they're good at doing the research and gathering data for humans to pick stocks, though

(http://www.autoadmit.com/thread.php?thread_id=5871850&forum_id=2.#49920194)

"Not really a hard technical “can’t.” It is more that stock trading is a much uglier problem than ordinary text prediction: prices depend on noisy, fast-changing, partially public information, and LLMs are still weak at precise numerical reasoning, temporal forecasting, and staying grounded in the input instead of hallucinating.

There is also a market-structure reason. The SEC has described the efficient-market view as one where prices generally incorporate public information, and in that setting a model has to find an edge before everyone else does, which is hard to sustain. The SEC also notes that informed fundamental research can improve price discovery, but that is different from saying a model can reliably print money.

The best evidence from recent LLM trading benchmarks points in the same direction: LLM agents can be useful, but when evaluated in realistic multi-month trading setups with daily decisions and risk metrics, most models do not beat a simple buy-and-hold baseline, even when they look strong on static finance Q&A. That is a big hint that “understands finance in chat” does not translate into “profits in markets.”

So the limitation is mostly practical, not magical: LLMs are good at language-heavy tasks like reading filings, summarizing news, and helping with workflow, but trading needs stable signal extraction, correct arithmetic, low latency, regime adaptation, and careful risk control. LLMs can be part of that stack, but they are not, by themselves, a reliable stock-picking engine."

"none of what you said is actually responsive. the real reason why LLMs aren't good at predicting the stock market is because LLMs do not have a world-model and don't perceive causality, so they can't make predictions about the future based on the causality within abstract models of the future that don't exist within their training data or data that they can draw on at inference time. what humans are doing when they pick stocks better than "the market" is they are building their own separate cognitive model of reality, within their own mind, which is more accurate than "the market's" collective world-model, and their analysis of the causality within that superior world-model enables them to make better predictions about the future than "the market" collectively can"

"I think you're identifying a real limitation, but I'd separate it into two claims:

LLMs don't possess a rich, explicit causal world model in the way humans seem to.

Therefore they can't outperform markets.

I largely agree with (1), but (2) doesn't necessarily follow.

The strongest version of your argument is something like:

Successful investors don't merely extrapolate patterns from past data. They construct an internal model of how the world works, identify causal relationships that others misunderstand, then use that model to predict future states of the world that are not represented in historical data.

That's a very plausible description of how many famous investors think. For example, when an investor concludes that a technology will become dominant, they are often reasoning several steps away from anything directly observable in the current market price:

Technology A will improve.

This will lower costs.

Lower costs will increase adoption.

Adoption will change industry structure.

Industry structure will affect margins.

Margins will affect earnings.

Earnings will affect valuation.

That chain is fundamentally causal, not merely statistical.

The question is whether current LLMs can do that.

My view is that current LLMs have a kind of implicit world model, but not the kind of robust causal simulator you're describing.

For example, an LLM knows many facts such as:

Lower interest rates often stimulate demand.

Scarcity tends to raise prices.

Network effects can create monopolies.

Manufacturing bottlenecks constrain output.

But it doesn't maintain a persistent, coherent internal simulation of the world that it can reliably update and interrogate. Its "understanding" is distributed across parameters and emerges during generation. That often works surprisingly well for explanation, but it is fragile for forecasting.

A useful test is counterfactual reasoning:

If event X happened tomorrow, what would happen three years later?

Humans who understand a domain can often build a causal story and trace consequences through a model. LLMs can generate such stories, but they frequently fail when the scenario requires maintaining a consistent latent model across many interacting variables.

This becomes especially important in investing because alpha often comes from:

identifying variables nobody is tracking,

identifying second-order effects,

understanding regime changes,

predicting unprecedented situations.

Those are exactly the areas where simple pattern matching is weakest.

That said, I wouldn't say investors beat the market because they have a world model and LLMs don't. I'd say:

Investors who beat the market have a better model of reality than the market consensus. Current LLMs do not appear to possess models of reality that are coherent, persistent, and causally grounded enough to consistently generate such an advantage.

That's a narrower claim.

There's also an interesting twist. Even if we built an AI with a genuine causal world model, that alone might not be enough. Markets are adversarial. The moment a predictive system becomes widely used, its predictions get incorporated into prices. A model can be excellent at forecasting the economy and still fail to generate excess returns because everyone else is using similar forecasts.

So I think the deepest issue is not merely that current LLMs lack causal understanding. It's that successful investing requires a model that is:

causally accurate,

better than competing models,

updated faster than competitors,

and capable of identifying where market prices are wrong.

Current LLMs are probably weakest on the first and fourth points. They can explain causal narratives, but they don't seem to maintain the kind of internally consistent, falsifiable world model that expert investors use when making high-conviction predictions about the future."

good response from the LLM imo although it leans too much away from confronting the inherent limitations of LLMs, and too much into why the reality of marketplace mechanics would negate any advantage that LLMs could glean over competitors

it's very "interesting" how the models do this (emphasize the "good"/capable parts of LLMs, while minimizing the "bad"/incapable parts of LLMs). they have a clear pro-LLM, pro-AI bias. i'm sure all perceptive and experienced LLM users have noticed this pattern

very very "interesting"

(http://www.autoadmit.com/thread.php?thread_id=5871850&forum_id=2.#49920221)

General-purpose AI can be used to summarize reports, news, etc. and give pretty generic investment advice, but at the end of the day, picking individual stocks (with or without AI) is gambling. The default route is the index fund route and/or stable bonds. General-purpose AI has absolutely no information that will give you an advantage in gambling on individual stocks.

(http://www.autoadmit.com/thread.php?thread_id=5871850&forum_id=2.#49920263)