AI can pick stocks better than you.

Unlock Editor Digest for free.

Rola Khalaf, editor of the FT, selects her favorite stories in this weekly newsletter.

Good morning. French markets fell slightly after President Emmanuel Macron called for early parliamentary elections, but the reaction was more of a resigned Gaelic shrug than any panic. Further proof of the unsettled theory that politics doesn't matter much to markets in the short term (except in extreme cases). Email me: robert.armstrong@ft.com.

Robots are here.

Everyone who works in the information industry — a category that includes journalists, software coders and stock pickers — must wonder if, or if, a computer will take their job.

A large language model, trained on writing for the Financial Times, can write newsletters that sound like mine. The letters may not be convincing enough today, but they likely won't be long in coming. Maybe people don't want to read newsletters written by LLMs, in which case my trip to Knicker's Yard isn't book enough. But the danger is clear.

Insecure readers may be less interested in the future of journalism than analysts and portfolio managers. Which brings me to a recent paper by three scholars at the University of Chicago School of Business, Alex Kim, Maximilian Muhn and Valeri Nikolaev (I'll call them KMN). The paper, “Financial Statement Analysis with Big Language Models,” puts ChatGPT to work on financial statements. With a few subtle hints, LLM turned these statements into earnings forecasts that were more accurate than analysts' — and the forecasts formed the basis of model portfolios that, in backtests, yielded much higher returns. created.

“We provide evidence consistent with large language models of human-like abilities in the financial domain,” the authors conclude. “Our findings indicate the potential of LLMs to democratize financial information processing.”

KMN fed ChatGPT with thousands and thousands of balance sheets and income statements, pulling dates and company names from a database spanning 1968 to 2021 and covering more than 15,000 companies. Each balance sheet and accompanying income statement contained the standard two years of data, but this was an individual input. The model was not “told” about the company's long-term history. KMN then prompts the model to perform fairly standard financial analysis (“What has changed in the accounts since last year?”, “Calculate the liquidity ratio”, “What is the gross margin?”).

Next—and this turned out to be very important—KMN persuaded the model to write an economic narrative that explained the results of the financial analysis. Finally, they asked the model to predict whether each company's earnings would go up or down in the next year. Whether the change is small, medium or large; And how sure was this prediction?

Predicting the direction of earnings, even in a binary way, is not particularly easy for man or machine. To oversimplify significantly: Human predictions (made from the same historical database) were about 57 percent accurate this time, when measured halfway over the previous year. This is better than what ChatGPT did before pointing out. However, after pointing, the model's accuracy increased to 60 percent. “This suggests that GPT comfortably dominates the performance of an average financial analyst,” KMN wrote in forecasting earnings direction.

Finally, KMN constructed long and short model portfolios based on the companies for which the model most confidently predicted significant changes in earnings. In past tests, these portfolios outperformed the large stock market by 37 basis points per month on a capitalization-weighted basis and by 84bp a month on an equal-weighted basis (the model suggests its prediction of earnings for small stocks adds more value with guises). It's very alpha.

I spoke with Alex Kim yesterday, and he was quick to emphasize the preliminary nature of the results. This is proof of concept, rather than proof that KMN has invented a better stock-picking mousetrap. Kim was equally keen to emphasize MN's finding that asking the model to write a narrative to explain the implications of the financial statements proved to be the key to unlocking greater predictive accuracy. This is the “human-like” aspect.

The study raises a number of issues, especially for someone like me who hasn't spent much time thinking about artificial intelligence. In no particular order:

1. KMN's result doesn't surprise me overall. There has been a lot of evidence over the years that first computer models or even plain old linear regressions can outperform the average analyst. The most obvious explanation is that the model or regression simply finds or follows the rules. They are therefore not subject to the biases that are only encouraged or confirmed by the information humans have access to (corporate reports, executive builders and so on).

2. What is perhaps a bit more surprising is that an out-of-the-box LLM was able to significantly outperform humans with fairly basic indicators (the model also outperformed basic statistical regression and Also performed on special “neural net” programs trained specifically to predict revenue).

3. Obviously, all the general qualifications that apply to any study of the social sciences apply here. Many studies are conducted; A few have been published. Sometimes the results don't last.

4. Some of the best stock pickers specifically avoid Wall Street's obsession with what's going to pay off in the near term. Instead, they focus on the structural advantages of business, and on the ways the world is changing that will benefit some businesses over others. Can ChatGPT make “big calls” as effectively as it can predict short-term earnings?

5. What is the job of a financial analyst? If an LLM can predict earnings better than its human competitors most of the time, what value does the analyst provide? Is he there to explain the details of the business to the portfolio manager who makes the “big calls”? Is it the information channel connecting the company and the market? Will it still have value when human buying and selling calls are a thing of the past?

6. Perhaps the ability of AI to outperform the median analyst or stock picker will change nothing. As Joshua Ganz of the University of Toronto pointed out to me, the low cost of the median stock picker was demonstrated years ago by an artificial intelligence technology called the low-fee Vanguard index fund. What matters is LLM's ability to compete with or support the smartest people in the market, many of whom are already using the power of computers to do their jobs.

I want to hear from readers on this topic.

A good read

More on Elon Musk's salary.