Strategy due diligence
A ten-year test of the QQQ first-five-minute strategy. It is not overfitted — that much is now settled. Whether it can actually be traded is a different question, and the answer turns on a single number.
Selection bias, in Sharpe points. Re-choosing the configuration on every training split and scoring it only out-of-sample costs almost nothing — and 96% of 455 splits independently picked the same one.
Break-even stop slippage against the range actually measured on this strategy. The edge lives entirely inside an optimistic fill assumption, and real fills sit at or past the point where it disappears.
Growth of $1, 2016–2026, by stop-fill assumption
The backtest fills stops at the stop price. A stop is a market order once touched, so real fills land worse. Each curve applies a fixed haircut — a fraction of the stop distance — to the 40.8% of sessions that stop out. Log scale.
Annualised Sharpe against stop slippage
Break-even sits at 29%. Independent measurement of this strategy’s own stop fills put real slippage at 28–41% — the shaded band. The strategy is priced exactly at the edge of its own viability.
Why it is this sensitive: the stop is hit on 40.8% of sessions (1,024 of 2,510). This is not a rare correction applied to a handful of trades — four sessions in ten pay it.
| Stop slip | Sharpe | Sortino | Return p.a. | Vol p.a. | Max DD | CVaR 95 |
|---|---|---|---|---|---|---|
| 0% | +1.53 | +2.63 | +29.9% | 18.2% | 15.2% | 1.01% |
| 20% | +0.52 | +0.87 | +8.6% | 19.2% | 25.9% | 1.21% |
| 28% | +0.15 | +0.25 | +1.1% | 19.6% | 45.5% | 1.29% |
| 35% | -0.16 | -0.27 | -5.1% | 20.0% | 62.8% | 1.36% |
Computed with skfolio’s Portfolio
analytics on compounded returns. Highlighted rows fall inside the measured slippage band.
The concern was ordinary and serious: the stop multiplier was chosen as the best of five by full-sample Sharpe, and QQQ as the best of eight instruments the same way. Reporting the winner of twenty-five noisy draws as an estimate of future performance is biased upward even when nothing has any edge.
Combinatorial Purged Cross-Validation re-runs that choice honestly: split the decade into 15 folds, hold out every combination of 3, and reassemble the held-out blocks into 91 complete out-of-sample paths, purged and embargoed so nothing leaks across a boundary.
Sampling uncertainty
Moving-block bootstrap, 21-day blocks, 5,000 resamples. No resample is negative.
Tail concentration
Sharpe after removing the best sessions. About 4% of sessions carry the whole edge.
The tail result is an operational warning, not a statistical one. Missing sessions at random is survivable; missing the wrong ones is not. Downtime is therefore a risk exposure, not an inconvenience — and this system has lost whole sessions to a sleeping laptop and a rejected order inside the last week.
Below about 20%, the strategy is viable and worth capital. At 28% or above it is not — and no amount of parameter work fixes it, because the lever is execution: wider stops, limit exits, or a time-based exit with no stop at all.
Position: paper trading only. The strategy has cleared the statistical bar that every other strategy in this book has failed. It has not cleared the execution bar, and until real fills say otherwise the headline figure of roughly +30% a year should be read as unproven.