Recently, scientists investigated whether OpneAI's GPT-4 could do financial statement analysis on par with human analysts. Their results were unexpected: even with only raw financial data and no further context, GPT-4 was able to anticipate changes in corporate profitability more accurately than human analysts and even outperform sophisticated machine learning algorithms.

In order to avoid the model from utilizing past information, the researchers meticulously removed all firm names and dates from the standardized financial records that they sent to GPT-4 for their study. They employed unique prompts to walk GPT-4 through the analysis step-by-step in order to replicate the workflow that human analysts generally employ. This method made sure that GPT-4's analysis resembled human thought processes as much as feasible.

The researchers contrasted the forecasts made by human analysts using the IBES database with the performance of GPT-4 using data from the Compustat database, which spans the years 1968 to 2021. The outcomes were instructive. GPT-4's prediction accuracy of 60.35 percent using the step-by-step instructions was noticeably greater than human analysts' accuracy of 52.71%. Furthermore, the F1-score of GPT-4 surpassed that of the human analysts, balancing forecast accuracy and relevance.

Testing GPT-4's capabilities in the absence of any textual data, such as the Management Discussion and Analysis (MD&A) that often goes along with financial statements, was one of the study's most important components. This made it possible for the researchers to ascertain if GPT-4 could synthesize and analyze just the numerical data and still produce reliable predictions. They discovered that it could, and that the detailed instructions were very helpful in enabling GPT-4 to analyze patterns, calculate financial ratios, and synthesize data in a manner that is similar to that of a person.

Situations where GPT-4 thrived were also emphasized in the study. The AImodel performed especially well in situations where human analysts usually falter, such tiny businesses or those with erratic profits. This implies that GPT-4 has an advantage in complicated circumstances because to its general knowledge and reasoning skills. The researchers found that human analysts are still valuable in spite of this, particularly when they have more time to process data. Predictions improved much further when the knowledge from human analysts and GPT-4 was combined.

The performance of GPT-4 was equivalent to that of more sophisticated machine learning models, such as artificial neural networks (ANNs). GPT-4 even marginally beat these specialized models in certain areas. 

These results have important applications in real life. GPT-4 might help analysts estimate corporate performance more quickly and accurately. The financial analysis process might be streamlined by its capacity to handle massive volumes of data and provide insights, which could make it quicker and possibly more accurate than depending just on human judgment.