Core Earnings LLM model

Has anybody read this paper? The talk about a sequential LLM prompting approach to Scrub earnings. Each company takes about a minute. There version of core earnings is better than compustat standardized operating earnings measure in many cases.

I was wondering if this approach is pursued at P123, for fundamental investors like me - could be very valuable.

5 Likes

Yeah. I think this is one of the most important ways to implement LLMs

Interesting stuff! Factset et al. must be working hard on this.

The number of ways to view earnings is kind of astounding. You can use operating income, income before taxes, income after taxes, net income (either including or excluding extraordinary items), EBIT, EBITDA, or EBTDA. You can add/subtract special items from income. You can use basic per share or fully diluted per share. You can get analyst estimates: median, high, or low. Compustat has a cleaned-up version called operating EPS, which supposedly excludes non-recurring items (this is the OPEPS mentioned in the paper; the OIADP is our OpInc). You can, of course, focus on quarterly, annual, TTM, or 3YAvg. And then there are the newly introduced Actuals.

I like the paper and think that there really is some good potential in LLMs reading through all the footnotes and adjusting financial statements accordingly. But FactSet and Compustat already do that, with real people in charge of the standardization. As test_user points out, they're very likely experimenting with LLMs to reduce staff costs.

When it comes to earnings, I think one of the important things is to get all the nonrecurring stuff out of the way. This is what analysts and actuals and Compustat's OPEPS and Special Items are all supposed to do. Throwing LLMs into the mix can't hurt, but it might be overkill.

It's interesting that when it comes to free cash flow, we have so few inputs (by comparison). Not that many companies have FCF estimates, none of the data providers break down CapEx into maintenance and growth, and you really have to read the financial statements carefully to get what the actual capital expenditures might be. I seem to remember a rent-a-car company spending money on buying new cars and counting it as CapEx, but when it sold its old cars, that didn't diminish CapEx at all. One should probably increase CapEx for companies with steady annual acquisitions costs, but not for companies that make one-time acquisitions. All of this has to be adjusted by hand (or with the help of a well-trained LLM), and as far as I can tell, nobody is doing this work.

3 Likes

Thanks everyone.

Right but a delay for micro-caps? Pretty significant at times if I recall correctly?

Since the authors argue for listening to ChatGPT I wondered what it would think of the overall approach RNTL outlines (maybe some additional specifics below): "your approach is compelling"

So wouldn't the real benefit be to upload a PDF file of an earnings report before the market opens—especially for a micro-cap with little analyst coverage? It may or may not be better than what an SP500 analysts would write but at least I have it before the market open. Making the comparing the 2 a bit academic? At least I have it whatever the academics finally decide is better.

The paper uses ChatGPT so maybe its opinion counts here too. I asked it about the paper and how I might best use it:

Q: "So you could do that for me now? Duplicate what the paper does before the market open?"

A: "Yes,….."

Q: "If I added Zacks most recent sentiment score do you think you could help Identify a surprise or change in sentiment before the market opens?"

A: " Yes, combining Zacks' sentiment score with a detailed earnings report analysis could provide a strong basis for identifying potential earnings surprises before the market opens…"

I would do something like that and those interested could probably paper trade it now. Maybe I would also give it the consensus earnings estimates (from P123) and ask it to explicitly predict the direction the stock will take after the open. Buy at the open or consider letting it finishing gapping down before I buy it if I like the stock and ChatGPT predicts an initial negative response.

I'm not talking about only a single metric, I was talking earlier about how analyzing financial reports rather than news sentiment is a better application scenario for LLM.

And even if we're only talking about core earnings metrics, I don't think that many microcaps or European companies that haven't had analysts/data for a long time are what you're referring to when you say that "analysts, actuals, and data providers are already trying to do this".

In many cases there is even a lack of "real people in charge". And this is precisely the case with stocks that are more likely to gain an advantage (lack of analysts and delayed or even missing reports from data providers).

Of course, so it's nice to fill in the missing data other than core earnings from the data providers.

And "real people in charge" doesn't mean it's better, just as active funds with "real people in charge" are not as good as index funds with "no real people in charge" on average.

Machine learning is stupid, but humans are even stupider.

1 Like

I didn't mean to reply to you, ZGWZ, earlier; I meant to comment on the paper, which really focused on earnings. Sorry about that. I agree with you about the LLMs being better for analyzing financial reports than news sentiment, and your point about companies lacking analysts is a good one.