AI trading programs will trigger huge losses

Artificial intelligence is moving fast into financial markets — and it may be walking straight into a statistical trap that has fooled economists for decades.
About $1.5 trillion in hedge fund assets are already under AI control, and that number is growing rapidly as more programs come online. The promise is seductive: superhuman pattern recognition applied to vast oceans of financial data, finding opportunities invisible to human analysts. But there are fundamental flaws in how AI programs interpret data — flaws that could make the problem worse, not better.
The first is the confusion of correlation with causation.
Consider a famous example: ice cream sales and drowning deaths both spike on hot summer days. A naive model might conclude that ice cream causes drowning. The real cause is a third variable — hot weather — driving both independently. An AI model processing millions of data points might draw the same wrong conclusion, with complete statistical confidence, and act on it with real money.
This type of mistake becomes more likely, not less, as AI systems grow more powerful. When a program analyses trillions of data points across hundreds of variables — stock prices, interest rates, government deficits, the day of the week, the phase of the moon — it will inevitably find relationships that are accidental rather than meaningful. Some combination of X, Y, and Z will appear to predict that stock A rises. Applied to future markets, it fails — because the relationship was never real.
The technical term is overfitting. Models that incorporate too many variables stop capturing genuine patterns and start capturing noise. The more variables added, the more convincing the false pattern looks — and the more catastrophically it fails in live trading.
This problem has a name in statistics: high-dimensional data, where the number of variables exceeds the number of meaningful observations. The S&P 500 has 500 stocks and roughly 2,500 trading days in a decade. Add enough variables and any model can be made to fit historical data perfectly. That perfection is the warning sign, not the achievement.
The deeper problem is that economics and finance are already riddled with causal claims that are probably just correlations — and AI will inherit every one of them. The most consequential example involves central banks. When a central bank raises interest rates and the economy slows shortly after, observers conclude the bank caused the slowdown. This is the accepted narrative in every financial newspaper every day.
But serious academic research tells a different story: central banks are almost always followers, not leaders. They raise rates because the economy is already overheating — the slowdown was coming regardless. The interest rate hike gets the credit, or the blame, for something that was already in motion. An AI model trained on decades of this data would learn the false causal story, not the real one — and trade accordingly.
The irony is sharp. AI's greatest strength — finding patterns in vast, high-dimensional data at superhuman speed — is also its central weakness. More data, more variables, and more compute power mean more spurious correlations found with more statistical confidence. The financial system is about to make very large bets on the outputs of models that may be solving the wrong problem — with extraordinary precision.

Hilliard MacBeth

The opinions expressed in this report are the opinions of the author and readers should not assume they reflect the opinions or recommendations of Richardson Wealth Limited or its affiliates. Richardson Wealth Limited is a subsidiary of iA Financial Corporation Inc. and is not affiliated with James Richardson & Sons, Limited. Richardson Wealth is a trade-mark of James Richardson & Sons, Limited and Richardson Wealth Limited is a licensed user of the mark. Richardson Wealth Limited, Member Canadian Investor Protection Fund.