to identify grammatical errors, improve sentence structure and suggest more precise or appropriate terms are also mentioned among the tool’s adoptions by managers.

“In summary, our experience with ChatGPT at Kadima demonstrated that the tool has the potential to improve the efficiency and quality of various tasks in different areas of the company”, write Kadima’s managers, adding that it is essential that users understand the subject and make a critical reading of the responses obtained, as ChatGPT does not always provide accurate solutions.