Train: 12 flagship | Test: 6 unseen (seed 42) | 12y | region-A
| Universe | Model | Total | CAGR | MaxDD | PF | Trades | 2023-26 |
|---|---|---|---|---|---|---|---|
| Train (flagship) | pooled+adapter | -10.1% | -0.90% | 44.8% | 1.02 | 932 | -23.5% |
| Test (unseen) | pooled+adapter | +78.4% | +5.04% | 16.6% | 1.37 | 563 | +37.2% |
| Train (flagship) | pooled ref (P3a) | +280.6% | +7.00% | 41.0% | 1.25 | 1,400 | +16.6% |
| Test (unseen) | pooled ref (P3a) | +561.4% | +10.04% | 46.0% | 1.33 | 1,414 | +83.2% |
| Train (flagship) | per-symbol ref | +2161.8% | +17.10% | 36.7% | 1.46 | 1,621 | +111.9% |
Verdict
MARGINAL (train CAGR>=10%: False, holdout PF>=1.30 & 2023-26>0: True)
elapsed: 19.4 min
Ablation: pooled-only in same harness (from cache)
| Universe | Model | Total | CAGR | PF | Trades | 2023-26 |
|---|---|---|---|---|---|---|
| Train | pooled-only | +47.2% | +3.34% | 1.10 | 1,336 | -11.6% |
| Holdout | pooled-only | +80.1% | +5.12% | 1.25 | 915 | +50.8% |
Harness (12 names/12y) explains part of the gap vs P3a (3.34% vs 7.00%). Adapter effect: train WORSE (-0.90% vs +3.34%), holdout PF better (1.37 vs 1.25) but 2023-26 worse (+37.2% vs +50.8%).
Final verdict
REJECT as dilution fix: train CAGR -0.90% far below the 10% bar; adapter degrades train vs pooled-only in-harness. Holdout PF 1.37 on 6 names is noted but not significant. Next: constrain adapter (fewer trees, pooled floor 0.6) or scale pool to 51/20y before re-judging.