What AI Cannot Predict About Financial Markets
The most useful thing to understand about AI and markets is where it stops working. Five categories of thing no model can see, and what that means for you.
There is nothing anti-technology about saying that AI cannot predict markets. It is the same observation that applies to human analysts, and for the same reason: prices depend on decisions that have not been made yet by people who have not decided yet.
It is worth being specific, though, because AI cannot predict markets is easy to nod along to and then forget the moment somebody shows you a chart. Here are five categories of thing that sit permanently outside any model.
1. Decisions that have not been taken
A central bank board has not yet voted. A company board has not yet decided whether to write down an asset. A government has not yet chosen whether to change a tax. No amount of historical data contains the outcome of a meeting that has not happened, and once those decisions are made they can reprice a market in seconds.
2. Genuinely new events
Models learn from the past, which makes them structurally bad at things without precedent. A pandemic, a novel regulatory regime for digital assets, a war in a place that supplies a critical commodity — these are not rare versions of familiar patterns, they are new inputs. A model asked to interpret them will confidently map them onto the closest thing it has seen before, which is exactly the wrong answer.
Example. A model trained entirely on a decade of low interest rates will have learned that falling bond prices tend to coincide with rising share prices. Feed it a year in which both fall together and it has no framework for what it is seeing — but it will still produce an output, delivered in the same confident tone as everything else.
3. Crowd behaviour and reflexivity
Markets are not a natural system being measured; they are a system that reacts to being measured. If enough participants adopt the same model, its signals change the very prices it was trained on. Strategies that worked while few people used them stop working when many do. This is not a bug that better engineering fixes — it is a property of markets.
4. Liquidity at the moment you need it
A model can tell you the last traded price. It cannot tell you the price at which someone will actually buy from you in three minutes during a sharp fall. Bid-ask spreads widen, order books thin out and slippage appears exactly when it hurts. This is why a stop order is not a guarantee, and why a backtest that assumes clean fills flatters every strategy it touches.
5. Your own circumstances
Even a hypothetically accurate forecast would not tell you what to do, because that depends on facts a market model knows nothing about: your income, your other commitments, your time horizon, your tax position, and how you would actually behave if the position were 30% down on a Tuesday morning. This is the reason general information and personal financial advice are different things, and why we are careful to say that everything on this website is the former.
How to read an AI accuracy claim
When you see a figure such as 87% accuracy, ask what was being predicted. Direction over what period? Was the test run on data the model had never seen, or on the data it was trained on? Were trading costs and slippage included? Who checked it? A claim that cannot answer those questions is not evidence, and a claim that will not answer them is advertising.
Also be sceptical of the softer versions: AI-powered precision, proven system, predictable opportunities. These borrow the credibility of measurement without offering any. Our risk and safety page lists the patterns worth recognising.
What AI is still worth having
None of this makes AI useless in markets — quite the opposite. It is excellent at reducing the cost of understanding what has already happened, which is most of what a careful person needs. Knowing that a decline was concentrated in one sector, that three announcements today were about the same theme, or what a particular fee actually means, is genuinely valuable. It is just not a forecast.
In summary
- No model can contain decisions that have not been made yet.
- Genuinely new events are where models are least reliable and most confident.
- Markets react to being modelled, so widely used signals decay.
- Liquidity and slippage break assumptions exactly when it matters.
- A forecast, even a good one, still would not tell you what is right for your circumstances.
If you would rather deal with a service that tells you where the limits are, register your interest with Murray Capholm. It is free, and it begins a conversation rather than a trading account.