Increase size when the evidence grows, not when confidence does. Three gates: the sample is large enough to trust, the edge holds up out of sample, and the worst case at the new size still fits inside your floor. Test the new size on your own trades before a single live order carries it.
When three gates clear at once, and the order matters less than the count.
| Gate | What it requires | How to check |
|---|---|---|
| Enough evidence | A few hundred trades at the current size | Confidence multiplier near full weight at 500 |
| Edge survives fresh data | Live or holdout results track the backtest | Stability Score at 60 or above |
| Worst case fits the floor | Drawdown range at the NEW size stays inside your limit | Resample your trades at the new risk percent |
Two gates out of three is not a pass. A validated edge with a worst case that overruns the floor at the new size is a future breach, and a comfortable worst case on 60 trades is a guess wearing a seatbelt.
Because streaks are variance and size is forever. Scaling up after a hot month means buying the most exposure at the exact moment results sit furthest above their average, and an ordinary reversion then lands on the biggest size you have ever traded. The streak did not change your edge. It changed your mood.
Most size increases are mood swings with leverage. The gates exist to take the mood out of it: evidence, out of sample survival, worst case inside the floor. None of those move because last week went well.
Resample your own history at the candidate risk percent and read what changes. Your trade record in R is fixed; the new size only rescales it, so the question is arithmetic before it is courage. Our demo NQ system carries a 29R worst run across 760 trades: at 0.5% risk that run costs roughly 14.5% of the account, at 2% it costs roughly 58%, give or take compounding. Same trades, same edge, one dial.
Quantprove's simulator runs this directly: set the new risk percent, read the drawdown range, the final capital spread, and the risk of ruin at that size. If ruin moves off zero, the answer was no.
Size is the only trade you place before the market opens.
One notch at a time, re-checked after a real block of trades at the new level, fifty to a hundred before the next move. Kelly's growth optimal sizing says the ceiling is far higher, and full Kelly drawdowns are violent enough that practitioners run fractions of it for a reason. The ceiling is not the target.
One caveat from the slower side of this: the math clears a raise long before your hands do. Trading 2% does not feel like trading 1% with bigger numbers, it feels like a different job, and execution wobble at the new size shows up as a real performance change. Scale at the pace of the slower of the two, the data or you.
The rolling record keeps reporting as the new trades arrive. A size change executed cleanly leaves the R distribution untouched, so the Health Score should read the same system it read before. A size change that shakes execution shows up as drift in the recent windows, the same signature as an edge that is fading, and the record cannot tell the difference. Monitor describes what the windows show; reading the cause stays with you.
Out of sample proof. Every gate above leans on the same foundation, evidence that the edge exists outside the data that built it, and that test has its own discipline and its own ways of going quietly wrong. How to test a strategy out of sample is the piece that covers it, and it belongs before the first size increase, not after.