Model progression
All models in one table, in the order they were built. Every model was tested with Grouped cross-validation, with a 10 m search distance unless the row says 15 m.
How to read it: Accuracy is measured on the training data. CV accuracy is measured on unseen data and is the fairer number. Condition number shows how stable the model is: higher means less stable.
| Model | α | Params | Accuracy | CV accuracy (grouped) | Condition no. |
|---|---|---|---|---|---|
| Model 1: base | 1 | 4 | 0.518 | 0.514 ± 0.089 | 36 |
| Model 1: base, 15 m | 1 | 4 | 0.509 | 0.512 ± 0.098 | 26 |
| Model 2: share-type and sigma | 1 | 7 | 0.668 | 0.664 ± 0.046 | 117 |
| Model 2: share-type and sigma, 15 m | 1 | 7 | 0.654 | 0.647 ± 0.042 | 98 |
| 3b, Directional scope ablation (3a vs 3b) | 1 | 7 | 0.666 | 0.658 ± 0.040 | 135 |
| Model 4a: alternative-specific parameters | 1 | 13 | 0.717 | 0.707 ± 0.069 | 775 |
| Model 5: alternative-specific constants (rejected) | 1 | 15 | 0.728 | 0.717 ± 0.068 | 37,583 |
| Model 6, Interaction term screen (α = 1) | 1 | 16 | 0.724 | 0.711 ± 0.077 | 20,695 |
| Model 4a at α* | 2.43 | 13 | 0.723 | 0.717 ± 0.065 | 2,338 |
| Model 6, re-derived at calibrated alpha | 2.43 | 15 | 0.723 | 0.710 ± 0.075 | 14,851 |
| Stepwise VIF elimination | 2.43 | 14 | 0.724 | 0.709 ± 0.075 | 2,430 |
| Final model | 2.43 | 12 | 0.721 | 0.716 ± 0.076 | 1,373 |
Accuracy rises by about 20 percentage points in total. Most of it comes from two steps: adding share_type and σ (+15) and giving each type its own weights (+5). After that, extra complexity mostly changes how stable the model is, not how accurate it is.
Part of Calibration.
Linked from Calibration · Final model · Grouped cross-validation · RQ1 rebuilt with grouped CV