Artificial intelligence is becoming part of the machinery of modern markets. Banks, hedge funds and trading firms are using increasingly sophisticated models to identify opportunities, manage risk and execute trades, writes Amir Naser Hojati.

Most discussion focuses on one question: What happens when an AI model gets something wrong?

I think there is another risk.

What happens when many good models reach the same conclusion at the same time?

After roughly 15 years of trading futures, I have learned that markets often become dangerous not because one participant is wrong, but because too many become convinced of the same thing.

As traders sometimes say: when everyone moves to the same side of the boat, start watching the waterline.

Different models, same trade

Financial institutions may believe they are diversified because they use different models, vendors and strategies.

But those systems may still be learning from similar information: price history, volatility, order flow, economic data and news sentiment. They may also be optimizing toward similar goals.

Different models do not need to be identical to produce the same trade.

Several independent systems can simultaneously decide to reduce the same exposure, buy the same asset or hedge the same risk.

That creates hidden dependence.

A market can appear diversified while many supposedly independent systems are reacting to the same information in similar ways.

The real vulnerability may therefore not be one bad model, but the connection between many good ones.

Markets have seen this before

Crowded trades are nothing new.

Portfolio managers react to the same data. Traders watch the same levels. Risk desks cut exposure when volatility rises.

Technology increases the speed of those reactions. AI may also increase their consistency.

Humans disagree, hesitate or sometimes simply choose to do nothing.

Machines are less likely to disagree unless that behaviour is deliberately built into them.

If many models detect the same deterioration in risk and begin reducing exposure simultaneously, individually rational decisions can produce an irrational market outcome.

Each model may be doing exactly what it was designed to do.

The market can still become unstable.

Historical success can be dangerous

A strategy can perform well for years because market conditions remain broadly familiar.

That history creates confidence.

But markets do not owe us tomorrow because something worked yesterday.

One of the most dangerous moments is not when a model is obviously failing, but when it has accumulated enough successful history that people stop questioning its assumptions.

A calm sea can make almost any captain look skilled.

For AI-driven markets, the important question should therefore not only be:

How accurate is this model?

It should also be:

What assumptions does this model share with everyone else’s?

More intelligence is not always more resilience

The usual response to model risk is to add more: more data, more parameters, another model, another control.

But resilience does not always come from adding complexity.

Sometimes it comes from removing exposure.

A system may become safer not because it predicts markets more accurately, but because it knows when conditions are unsuitable for taking risk.

That is one of the hardest lessons to translate from discretionary trading into automation.

Finding a trading signal is relatively easy.

Teaching a machine when a valid signal should be ignored is much harder.

Experienced traders know that a setup can look perfectly good and still not deserve capital. The market may be too crowded, liquidity may be changing, or correlations may be breaking down.

Sometimes the best trade is no trade.

Diversity of behaviour matters

Using several different AI models does not necessarily provide diversification if all of them react to stress in the same way.

Real diversification may require different behaviour, not just different architecture.

When should a model refuse to trade?

When should it reduce confidence?

When should it respond more slowly?

And how should it behave when the actions of other models are themselves changing the market?

That matters because financial markets are not ordinary prediction problems.

Predicting the weather does not change the weather.

Predicting a market can change the market because participants act on the prediction.

As more intelligent systems enter that feedback loop, their collective behaviour may become as important as the quality of any individual model.

The next edge may be restraint

Financial institutions will continue building faster and more powerful models.

But I suspect the next competitive advantage will not come only from making machines better at finding trades.

It will also come from making them better at recognizing when everyone else has found the same trade.

Markets have always punished excessive certainty eventually.

AI does not remove that principle. It may simply allow certainty to spread faster.

The smartest trading systems of the next decade may therefore be distinguished not only by how quickly they act, but by something harder to program:

Kknowing when not to join the crowd.

Amir Naser Hojati is a futures trader and fintech entrepreneur with around 15 years of experience in financial markets. His work focuses on algorithmic trading, automated risk management, and AI-driven trading systems.

Share this article: