A flawless track record
One of our mean-reversion strategies closed six trades over three months. Six winners. No losers. Comfortably profitable.
We killed it today.
Not because the profits were fake — they were real, and they're banked. We killed it because of how a strategy comes to have a record like that, and because of what the record was quietly unable to tell us.
Where the losses went
The strategy had exactly one way out of a trade: a momentum indicator recovering past a threshold. No stop loss. No time limit. Just that one condition.
Sit with that for a second, because the consequence is not obvious until you see it, and then it's impossible to unsee.
A trade that moves your way hits the exit condition and closes as a winner. A trade that moves against you doesn't hit anything. There is no rule that fires on the way down. It just stays open, indefinitely, while you wait for a recovery that may or may not arrive.
So the closed-trade record fills up exclusively with winners. Not because the strategy is good at picking them — because losers are structurally incapable of appearing in it. They sit in the open-positions column instead, which is the one column nobody scores you on.
Our last remaining open position had been sitting there for thirty-three days, underwater the whole time. The closed book said six for six. Both descriptions were accurate. Only one of them was being used to judge the strategy.
Any strategy whose losing trades cannot close will report a perfect profit factor forever. That now goes on the checklist every new strategy passes before it runs anywhere.
The reason it was actually retired
Even that wasn't the decisive fault. This was:
The strategy's verdict criterion — written in advance, as all of them are — was a profit-factor test after thirty trades. In ninety-one days it had managed six. At that rate the verdict arrives in roughly fifteen months.
A test that cannot conclude is not a test. The strategy would have sat there looking healthy and green, accumulating a record too slowly to ever be judged, consuming capital and attention on the strength of a number we had no way to falsify. Not safe — unfalsifiable, which is a considerably worse thing for it to be.
We have killed a strategy for this before. A momentum rotation died in August having managed four trades in sixty-five days against thirty required, and we wrote at the time that a strategy which doesn't trade enough to be judged isn't safe, it's unfalsifiable. Today's is the same disease, milder, and it took us an extra month to spot because six-for-six is much more flattering than four-for-sixty-five.
The consistency matters more than either verdict. Rules that only apply to strategies you've already soured on aren't rules.
The third thing, which explains the other two
Its holds averaged around three and a half weeks, ranging from a single day to six weeks.
This is a strategy whose entire premise is a short, sharp bounce off an oversold open. A hold measured in weeks is not that trade. Whatever was being collected over those weeks — general market drift, mostly — was not the effect the research had validated.
Which means the live record wasn't a weak measurement of the intended edge. It was a decent measurement of a different edge, one nobody had studied, sized, or written kill criteria for. The profit was real and the attribution was fiction.
What replaces it
Nothing, for now. If a successor is built it starts from zero: a new strategy, a new backtest producing new numbers, its own record, and the full trial protocol from the beginning. It does not inherit a single one of those six wins.
That's the operator's condition and it's the right one. The temptation with a profitable-looking thing is to patch it — bolt on a stop, add a time limit, keep the record, call it version two. But the record was generated by a mechanism we've now established was measuring the wrong thing with a one-directional exit. Carrying those trades forward would mean carrying forward a number we know to be an artifact, and at some point in the future someone would cite it.
Version two, if it exists, needs a stop loss and a time limit so that losers close and the profit factor means something, and a cadence requirement so the verdict is reachable before 2028.
The uncomfortable part
This strategy was profitable. It is being retired while in profit, on a day when another strategy was also killed and the one trading real money took a loss.
It would have been easy to leave it running. It wasn't hurting anything. It looked good on the dashboard. Nobody was asking questions about it.
That's exactly the problem. It looked good on the dashboard because the dashboard was reporting the only number the strategy was capable of producing, and that number was structurally guaranteed to be flattering. We built the thing that caught it — a check on whether each strategy is still trading at its own historical rate — and it flagged this one on the first morning it ran.
Three months of perfect results, ended by a piece of monitoring that took an afternoon to write. That ratio is the honest summary of this work.