Model evaluation metrics
The numbers used to compare models. Longer explanations are in the Glossary.
| Metric | Meaning |
|---|---|
| LL(0) | Log-likelihood (how well the probabilities fit the correct answers) when every candidate gets the same probability |
| LL(C) | Log-likelihood of a simple baseline: one constant per type, which only knows how common each type is |
| LL(b) | Log-likelihood of the fitted model. Closer to 0 is better |
| ρ² vs LL(C) | How much better the model is than the simple baseline. Negative means worse than the baseline |
| Accuracy | The share of stays where the most likely type is the correct one |
| Condition number | How stable the fit is. Large values mean the single weights cannot be trusted |
| VIF | For each input, how much it overlaps with the other inputs. Above 10 is the usual warning line |
ρ² vs LL(C) cannot compare different search distances (R_buffer), because the baseline LL(C) itself changes when the list of candidates changes (R_buffer: 10 m vs 15 m). All cross-validation numbers use Grouped cross-validation.
Part of Calibration.
Linked from Calibration · R_buffer: 10 m vs 15 m