In-sample winner → out-of-sample reality
For each horizon we take the config a naive optimiser would have shipped (best in-sample $) and show what it did on the held-out window. Green in, red out — that’s overfitting, shown on purpose.
| Horizon | Lev | Win | In-sample $ | Out-of-sample $ | Verdict |
|---|---|---|---|---|---|
| 1d 1d dip-buy -7% · ride to target / wide stop (position) |
3× | 32% | +$8639 | $-5450 | REGIME_DEPENDENT |
| 1d 1d dip-buy -7% · ride to target / wide stop (position) |
3× | 37% | +$8920 | $-3278 | REGIME_DEPENDENT |
| 1d 1d dip-buy -7% · ride to target / wide stop (position) |
3× | 38% | +$7530 | $-1975 | REGIME_DEPENDENT |
| 1h 1h dip-buy -4% · ride ≤30d to target / wide stop |
3× | 33% | +$144 | $-404 | NET_NEGATIVE |
| 1h 1h dip-buy -4% · ride ≤30d to target / wide stop |
2× | 29% | +$-130 | $-88 | NET_NEGATIVE |
| 5m 5m dip-buy -2% · ride ≤7d to target / wide stop |
2× | 47% | +$-19 | $-33 | NET_NEGATIVE |
| 5m 5m dip-buy -2% · ride ≤7d to target / wide stop |
3× | 62% | +$5 | $-25 | NET_NEGATIVE |
| 1h 1h dip-buy -4% · ride ≤30d to target / wide stop |
2× | 38% | +$-70 | $-12 | NET_NEGATIVE |
| 5m 5m dip-buy -2% · ride ≤7d to target / wide stop |
2× | 50% | +$-13 | $0 | REGIME_DEPENDENT |
How it’s modelled: WITH hard stop (operator spec). A stop caps the normal loss, but a gap/crash beyond the ~0.9/leverage liquidation buffer at 2-3× still wipes the margin — modelled. Realistic frictions: 5 bps/side + 2 bps/side slippage + perp funding over the hold. Dip-buy entry on the bar low, no look-ahead. Configs SELECTED on in-sample only; OOS = held-out last 30%, the honest judge. 5m history ~2 weeks (30% targets rarely hit intraday). Demo/paper experiments — live stays BLOCKED behind the proof gate.