To the clients and friends of ProfitScore:

AI And Portfolio Research?

Lessons from the trenches

It seems like almost everyone is using AI now to try to improve portfolios.  For good reason, you can do a ton of research in the shake of a squirrel’s tail.

This seems like a new phenomenon.  It is not.  Artificial Intelligence has been used in finance for a very long time. Firms like Renaissance Technologies have been using machine learning since at least the late 80’s. The book “The Man Who Solved the Market” by Gregory Zukerman is a fascinating read about Jim Simons and Renaissance Technologies.  About 10 years ago, the craze for Machine Learning was at its peak in the institutional money management world.  It was widely believed that if you were not using machine learning, you would get left behind.  Amazingly, it ended up hurting more than helping in most cases. 

My goal today is to share some of the lessons we have learned from the past.  I have shared before the story of us meeting a leading allocator for machine learning strategies.  He explained that these AI systems would perform incredibly well at first, but between the first and second years, they usually stopped working.  Sometimes crashing spectacularly.  Alpha Decay!

The good news is that computers are really good at finding relationships; the bad news is that computers are really good at finding relationships.  This means that if one computer can find a relationship, so can others, leading to Alpha Decay.

Hedge funds, quant shops, prop trading desks, and even independent RIAs are dabbling in AI-based trading. Artificial intelligence is being deployed across the investment landscape to identify patterns, back-test strategies, and locate edges that human analysts would never find on their own. The processing power is staggering. The data sets are massive. The optimism is even bigger.

So, based on what I just shared- What happens when everyone finds the same edge?

The Crowding Problem

Here is the dirty secret of AI-driven investing: the models are training on the same data.

When thousands of AI systems ingest similar price histories, economic data, and alternative datasets, they begin to identify similar patterns. They build similar signals. They take similar positions. And when the regime shifts, when the world changes in a way the models did not anticipate, they all head for the exit at the same time.

As mentioned above, we have seen this movie before. The 2007 Quant Quake was a preview. Highly sophisticated, completely uncorrelated strategies turned out to be anything but uncorrelated when liquidity evaporated. The unwind was violent, fast, and deeply painful.

AI amplifies that risk. Scale a bad assumption across a thousand models, and you have not built a smarter market. You have built a bigger bomb.

Figure 1: As AI signal adoption rises, individual strategy alpha collapses while systemic instability surges; the red zone is where both risks converge.

The New Inefficiency

Here is the irony: as AI homogenizes the inputs, it may actually create new inefficiencies depending on volatility states.  Calmer markets become more efficient, and more volatile markets become less efficient.

AI models are, at their core, pattern recognition machines. They are trained on historical data. They are extraordinarily good at identifying what has worked. They are considerably less proficient at recognizing when the world has fundamentally changed and the old patterns no longer apply.

Market state transitions, from low volatility to high volatility, from risk-on to risk-off, from expansion to contraction, are precisely the moments when AI consensus trades provide the most risk due to the dash for the door mentality. When they unravel together, the potential for sharp reversals, down and then violently back up, becomes very real.  We have seen glimpses of this in recent markets.

That is not inefficiency being arbitraged away. That may be inefficiency being manufactured at scale.

What This Means for the Disciplined Investor

The answer is not to abandon AI tools. The answer is to use them with clear eyes about what they can and cannot do.  You need to develop a sixth sense in determining if what AI is sharing is likely the same things it has discovered for many other researchers.  The best policy, in my opinion, is to generate out-of-the-box ideas, then use AI in a sandbox to validate them.  I am not sure I trust most AIs to keep my trading secrets private, or to avoid using them for training.

Strategies designed to identify and adapt to changing market states, rather than to optimize for the full market history, may carry a structural advantage. My guess is that the majority of the crowd is chasing the same patterns.

Markets may be getting smarter. But smarter is not always wiser. And in investing, that distinction has always been the one that matters.