The weekly piece · Number one
The best result we ever produced was fake.
In June a test of ours scored better than the most famous fund in history. Three faults were holding the number up. Here is how we found them, and what the honest version came to.
One word first
A reward for risk score answers one question: how much steady return did this earn for each unit of stomach churning wobble along the way? A gain that whipsawed to get there scores low. A gain that climbed in a calm line scores high.
As a rough rule of thumb, around one is good. Around two is world class. There is a famous fund that ran near two for decades, after enormous fees. That two is the number everybody in this field is quietly chasing.
So a 2.34 should have been a celebration
On 22 June 2026 we were doing what anyone would do: testing ideas against years of history and reading the score. One candidate came back at 2.34. On paper we had beaten the legend. The book made about 44 percent a year, it was positive in every calendar year, and its worst fall from a peak was under 11 percent.
Every one of those numbers is a backtest. None of them survived the week.
The most expensive mistake in this business is not a bad idea that looks bad. It is a bad idea that looks great.
We did not celebrate, because a result that good is not evidence of a discovery. It is evidence of a bug, until you have gone looking for the bug and failed to find one. So we spent the next run trying to break our own number. That is the whole job, and it is the part nobody posts.
It broke three ways
One. It was selling short the names nobody will lend you
To sell a stock short you first borrow the shares. That step is not a formality. For the wild names, the ones that go vertical on attention, there are often no shares to borrow at any price. When there are, the borrowing cost is enormous.
About 88 percent of this strategy's edge came from its short side. That short side was leveraged funds, inverse funds, volatility products and tiny hot names. It was earning its money on trades a real person could not place. The test charged a flat borrowing cost of ten percent a year. On that basket the real cost runs from fifteen percent to well over a hundred.
Two. A safety filter was deleting the losses
Price data has bad ticks in it. A raw feed will occasionally report a daily move of a million percent, so every serious test throws away moves beyond some limit. Ours threw away anything beyond thirty percent in a day.
That sounds careful. It is not, on a short book. The days a short position loses catastrophically are exactly the days a name doubles, and those days were being deleted as bad data. The filter was quietly removing the worst outcomes of the trade the strategy depended on. A short book that never has a bad day is not a short book. It is a filter.
Three. There were about fifteen independent bets in it
The test spanned four years, which sounds like plenty. But positions were held for a long time, so the number of genuinely independent holding periods was roughly fifteen. Fifteen is not a track record. It is calling a coin lucky after fifteen flips.
There is a standard test for this. It asks whether a result stands up given how many things you tried before you found it. This one did not pass, and neither did anything near it.
What survived
One thing did. The first suspicion in a case like this is that the model peeked at the future. That means using information nobody had at the moment of the trade. We checked that first, and it was clean. The signal was built one day at a time, on data available on that day, in the order the days actually happened.
That mattered, because it meant there was something real underneath the wreckage, and it was worth finding out how much.
The honest number
We ran it again. Shorts restricted to ordinary shares a person can actually borrow. The safety filter widened so the squeezes stayed in the results where they belong. Borrowing charged at rates a real account pays, and then charged again at a punishing rate to see what broke.
The long side came back at 0.84, making roughly 15 percent a year, positive in every year, with a worst fall of about 15 percent. A version that leaned long scored 0.96. The full long and short book fell apart entirely once the squeezes were allowed back in.
So: 2.34 became 0.84. We retracted the first number the day we found the faults, and we kept the second. Still a backtest. Still hypothetical. But it is a number we can defend line by line, and the other one was not.
What to take from it
When somebody shows you a strategy result, three questions get you most of the way:
- Could these trades actually be placed? Ask specifically about the short side, and about borrowing.
- What was thrown away as bad data? A cleaning rule that removes the worst days is not cleaning.
- How many independent bets is this? Not how many years, and not how many trades. How many genuinely separate ones.
A number nobody has tried to break is not a result. It is a lottery ticket somebody is excited about.
Where these numbers come from
Our own strategy lab log for 22 June 2026, rows E5, E6 and E7: the original candidate, the adversarial audit that broke it, and the honest re-test. The engines are strategy_hunt_cycle4.py (the 2.34) and strategy_hunt_cycle5.py (the 0.84), with the findings written up in our internal strategy hunt report. The retraction is recorded in our own strategy book alongside two earlier headlines we withdrew for the same family of reasons.
All of it is simulation on historical prices, and none of it is an account statement. What an account actually did, in paper trading, with the misses in it, is on the record page.
Next week: why a strategy that wins most of the time can still lose money, and why we publish the win rate next to the coin flip. Meanwhile, the live record is here, and the weekly piece is free.