A recent Harvard-led study showed that an algorithm can learn to predict what a fund manager is attempting when buying, selling, or holding a stock.

In other words, most active management follows a pattern a machine can copy and potentially automate.

Despite this, there are still aspects of investing that an AI cannot replicate nor take away from humans.

Investor Essentials Daily:
The Monday Macro Report
Powered by Valens Research

There is still hope that human judgment can beat AI investing.

At least for now.

A recent Harvard-led study set out to test if an algorithm can learn to predict exactly what a fund manager is trying to do when they buy, sell, or hold a stock.

On the surface, the answer looked like yes. Using data from 1990 through 2023, the model was able to predict about 71% of mutual-fund trading decisions.

That suggests most active management follows a pattern a machine can copy and potentially automate.

Here’s the good news: The 29% of trades the model could not anticipate turned out to be the trades more closely associated with outperformance.

In other words, AI proved it could mimic the routine part of the job. The harder-to-see, non-routine decisions were the ones that helped managers beat the market.

This shows the best investing ideas continue to come from what machines still can’t model.

A lot of this comes back to what Howard Marks had to say about AI and investing.

Good investing is partly about pattern recognition. Great investing is about knowing when the pattern no longer applies.

The Harvard model seems to have learned the first part very well. According to the study, it was especially good at understanding how managers react to repeated stimuli like changes to economic conditions including interest rates and inflation, plus the behavior of their peers.

It’s not hard to see why the model gets this. A huge portion of investment management is systematic.

The study highlights that many managers simply follow the crowd. The model could predict whether a fund would buy or sell a particular stock based on what other peers in its fund category did.

When everyone is making the same moves, though, it’s hard to beat the market. That’s why the “unpredictable” part of investing is what sets good managers apart.

A great example is the role of conviction. Most fund managers put their biggest weights in the ideas they believe in most strongly. The top five companies in a given portfolio were the hardest for the model to predict.

In many cases, that is because the edge is slightly speculative.

It depends on a judgment about what the market is missing before the evidence is obvious.

Managers who leaned into companies with high sales growth, high R&D relative to earnings, and positive earnings surprises outperformed the market by roughly 4.25%.

Those are the kinds of decisions that can confuse a model trained in the past. If a company is pouring cash into R&D, it’s hoping for a breakthrough that leads to earnings in the future. An algorithm can’t predict whether those investments will bear fruit or not.

That’s where human judgment earns its keep. Like Marks said, there are non-quantitative factors—basically trying to predict a company’s future—that AI cannot help with.

The study makes one thing clear.

Investors can use AI for the basics. It can speed up the “rote” parts of the job. It can pull together data that shows how interest-rate changes affect portfolios and it can summarize the habits of the crowd.

It is a great tool for the parts of investing that are repetitive and rely on data everyone has access to.

That said, don’t expect it to find the next great surprise before the market does. It doesn’t know how to account for R&D spending, and it can’t predict when a company will surprise Wall Street. For that, use the factors it’s unable to grasp.

Best regards,

Joel Litman & Rob Spivey
Chief Investment Officer &
Director of Research
at Valens Research