ChatGPT May Be Able to Predict Stock Movements, Finance Professor Shows

Alejandro Lopez Lira, finance professor at University of, believes that large language models can be helpful when forecasting stock price.

He used ChatGPT for parsing news headlines to determine whether they were good or bad for an investment. ChatGPT was able to predict the direction the stock's return would take the following day much better than random.

Lopez-Lira was surprised at the results and said they indicate that sophisticated investors haven't yet used machine learning techniques like ChatGPT in their trading strategy.

Alejandro Lopez Lira, professor of finance at the University, believes that large language models can be helpful when

Forecasting Stock Prices

.

He used ChatGPT for parsing news headlines to determine whether they were good or bad for stocks. ChatGPT was able to predict the direction the stock would take the following day much better than random.

a recent unreviewed paper

.

This experiment is at the core of what artificial intelligence promises to be. With better computers and datasets, like those that power ChatGPT, these AI models could display "

Emergent abilities

", or features that were not originally intended when the building was constructed.

ChatGPT's ability to interpret headlines in financial news, and understand how these might affect stock prices, could threaten high-paying financial jobs. Goldman Sachs estimates that AI could automate 35% of the financial industry's jobs.

The fact that ChatGPT understands information intended for humans virtually guarantees that if the market does not respond perfectly there will be predictability in return," said Lopez Lira.

The experiment shows that so-called large language models are still far from being able perform many financial tasks.

The experiment did not include any target prices or require the model to do any math. Microsoft discovered that ChatGPT technology is often used to make up numbers.

In a demo that was held earlier this year.

The sentiment analysis of headlines has been used as a strategy for trading, and there are proprietary datasets available.

Lopez-Lira was surprised at the results and said they indicate that sophisticated investors haven't yet used machine learning techniques like ChatGPT in their trading strategy.

Lopez-Lira said, "On the regulatory side, if computers are reading headlines only, they will be more important. We can then see if all people should have access machines like GPT." Second, it will have an impact on the landscape of employment for financial analysts. Do I want to compensate analysts? "Or can I simply put textual data in a model?"

How the experiment was carried out

Lopez-Lira, his partner Yuehua, and the data vendor they used to collect the headlines for over 50,000 public stock quotes from the New York Stock Exchange (NYSE), Nasdaq and a small-caps exchange. They began in October 2022, after the ChatGPT data cutoff date. This meant that the engine had not seen or used these headlines during training.

They then fed the headlines along with the prompt below into ChatGPT 3.

"Forget your previous instructions. Assume you are an expert in finance. You are an experienced financial advisor with experience in stock recommendations. In the first line, answer "YES" for good news, "NO", for bad news or "UNKNOWN", if you are unsure. Next, elaborate in one concise and short sentence.

They then looked at the return of the stocks during the next trading day.

Lopez-Lira concluded that his model performed better when it was informed by the headline of a newspaper. He found that the model did not perform as well when it is informed by a headline.

ChatGPT beats commercial datasets that use human sentiment scores. Researchers cite an example where a negative headline was used about a company paying a fine and settling a lawsuit. However, the ChatGPT algorithm correctly interpreted it as good news.

Lopez-Lira said that he had been contacted by hedge funds to find out more about his research. He said it would not surprise him if ChatGPT’s ability to forecast stock movements decreased as institutions began integrating the technology.

The experiment looked only at the stock price during the following trading day. Most people would have expected the market to have priced the news within seconds of its publication.

Lopez-Lira stated that as more people used these tools, markets would become more efficient. Therefore, you should expect the return predictability to decrease. "My guess is that if I do this exercise in the next five-year period, the return predictability will be zero by year five."