Like a lot of outlets we’ve been writing a fair bit about generative AI lately, and specifically its implications for the finance industry. With so much going on, and so many hucksters jumping on the bandwagon, it’s difficult to separate AI hype from reality.
The good news is that the likes of ChatGPT would probably fail a CFA exam, can’t hold a (monetisable) tune, and tend to act like dumb momentum jockeys when it comes to investing. The bad news is that they could get degrees in economics and the law, and would probably be able to replace a junior sellside analyst.
Adding to the scene, Apollo’s chief economist Torsten Sløk has brought a couple of new academic papers to FTAV’s attention. The first explores how well ChatGPT parses the most obtuse, recondite, vapid but market-moving verbiage on the planet: central banker speeches.
Can ChatGPT decipher Fedspeak, the paper asks, and answers:
Yes! This paper investigates the ability of Generative Pre-training Transformer (GPT) models to decipher Fedspeak, a term used to describe the technical language used by the Federal Reserve to communicate on monetary policy decisions. We evaluate the ability of GPT models to classify the policy stance of Federal Open Market Committee (FOMC) announcements relative to a human classified benchmark. The performance of GPT models surpasses that of other popular classification methods.
Also in Sløk’s latest missive is a paper exploring whether ChatGPT and other “large language models” can predict equity prices by analysing the sentiment of news headlines.
Researchers Alejandro Lopez-Lira and Yehua Fang of the University of Florida found that earlier iterations of LLMs — such as GPT-1, GPT-2 and BERT — do a bad job, but ChatGPT apparently outperforms other commercial sentiment analysis systems already out there.
Here’s their conclusion:
First, it highlights the importance of continued exploration and development of LLMs tailored explicitly for the financial industry. As AI-driven finance evolves, more sophisticated models can be designed to improve the accuracy and efficiency of financial decision-making processes.
Second, our findings suggest that future research should focus on understanding the mechanisms through which LLMs derive their predictive power. By identifying the factors that contribute to the success of models like ChatGPT in predicting stock market returns, researchers can develop more targeted strategies for improving these models and maximizing their utility in finance.
For some sobriety amid all the hype, read Greg Zuckerman’s piece “AI Can Write a Song, but It Can’t Beat the Market” in the WSJ last week, which is excellent, as usual.
One major problem: there’s actually fairly limited data universe in financial markets. In physics you can run multiple experiments that can each produce billions of subtly different datapoints. In markets, there’s basically just one possible database: what securities have already done. And if you go back much further than a decade the data starts getting pretty coarse.
Markets are also noisier, more dynamic and more adversarial than many other realms where AI is being deployed. You’re training these models on data from what might as well be the dark ages of investing. Both fundamental and quant strategies constantly evolve.
However, it’s a bit of a straw man to say that any investment firm could, would or should turn “all their operations over to machines”.
That’s not what anyone serious in the industry is talking about. The idea of a big physical SUPERCOMPUTER! sitting in a basement somewhere that merely needs the flick of a switch to untangle the mysteries of the markets is Aronofsky stuff.
If you talk to the top people at the top quant hedge funds in the world, what they all say is that AI — whether machine learning, natural language processing etc — is just another tool. For some tasks it’s redundant, like using a sledge hammer to hit in a nail, or useless, like using a sledge hammer to paint a wall. But for certain tasks it’s either essential or will let you get a lot more work done a lot quicker.
The tools are now getting a lot better, and becoming easier for beginners to handle. That will unquestionably have a lot of implications for the investment industry.
Further reading:
— When markets become self aware (FTAV, 2015)
