A Sharpe of 2.1 That Didn't Survive Contact

The 1.7-point gap between gross (4.8%) and net (3.1%) in-sample is the first clue. With 520% annual turnover, that implies a realized cost of about 33 bp per round trip (1.7% / 520%), not the 12 bp per side (24 bp round trip) assumed. So the backtest is already understating costs by roughly 40% - and that is before the live period.

Four separable causes:

1. Selection / overfitting. 144 combinations on one five-year sample means the reported Sharpe is a maximum of 144 noisy draws. Under the null of zero skill, the best of 144 roughly independent trials has an expected maximum Sharpe well above 1; you do not need the exact deflated-Sharpe number to see that 2.1 selected from 144 is not the same evidence as 2.1 tested once. The absence of a parameter plateau (a stability neighborhood) makes this worse.
2. Cost misspecification. Mid-caps in a 20-day-reversal signal trade exactly when spreads are widest and impact is highest - after a sharp move. 12 bp per side is a calm-market number. The live net of 0.2% on gross 2.9% implies ~2.7% of real cost, i.e. roughly 52 bp per round trip.
3. Universe drift. The universe was redefined twice and three of the top five in-sample contributors left it. The backtest's returns were partly earned on names that are no longer tradable - that is look-ahead survivorship in the selection of the universe, not just in the data.
4. Alpha decay / crowding. Mean reversion of this simple a form decays as it becomes known. Live gross of 2.9% vs backtest 4.8% is a 40% haircut, consistent with decay plus the universe change.

Net: the live result is not 'bad luck in seven months.' It is what this backtest should be expected to produce.

Validation requirements:

1. A genuine holdout. Reserve the most recent 24 months (or a separate asset class/universe) untouched during all parameter work. Report the test-set Sharpe and, critically, report the number of trials so a deflated Sharpe ratio or equivalent multiple-testing adjustment can be computed. A strategy that only survives with 144 trials of selection does not survive.
2. Parameter plateau, not a peak. Plot Sharpe across the 12 × 12 grid. A real effect has a broad neighborhood of positive, similar-Sharpe cells. A single hot cell surrounded by zeros is noise. Require the chosen point to sit in a plateau whose median Sharpe is at least ~70% of the peak.
3. Walk-forward with per-period reoptimization - parameters permitted to change each period, with the out-of-sample segments stitched together. This tests the process, not one lucky fit.
4. Cost realism. Replace 12 bp with a model that depends on the realized move: spread plus square-root market impact scaled by ADV, applied with a delay (signal at close, trade at next open) so no look-ahead execution. Recompute net Sharpe; if net Sharpe is under ~0.8 after realistic costs at 520% turnover, the strategy is not investable as specified. Lower turnover by widening the entry threshold or lengthening the hold - but then re-run the whole validation, because that is a new set of trials.
5. Capacity check. Estimate the AUM at which impact consumes half the gross alpha, using ADV of the smallest names in the universe. Mid-cap reversal is capacity-constrained; state the ceiling explicitly.
6. Universe as-of. Rebuild the universe as it actually existed on each date, not as it exists now, and confirm the top contributors were members at the time.

Decision rule stated in advance: allocate only if out-of-sample net Sharpe ≥ 1.0 with realistic costs, the parameter point sits in a plateau, and the capacity ceiling exceeds the proposed allocation by at least 3x. Otherwise, paper-trade and revisit.