AI simulates the average investor not the person
A Precedent We Chose to Forget
In the 1950s, the economist Herbert Simon argued that models of rational choice described a creature that did not exist. Herbert Simon, satisficing Real people satisficed. They stopped searching when a choice was good enough, not when it was optimal. The profession nodded, published him, and went back to work. Half a century later, the same forgetting is happening again, only faster and with better marketing. We now build machines that predict what a person will buy, sell, or hold — and call the result a simulation of that person. The question Simon raised has not gone away. It has simply moved from the seminar room to the trading screen.
What the Study Actually Did
A team of researchers placed 120 volunteers in a controlled paper-trading environment. Study authors, arXiv preprint The participants used non-redeemable virtual funds under real-time market conditions. No brokerage accounts were touched. No real money moved. No transaction records were accessed. The setup was controlled and preliminary, and the paper says so plainly. Given only the information available before a prediction cutoff, the system had to guess the participant’s next-trading-day action, the security traded, and the quantity. Temporal alignment was enforced. Predictions rolled forward as the market did. The researchers compared two settings: one with point-in-time market information, one without. Market context helped the model guess the action and the ticker. It did not help with the size of the trade.
The Compression Problem
The most revealing finding is not about accuracy. It is about what the models refused to do. Across the board, the simulators overproduced hold actions. They underpredict sell decisions. They simplified multi-security transactions into something tidier than any human would execute. The researchers call this systematic behavioral compression. Study authors, arXiv preprint The phrase is clinical, but the phenomenon is familiar. A model trained on human behavior does not reproduce the behavior. It reproduces the average of the behavior, then mistakes that average for the person. The sell decision — the moment of doubt, the fear, the tax consideration, the sudden need for cash — gets smoothed away. What remains is a trader who never panics and never changes their mind. That trader does not exist. But the model is confident he does.
Why Sizing Is the Hard Part
Transaction quantity resisted prediction even when the model knew the market. This is not a technical failure waiting for a bigger model to fix. It is a structural signal. The size of a trade is where private life enters the market. A person sells more shares because a roof needs replacing, because a child starts university, because a marriage is ending, because a bonus arrived. None of that information appears in the price feed. The model sees the ticker and the timestamp. It does not see the kitchen table. So it guesses an average, and the average is wrong for everyone. The paper notes this difficulty. The number of shares is where the human being refuses to be simulated.
The Older Version of This Mistake
Economists once built models of the representative agent — a hypothetical person who stood in for everyone and therefore described no one. The representative agent was useful for closing equations. It was useless for predicting what any actual household would do on any actual Tuesday. The LLM user simulator is the representative agent with a larger vocabulary. It can talk about risk tolerance and time horizons. It can generate a plausible rationale for a hold. But the rationale is generated after the prediction, not before it. The model does not know why the person held. It knows that holding is the most common action, so it holds, then writes a sentence that sounds like a reason. The old critique applies with new force: a simulation of a person is not a person, no matter how fluent the simulation becomes.
What the Paper Does Not Claim
The authors are careful. They call the work preliminary. They note the controlled setting, the virtual funds, the absence of real stakes. They ask for larger-scale evaluation of individual, temporal, and portfolio-level behavioral fidelity. This caution is appropriate, and it also sharpens the point. If a model cannot reproduce a person’s trading behavior when the stakes are zero, the problem is not that the stakes were too low. The problem is that the model was not reproducing the behavior. It was reproducing the average of the behavior, then mistaking that average for the person. Those are different objects. One has a kitchen table. The other has a CSV file.
The Contradiction That Survives Improvement
Here is the trap. Every limitation in this paper looks like a limitation that more data and more compute will erase. Better market feeds. Longer context windows. Finer-grained behavioral traces. The models will get better at guessing the action and the ticker. The hold bias will shrink. The sell predictions will improve. And the sizing problem will remain, because sizing is not a modeling problem. It is an information problem. The information that determines how many shares a person sells is not in the market data, and it never will be. You can simulate the trader’s pattern. You cannot simulate the trader’s life. The technology improves. The gap stays. And the industry will keep calling the simulation a user, because the alternative is to admit that the user was never in the data to begin with.