Jim Simons
Slaves to the model
Discipline handed to the system
Simons in public interviews, on how Medallion is run
The best-returning fund on record forbids its brilliant people from improvising. This piece asks who that rule is actually protecting against.
What actually happened?
Medallion has compounded at better than thirty-nine per cent a year after fees for decades, and Simons described the method in one rule. If you're going to trade using models, you just slavishly use the models; you do whatever the hell it says, no matter how smart or dumb you think it is. The reason is mechanical. Allow one override and you can never again say whether a loss belonged to the model or to the person, so the system stops improving. And the urge to override peaks exactly when the model most needs leaving alone.
One override and attribution is dead
Improvement depends on clean attribution: was this loss the model or the execution? The moment a person intervenes the two are inseparable, every loss can be blamed on the other party, and nothing gets changed. The rule's real job is protecting traceability.
Operations tweaked the agreed campaign parameters three times. At the quarterly review nobody could say whether the strategy or the tweaking produced the result, so the quarter taught nothing.
Write the rule when calm, run it when frightened
The urge to intervene scales with how extreme things get, and extreme is precisely when the system most needs to be trusted. So the authority has to be fixed in advance: what lets a person touch it — a code error, corrupted data — and what never does.
Put the three conditions that permit manual intervention on the wall in writing. On the day it crashes, check the list. Fear is not on the list.
People move out of execution and into improvement
This division of labour is not an insult to judgement, it is a promotion for it. Instead of a thousand decisions taken while emotional, you make one change to the model that alters a thousand decisions. The leverage sits in the system, not in the seat.
Reviewing a hundred tickets a day by hand consumes judgement. Spending a week turning that review into a written rule copies it instead.
How do I use it today?
Where you are: the system's answer conflicts with your instinct right now, and you want to override it just this once.
Ask first: am I changing this decision or the model? Can the reason I would override be written as a rule that applies every time from now on?
Where it goes wrong: turning follow the model into never servicing the model, so a bad rule runs forever; or applying the discipline to everyone except yourself.
Lines to keep
Slavishly use the model, however smart or dumb it looks today.
Override it once and you never learn what actually failed.
We do not overrule the model. We improve the model.
Same situation, other people are asking
Can I trust what my gut is telling me this time?Everyone in the room agrees. I don't. Am I seeing something or just being difficult?The model's answer and my instinct disagree. Do I override it?It all fits together. Is that evidence or is that a story?Do I trust my gut · all 4 questions →If this one named what you are going through,
send it to someone who needs it, or keep it somewhere you will find it again.