AI stock analysis: what it can and can't do

"AI picks stocks" is one of the most over-promised phrases on the internet. At the same time, large language models have become genuinely useful for investment research. Both statements are true, and the difference between them is what this guide is about.

We build an AI analysis tool, so we have a bias. We will try to make up for it by being specific about the limitations, including our own.

What AI is genuinely good at

Reading a lot, quickly. An analysis that touches five years of financial statements, the last 30 news articles, the price history and a sentiment scan used to take an analyst an afternoon. A pipeline of agents does it in a minute or two.

Structuring an argument. Language models are good at turning scattered facts into a coherent thesis with explicit assumptions. That makes it easier for you to spot the weak link.

Arguing the other side. This is the most underrated use. Humans are bad at criticising their own ideas. A model has no emotional attachment to your favourite stock and will write the bear case on request, with the same care as the bull case.

Being consistent. A checklist applied by software is applied every time. It does not skip the balance sheet because the chart looks exciting.

Where AI fails

It does not know the future. Markets move on information that does not exist yet. No model, however large, removes that uncertainty. Anyone who implies otherwise is selling something.

It can be confidently wrong. Language models produce fluent text whether or not the underlying reasoning is sound. A well-written paragraph is not evidence. Check the numbers that matter to your decision.

It inherits the data's blind spots. If the input is stale, incomplete or wrong, the output will be too. Newly listed companies, small caps and anything that depends on non-public information are hard.

It is easy to overfit. Many AI trading products show spectacular backtests that were tuned on the same history they are tested on. We explain why that is misleading in backtesting pitfalls.

Single chatbot vs. multi-agent analysis

Asking a general chatbot "should I buy NVDA?" gives you one pass of reasoning, usually hedged, with no guarantee it looked at current data. A multi-agent system splits the work:

The benefit is not that more agents are smarter. It is that disagreement is built into the process instead of being an afterthought, and every intermediate report can be read and checked.

Five questions to ask any AI stock tool

1. Do you publish every call, including the wrong ones? A tool that only shows winners is marketing, not evidence.
2. Are your numbers out-of-sample? Results on data the rules were tuned on are optimistic by construction.
3. How big is the sample and over what period? Thirty trades in a bull market prove very little.
4. Can I see the reasoning, not just the score? A number without an argument cannot be challenged.
5. Is it clear this is not advice? Serious tools say so plainly, because it is true.

How we try to handle this at Kairon

• Every directional call is logged and compared with the real price later. The results are in the public track record, wrong calls included.
• We separate in-sample from out-of-sample figures and only consider the latter meaningful.
• The portfolio-manager agent is allowed to say Neutral. Forcing a direction on every stock would look more decisive and be less useful.
• Every analysis shows the full debate, so you can disagree with specific points rather than with a black box.

The right way to use it

Treat an AI analysis like a research note from a smart, fast colleague who has no stake in your decision. Read it, check it, argue with it. The most valuable output is usually the part you did not want to hear.

This guide is educational and not investment advice.