arXiv:2601.23124v1 ·
OpenReview Xf9hJMGwDd ·
authors' code AngelReyero/loss_based_KO
| Claim | Subject | Headline |
|---|---|---|
| 1 | Valid p-values, no train–test split | type-I 0.017 at α=0.05 |
| 2 | FDR control at level q | FDR 0.159 ± 0.024 ≤ 0.2 |
| 3 | Null-feature optimisation stability | n−0.654, R²=0.996 |
| 4 | Double robustness, compound rate | n−0.891, R²=0.981 |
| 5 | Higher power than HRT + derandomisation | beats HRT 6/6 cells |
| 6 | Model-agnostic on Wisconsin Breast Cancer | type-I ≤ 0.067 across RF/NN/GB |