Gulfstream Capital
← Gulfstream Intelligence

Research perspective / AI and investment judgment

AI can learn the routine. Where does the edge come from?

When an algorithm can anticipate much of a fund manager’s behavior, what deserves closer attention? A new study and a question from our own research point toward the work that comes after pattern recognition.

AI can make investment research faster. It also raises an awkward question for active management: how much of the work is genuinely distinctive, and how much follows a pattern someone else can reproduce?

That question surfaced in Alex Edmans’s October 6 conversation with Peter Thal Larsen on Reuters Breakingviews’ The Big View. Edmans discussed research linking harder-to-predict investment decisions with stronger performance. It is a useful starting point for thinking about the role of technology, human judgment and genuinely differentiated research. [1]

An organized information grid and an evidence-review lens connect technology with investment judgment.
Technology can organize and compare. Judgment has to test the evidence and the decision.

What the study actually found

In Mimicking Finance, Lauren Cohen, Yiwen Lu and Quoc H. Nguyen report 71% prediction precision for mutual funds’ quarterly buy, sell or hold decisions, versus 52% for a benchmark based on the manager’s most common past action. The working paper also finds that holdings with incorrectly predicted decisions outperformed holdings with correctly predicted decisions. [2]

The distinction matters. Predicting a manager’s behavior is different from predicting a stock’s return. And “harder to predict” means this particular model missed the decision. It does not establish a permanent boundary on what AI can learn, or show that surprise itself causes better performance.

Our interpretation is that access to a capable model is becoming part of the research toolkit. The harder question is what an investment process adds: better evidence, a more complete understanding of the business, a differentiated explanation of what could change, and a disciplined assessment of the price paid.

A manager decision and an uncertain investment outcome are shown as separate research questions.
Predicting a decision and predicting a return are different questions. No returns or forecasts are plotted.

What a few outliers can change

Our head trader’s preliminary, AI-assisted work on insider-buy signals raises a related question. The observation from that work is that a relatively small number of strong outcomes can have a large influence on the average, while the typical result looks much more modest. The work remains exploratory and has not been independently validated.

That observation concerns the shape of outcomes. The fund-manager study concerns the predictability of decisions. They are different questions, but together they encourage a useful discipline: look beyond how often a process appears to be right, and examine what actually drives its results.

A high success rate can coexist with poor economics if occasional losses are large. A promising average can also be fragile if it rests on only a few observations. The relationship among the typical outcome, the strongest and weakest outcomes, position sizes and costs deserves close attention.

A large gap between average and median is a reason to inspect the underlying distribution. By itself, it cannot establish that the strongest outcomes were identifiable in advance, that they will recur, or that a particular selection process has a durable advantage.

Typical cases and exceptional cases are examined alongside position size, costs and uncertainty.
Examine the typical case and the exceptions, then account for sizing and costs. Conceptual, not performance data.

The discipline after the signal

For us, the useful response is to sharpen the questions.

What information was actually available when a decision could have been made? Does the apparent opportunity survive a different period and realistic trading costs? Are several attractive observations really the same underlying event? How dependent is the result on one exceptional winner?

There is a second question too: could a seemingly sensible filter remove unusual situations before they receive proper attention? Outliers deserve investigation. That includes checking the data and understanding the circumstances, alongside testing how much the conclusion depends on them.

The practical role for AI is broad: organize public information, compare cases, challenge assumptions and make gaps easier to see. Human judgment must then be explicit enough to test. A compelling story needs evidence; a confident model output needs scrutiny. Neither earns an exemption from risk, valuation or portfolio fit.

Available information, a research decision and later evidence appear in a clearly ordered sequence.
Keep information available at the decision separate from evidence learned afterward. Schematic sequence only.

Where we put the emphasis

The aim is a research process that can recognize its own limits, investigate what is unusual and remain disciplined when outcomes are uneven. That leaves room for exceptional results without assuming we can identify them with certainty.

As routine analysis becomes easier to reproduce, we believe the burden of proof on investment judgment should rise. Better questions, stronger evidence and careful risk decisions are how that judgment has to earn its place.

Evidence, costs, valuation and risk surround an investment thesis.
An investment thesis needs scrutiny from more than one direction. Conceptual research lenses.

Source notes

  1. Reuters Breakingviews, The Big View, How financial markets became even more manic, October 6, 2026. Guest: Alex Edmans; host: Peter Thal Larsen.
  2. Lauren Cohen, Yiwen Lu and Quoc H. Nguyen, Mimicking Finance, NBER Working Paper 34849, February 2026. Accessible author manuscript, dated January 27, 2026.