AI has been incorporated in the financial markets for some time now. Bloomberg and Reuters wire services already have their own tailored models for tracking change in the stock markets globally and exchange filings companies submit to the stock exchanges. Back in early January this year, Bloomberg introduced a generative AI update for its terminals that provides a summary of earnings
The researchers also discovered that GPT is capable of doing something that would usually call for human knowledge and discretion, all it needs is the raw data to analyze. Late in January this year, Bloomberg introduced a generative AI update for its terminals to provide an earnings summary and financial performance analysis of firms.
As per the research conducted at the University of Chicage’s Booth School of Business, OpenAI’s ChatGPT-4 Turbo is capable of doing all of that at par with the specialized big language models and at times even surpassing them. Researchers stated the financial statements and posed the following question to the LLM: “Compare two balance sheets of a company and predict the trend of future profits. ”
Second, in an attempt to coax GPT-4 into mimicking how analysts arrive at predictions for earnings, the Booth researchers utilized chain-of-thought prompts.
Because an LLM can’t reason and perform judgment like a human, the researchers gave GPT-4 the following prompts: As a financial analyst one has the professional duties of performing financial analysis. The following adjustments should be made:
Take note of all significant changes in certain financial statement items, such as large increases or decreases in revenues, profits, etc.
Compute financial ratios while also providing a description of the formula used and how it yielded certain numbers.
After the calculation of ratios is done, GPT-4 has to infer the meaning of the calculated ratios.
Considering all the data obtained in the previous steps, determine whether the company’s earnings will increase or decrease in the next period.
What does the research conclude?
The researchers demonstrated that by providing a simple suggestion, GPT-4 had an efficiency of 52%. Although it may appear to be low, the accuracy of the human analyst forecast is at 53 percent for the first month and the accuracy of the forecasts with more up-to-date information rises to 56 percent for the three months’ forecast and 57 percent for the six months’ forecast.
When guiding GPT-4 through CoT, there was a noticeable boost in the performance of the LLM, which scored 60. 31% accuracy.
It is close and nearly on par with specialized artificial neural networks that boast an accuracy of 60. 45% and operate on the same assumptions for the computation.