Your MoneyPREMIUM

SIMON BROWN: Mind your language when it goes large

Speaking of the subtle art of AI investing

A striking visualization of human and artificial intelligence collaboration, highlighting financial growth and data analysis in a modern business setting with futuristic elements. (123 R F/pitinan)

I have been using the various different LLMs from Alphabet, OpenAI and Anthropic pretty much since the launch of ChatGPT in late 2022.

For the past year I have been paying one of them the $20 monthly fee, but switching between the different offerings depending on which I felt was better.

My use case for them was fairly limited. Sure, they made great images and could help summarise large documents and answer questions. But whenever I tried to use them to help with investing they fell over.

They’d answer a question with absolute conviction but be so hopelessly wrong it was as if they weren’t even pretending to care.

I tried ChatGPT on a discounted cash flow (DCF) model late last year and it couldn’t even get the formula right, never mind coming close to the assumptions that would produce a useful answer.

Then, in February, Anthropic launched Claude Cowork, alongside some financial services plugins. One of the claims it made was the ability to do a good DCF. So I switched my subscription to Claude — you need to pay for Opus 4.6 for the DCF to work — and I tested it out.

And it worked. It nailed the formula (true, that’s a low bar). But it also nailed the assumptions required. These assumptions are very sensitive, so a high level of understanding is critical, and I was able to adjust them if I felt they were wrong.

I then started asking it to generate detailed reports. I was interested in Dell* and asked it to generate a report using my investment thesis but including other related stocks, valuations and risks. I asked it to poke holes in my thesis. It produced a quality report with a valuation 50% higher at $180, a level the stock just hit.

One important caveat: it is markedly better on US stocks than it is on those outside North America, including South Africa.

Of course, AI is never going to be 100% accurate. But then, neither am I. So I take the report as generated and upload it to one of the other LLMs, asking it to fact-check all the details. This seems to catch most errors.

Still, things do sometimes go completely haywire. One of the Cowork skills is an “initiation of coverage” and I asked it to do this on AdvTech*. It generated the report in five sections due to the complexity and the fact that I was getting rate-limited.

The reports individually were excellent. Then, when I asked it to merge them into one, Claude lost its mind. Suddenly AdvTech founded Curro in 1987 with Geoff Whyte as founding CEO.

So, yes, AI is a huge help in my investing workflow and has made me more productive and, I think, a better investor. But you need to be able to check things, because things can go wrong.

*The writer holds shares in AdvTech & Dell

Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.

Comment icon