Artificial intelligence

Immediately after the appearance of the chat application that works on artificial intelligence “ChatGPT”, many people began to ask him questions, each in his field, but when predicting stock prices, the answer to the application is almost general and unpredictable. Prices and markets began by explaining how you can trust it to deal with finance, but a University of Florida researcher had another point.

University of Florida finance professor Alejandro Lopez-Lera said Language Large Models, or LLMs — the programming languages ​​used to run AI bots — are useful when predicting stock prices.

He used ChatGPT to see if headlines were good or bad for stocks and found that ChatGPT’s ability to predict the direction of the next day’s returns was much better than random results. It was reported by “CNBC” network, and “Al Arabiya.net” reviewed it.

Experimentation hits at the heart of the promise of the latest AI: running bigger computers and better datasets — such as ChatGPT — these AI models may reveal “emergent capabilities,” or capabilities that weren’t originally planned for when they were built. .

And if ChatGPT shows a growing ability to understand headlines from financial news and how they might affect stock prices, it could threaten high-paying jobs in the financial sector. Goldman Sachs estimates that around 35% of finance jobs are at risk of being automated by AI.

“The fact that ChatGPT understands human information almost guarantees that there will be predictable returns if the market doesn’t respond well,” Lopez said.

But the details of the experiment also show how far the so-called “large language models” are from performing many financial tasks.

For example, the test did not include target prices for stocks or make the model perform any calculations. In fact, technology like ChatGPT mostly generates numbers, but sentiment analysis of headlines is also well understood as a trading strategy, with proprietary data sets already in place.

Lopez said they pointed out that experienced investors have yet to use ChatGPT-style machine learning in their trading strategies.

How did the test work?

In the experiment, Lopez and his partner Yuhua Tang looked at more than 50,000 headlines from data providers about public stocks on the New York Stock Exchange, Nasdaq, and a smaller stock market. And since they were launched in October 2022 – after the data cutoff date for “ChatGPT”, the engine did not see or use those headlines in training.

Then, they fed the headers to “ChatGPT 3.5” — the previous version of the chat application language — with the following instruction: Forget all your previous instructions. Let’s say you are a financial professional. You are an experienced financial professional recommending stocks. Answer “yes” if the news is good, “no” if the news is bad, or “don’t know” if you are not. . “Sure in the first line. Elaborate in a short, concise sentence on the next line.”

Then they looked at the stock returns on the next trading day.

Ultimately, Lopez found that the model performed well in almost all cases when informed by headlines. Specifically, it was found that the model had less than a 1% chance of randomly selecting the next day’s move over what was reported in a news headline.

ChatGPT also outperformed commercial datasets with human sentiment scores. One example in the paper showed a headline about a company settling lawsuits and paying fines, which had a negative connotation, but ChatGPT’s response actually treated it as good news, the researchers report.

Lopez said he was approached by hedge funds to learn more about his research. He added that he wouldn’t be surprised if ChatGPT’s ability to predict stock movements declines in the coming months as companies begin to integrate the technology. This is because this test only looks at stock prices during the next trading day, while most people expect the market to have priced the news within seconds of it going public.

In López-Lera’s view, as more people use this type of instrument, markets become more efficient and therefore less predictable.

Read more