Initial academic research is currently underway on the use of ChatGPT in finance. According to two of his recent studies, GPT looks like a promising technology for improving investment decisions and explaining those decisions. Perhaps the long-held dream of replacing humans in finance is coming true.
In December I wrote: Except it’s not. Financial management was one of the early goals of artificial intelligence (AI) research. It seemed like an easy and very rewarding job. But so far, AI has only been successful in niche applications in finance.
GPT stands for Generative Pre-trained Transformer, a five-year-old idea that could be a game-changer for AI applications. He has three broad approaches to extracting useful information from data. Structured data such as account numbers and price histories allow you to apply statistics and formal models. Completely unstructured data (a series of bits such as a photograph, physical measurements, text) has algorithms that can extract patterns and predict future input.
Language is in between. There are structures that mean that only certain letter combinations are comprehensible words, and there are grammatical rules for stringing words together. However, there are exceptions to the rule and there are nuances beyond the literal text. Understanding a text requires a lot of domain knowledge and context. There is an old story about his AI worker who built a program to translate between English and Russian. She attached the phrase “invisible, out of mind” to translate it into Russian, then translated Russian into English to get “invisible idiot”. There are no rules in the language that tell us that this phrase is a maxim about forgetfulness rather than a personal description, but native speakers make no mistake.
GPT models are currently the hottest approach for working with linguistic data, but quantitative trading and investing have long used cruder linguistic models. Human researchers carefully and slowly read relevant information such as company statements, news articles, surveys and research reports. Computers can read vast amounts of information in many languages and draw immediate conclusions. This is essential for high-frequency trading where milliseconds determine if a news headline is good news or bad news. Stock prices are the name of the game.
Most of the language models used in quantitative finance today treat it as structured data. Algorithms look for specific words or count words in headlines and press releases. Some algorithms look for specific patterns or structures. But none of the principal researchers try to make sense of the text, explain why they have reached their conclusions, or continue a further conversation on the subject.
There are currently two papers published titled “Can ChatGPT crack Fedspeak?” and “Can ChatGPT predict stock price movements?” We’re talking about whether to beat it and make quick decisions about short texts.
The first paper tells ChatGPT that individual sentences from the Federal Reserve’s statement are either “dovish” (suggesting that the central bank is more likely to cut rates than raise them) or “hawkish”. A high-frequency trading algorithm evaluates each sentence in a Fed release and uses its output, along with other data, to You may trade Federal Funds futures and other commodities before analysts have read the first word of the release.
In this study, ChatGPT clearly outperformed dictionary-based models searching only for specific words to match the conclusions of human analysts. When the researcher gave additional training on the Fed statement and tweaked his ChatGPT with feedback on how humans rated the statement, it was nearly as if the two human researchers agreed with each other. Concurred with human researchers with the same frequency. And that explanation for the decision was plausible.
This is not immediately useful for trading. The paper did not disclose how fast the model ran, nor whether the overall interpretation of the Fed’s releases as a whole matched well with the overall conclusions of humans (high-frequency traders are more It doesn’t matter if they match reality, because you’re trying to beat the new (not the right place in theory). However, it does suggest that the GPT model may have actually progressed toward understanding language. If that’s true—and his one study proves nothing—unleashed into a wider range of texts, his thesis such as inflation will likely continue to be a problem over the next 12 months. can be generated. High frequency trading. Also, instead of binary buy and sell signals, ChatGPT can hold conversations with human analysts to improve investment decisions. Finally, if this seems to work, the next generation of his GPT model can be trained on text and the full history of financial price movements.
The second paper is more directly related to trading. We used ChatGPT to rate news headlines with good or bad stock prices. After the headlines were released, we tested a strategy of buying stocks with good news at the opening price and selling at the closing price. Or, if the headline is bad, sell at the opening price and buy back at the closing price.
Results are inconclusive. The ChatGPT signal had a 0.01 correlation with the next day’s raw stock returns. But to assess the signal, we need to compare it with the residual return after adjusting for market returns and possibly known factors. A correlation of 0.01 may or may not be valuable in combination with other signals. The strategy tested showed positive returns without transaction costs from October 2021 to December 2022, but the authors argue that whether it beat the market strategy, the positive returns were statistically significant. did not provide data on whether His reported gross profit of 0.13% per trade suggests that transaction costs may not be overcome.
The authors also report regressions with forward-looking information, so they cannot be used to assess the effectiveness of making decisions based on information known at the time. The ChatGPT signal does not provide any additional information to the 3 decimal places indicated by the author, but seems to have a small positive value. But indecisiveness does not mean failure. This research suggests that ChatGPT outperforms popular alternative models, and research continues on GPT and other large-scale language models.
GPT is not an incomprehensible black box, but an AI tool that can work with, learn from, and teach humans. At the very least, it seems poised to replace outdated algorithms and increase the use of AI in both quantitative and qualitative investments. It’s a long way from taking over Wall Street, but there’s no reason to think it can’t be done.
Bloomberg Opinion Details:
• Even AI can’t beat the market these days: Aaron Brown
• Suspending the AI only hurts yourself: Tyler Cowen
• There is no such thing as artificial intelligence: Palmy Olson
This column does not necessarily reflect the opinions of the editorial board or Bloomberg LP and its owners.
Aaron Brown is the former Managing Director and Head of Financial Market Research at AQR Capital Management. He is the author of “The Poker Face of Wall Street”. He may have interests in the field he writes about.
More articles like this can be found at bloomberg.com/opinion.