We’re an investment preparation agency that’s been running for quite a while in Zurich, Switzerland. A team member once proposed to combine LLMs – financial statement analysis with large language models like GPT. Honestly, the opportunities if this works out are huge.
For one, public companies release their financial statements all the time. A proper financial analysis in a few minutes would allow a slightly more accurate market prediction. So we went to the drawing board, and looked at our options.
We tried ChatGPT, Claude, and Gemini. The three most dominant LLMs.
ChatGPT
Mostly, the question will first be what type of document are you uploading to analyse. If it’s a PDF, are the statements pictures or values? This will make a huge difference in the analysis. As far as we could tell, most LLMs fail in this.
If you’re uploading Excels, like we did, the results are better. But surprising the most perfect way to deal with this was to create a new project (on Claude or ChatGPT), then upload in the project knowledge screenshots of the financial statements (yes, you read that right.)
ChatGPT performed averagely in this area. We loved how it analyzed huge PDFs quite quickly. But when it came to excels, it (and all the others) was slightly imperfect.
Claude
This is what we ended up using. We created a new project and used Claude (with tons of screenshots) to remember everything in case we forgot. Overall, it saved time, but not as much as we thought it would be.
We imagined a completely, AI controlled environment where the statements would be analyzed, released, etc. That wasn’t the case. But some financial statement analysis were easier with this large language model.
Financial modeling is sort of an encyclopedia, some things are much easier than others.
Gemini
Gemini was at third place. It was good, but not as good as the other two when it came to their specific perks mentioned. If you have it in your subscription bundle, then you ought to try it before giving the others a shot.
However, Claude is our recommendation when it comes to financial statement analysis with large language models.
The Future of Financial Statement Analysis with Large Language Models
After months of testing, here’s the reality: financial statement analysis with large language models isn’t the magic bullet we hoped for, but it’s not useless either.
The biggest takeaway? We’re still in the early stages. None of these LLMs completely replaced our traditional analysis methods, but they definitely sped up certain parts of the process. Claude came out on top for our specific needs, mainly because of how well it handled our screenshot-heavy project setup.
What we learned is that the format of your financial data matters more than we initially thought. Excel files work better than PDFs, but surprisingly, screenshots in a dedicated project worked best of all. Go figure.
The time savings were real, just not revolutionary. Instead of completely automating our financial statement analysis, these large language models became powerful assistants that helped us catch things we might have missed and process information faster.
For investment firms like ours, the question isn’t whether to use LLMs for financial analysis—it’s how to use them most effectively. Right now, that means Claude with a well-organized project setup, realistic expectations, and keeping human oversight in the loop.
Will this change in the next year or two? Probably. But for now, we’re sticking with our hybrid approach: let the AI handle the heavy lifting, but keep the final analysis decisions where they belong—with experienced analysts who understand the nuances that algorithms still miss.
If you’re considering financial statement analysis with large language models for your firm, start small, test thoroughly, and don’t expect miracles. The technology is promising, just not quite there yet.