KERAS SIMPLE DENSE + GLOBAL AVERAGE
Close-H H10: MSE vs loss PnL
Input séquentiel 24×features, Dense temporel, GlobalAveragePooling1D, Dense, sortie 1. Target = return close à 10h normalisé par std train. Exécution timeout-only cap5. Seuil choisi sur Val, Test reporting-only.
Best robust2.38
Val/Test41.36/6.84
Test gain/DD39.4% / -21.7%
Avg/trade0.010%
Best variantridge_last
Classement
| Variant | Seuil | Val | Test | Robust | Gain | DD | Avg/trade | Trades | Open avg | Recent score | Recent gain | Recent DD |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Ridge baseline last-step ridge_last · avg epochs=0.0 | p80 | 41.36 | 6.84 | 2.38 | 39.42% | -21.72% | 0.010% | 3967 | 3.27 | -29.11 | -8.00% | -14.07% |
| Keras Dense+GAP, MSE loss keras_mse_gap · avg epochs=9.1 | p90 | 105.17 | 0.55 | 0.55 | 25.21% | -16.44% | 0.030% | 793 | 0.65 | -9.89 | 0.27% | -6.77% |
| Keras Dense+GAP, PnL loss keras_pnl_gap · avg epochs=9.0 | p90 | 141.17 | -3.01 | -5.00 | 28.57% | -21.05% | 0.017% | 1674 | 1.38 | -25.48 | -3.80% | -14.45% |
| Keras Dense+GAP, hybrid MSE+PnL keras_hybrid_gap · avg epochs=9.2 | p90 | 71.37 | -26.30 | -27.37 | 5.30% | -21.07% | 0.009% | 842 | 0.69 | -15.26 | 0.23% | -10.33% |
Loss PnL: minimise -mean(tanh(pred)*y_true) + petite pénalité pred². Hybrid = 35% MSE + 65% PnL. Tous les scalers et labels sont fit Train-only.
