TEST ISOLÉ WALK-FORWARD
BN / LayerNorm / GroupNorm / Cross-sectional / Portfolio loss
12 folds walk-forward 2025: Train12m, Val6m, Test1m. Train sur paniers synthétiques long/short normalisés + original; Test original principal, Full synthetic secondaire.
Configs6
Best0.89
Best gain29.5%
Best cfgexplicit_cross_sectional
Résultats
| Config | Seuil | Val orig score | Test orig score | Test gain | Test DD | Full synth score | Full synth gain | Temps |
|---|---|---|---|---|---|---|---|---|
| Cross-sectional norm explicite par timestamp explicit_cross_sectional · norm=none · loss=mse | p80 | -30.65 | 0.89 | 29.54% | -19.10% | 0.84 | 2.46% | 1.4m |
| Portfolio batch loss proxy portfolio_loss · norm=none · loss=portfolio | p80 | -14.25 | -8.93 | 2.08% | -7.34% | 2.42 | 3.62% | 1.2m |
| Base MLP no norm base_mlp · norm=none · loss=mse | p50 | -22.56 | -10.82 | 9.49% | -13.54% | 1.81 | 3.37% | 1.1m |
| LayerNorm layernorm · norm=ln · loss=mse | p50 | -25.08 | -12.79 | 1.03% | -9.22% | -1.27 | 0.43% | 1.5m |
| GroupNorm groupnorm · norm=gn · loss=mse | p80 | -16.17 | -18.43 | -4.27% | -9.44% | -2.32 | -0.57% | 1.6m |
| BatchNorm standard batchnorm · norm=bn · loss=mse | p80 | 18.68 | -37.10 | -12.65% | -16.30% | -1.35 | 0.66% | 0.9m |
Lecture
BN standard est testé comme contrôle risqué. Les variantes plus propres sont LayerNorm, GroupNorm, normalisation cross-section explicite par timestamp, et une loss portefeuille proxy. Test original reste le critère principal.
