Minimal Superior Penalties: Characterising and Calibrating Decision-Aware Proper Scoring Rules
Organizations: Department of Mathematics and Computer Science, Amirkabir University of Technology (Tehran Polytechnic), Iran
Abstract
Training a classifier means letting a validation metric decide which model is kept and when training stops. A natural requirement on that metric is superiority: every correctly classified forecast should score better than every incorrectly classified one, however close the underlying probabilities are. Neither the Brier score nor the logarithmic score satisfies this on its own. We show that superiority is governed by the separation between two endpoints -- the supremum of the score on correct predictions and the infimum on incorrect predictions -- and derive both in closed form for each score. Their difference is the infimal penalty, the greatest lower bound on the penalty needed to restore superiority. Prior work restores superiority with a penalty calibrated from the first endpoint alone; for the Brier base, our infimal penalty is smaller than that published penalty by a factor ranging from 4.00 at three classes to 2.11 at twenty. We further embed the published and infimal penalties in a one-parameter family and show that, under a single-valued tie convention, every member of the family is strictly proper. In a pre-specified two-dataset case study, mean selected-model macro-F1 rises with the penalty in the reported ten-class regimes, though not distinguishably from noise once seed/selection-mode nesting is accounted for; a fifth regime at forty-seven classes does not reproduce even this descriptive pattern. The intended no-superiority baseline is algebraically a rescaling of the family under test, so the experiment cannot attribute the observed pattern specifically to superiority. When superiority itself is the objective, any penalty strictly above the infimal value suffices; choosing a larger penalty is a separate calibration trade-off rather than a requirement for the guarantee.
Figures & tables
| Convention | |||
|---|---|---|---|
| A (permissive) | correct | correct | incorrect |
| B (single-valued, ) | correct | incorrect | incorrect |
| C (ties-to- ) | incorrect | incorrect | incorrect |
| Work | What it targets | Pointwise ordering guarantee | Propriety preserved | Closed-form penalty threshold |
|---|---|---|---|---|
| Kull and Flach (2015) | Decomposes a proper rule into calibration, refinement, adjustment, and grouping terms | Not addressed | N/A (decomposes an already-proper rule) | No |
| Resin (2023) | Elicitability bridge, single-class to full predictive distributions | Not addressed | Not the focus | No |
| Hu and Kash (2025) | Weighted sum of proper scoring rules for sub-properties | Not addressed | Weighting is designed to preserve it | No |
| Shahroudi et al. (2025) | Learned proxy aligning an upstream metric with downstream decision value (regression) | Not addressed | Not the focus | No (adaptive weighting) |
| Schneider et al. (2026) | Formal decision-alignment analysis; shows ECE-like metrics cannot align with pointwise-separable utilities | Diagnostic result, not a constructive guarantee | Not the focus | No |
| Lombardo et al. (2025) | Reweighting framework for cost-sensitive binary classification | Not addressed | Not propriety-preserving | No |
| Brier ( ) | Log ( ) | |
|---|---|---|
| Source def. | Eq. 42, | Eq. 43, |
| Published | ||
| Infimal | ||
| Ratio, published/ | not tabulated | |
| Implementation | Algorithm 2 | Algorithm 3 |
| Impl. vs. published | Mean-form; same | Matches published exactly |
| Published | Ratio | Implemented | Ratio | ||
|---|---|---|---|---|---|
| Grid point | Brier | Log | What it is |
|---|---|---|---|
| zero | No penalty; the base rule alone | ||
| critical ( ) | The infimal penalty (Corollary 4 , theory part 2) | ||
| supercritical | — | : strictly admissible | |
| published | The source paper’s own calibration | ||
| implemented | n/a | The exact magnitude the source paper’s own experiments ran under, on the Brier base; the log base has no comparable value to report (relation to published rules) | |
| limit | Approximates , i.e. monitoring error rate directly |
| Regime | What changes | Why |
|---|---|---|
| full | Nothing — full training data, full network capacity | The dataset’s natural difficulty |
| subset | Training data reduced | Fewer examples, more room for near-ties |
| noisy | Symmetric label noise injected | A different, independent way to create near-ties |
| hard | Reduced data and label noise together | Both pressures combined |
| Baseline | What it measures | Purpose |
|---|---|---|
| Monitor macro-F1 directly | Select the checkpoint with the best macro-F1, full stop | Upper bound on macro-F1-driven selection |
| Monitor error rate directly | Select on error rate alone | The limit of the family |
| Weighted mixture | Brier and error rate, blended at the same weights sweeps, but without superiority | Meant to isolate superiority from re-weighting; turned out not to (Section 8.6 ) |
| Outcome | Macro-F1 past | Calibration and NLL | What it would mean |
|---|---|---|---|
| flat | Statistically flat | Still degrade | Superiority is the active mechanism; the published penalty spends calibration sensitivity for no macro-F1 gain |
| rising | Keeps rising, even past published | Still degrade | Superiority is incidental to whatever benefit the penalty gives; monitor error rate or macro-F1 directly instead |
| peaked | Peaks at an interior | Still degrade | A genuine trade-off; the contribution is a -selection rule, not one recommended value |
| Dataset/regime | Macro-F1 | Informative | Near-tie rate | Penalty-fire rate |
|---|---|---|---|---|
| mnist/full | no | |||
| mnist/subset | no | |||
| mnist/noisy | no | |||
| mnist/hard | yes | |||
| cifar10/full | no | |||
| cifar10/subset | yes |
| Base | Mean diff. | Median | SD | 95% CI ( ) | 95% CI (bootstrap) | Cells |
|---|---|---|---|---|---|---|
| Brier | ||||||
| Logarithmic |
| Base | supercritical | published | Paired diff. | (raw) | (Holm) | |
|---|---|---|---|---|---|---|
| Brier | ||||||
| Logarithmic |
| value | Base | Macro-F1 | ECE | NLL | |
|---|---|---|---|---|---|
| zero | Brier | ||||
| zero | Log | ||||
| critical ( ) | Brier | ||||
| critical ( ) | Log | ||||
| supercritical | Brier | ||||
| supercritical | Log |
| value | Brier macro-F1 | Log macro-F1 | |
|---|---|---|---|
| zero | |||
| critical ( ) | |||
| supercritical | |||
| published | |||
| implemented | n/a | ||
| limit |
| value | Macro-F1 | ECE | NLL | |
|---|---|---|---|---|
| zero through subcritical | to | |||
| published | ||||
| implemented and limit | , |
| Superior ( )? | ||
|---|---|---|
| no | ||
| yes | ||
| yes | ||
| yes | ||
| yes | ||
| yes (limit) |
| Comparison | Base | Agreement |
|---|---|---|
| vs. | Brier | ( ) |
| vs. (limit) | Brier | ( ) |
| vs. (limit) | Logarithmic | ( ) |
| Baseline | Weight | Macro-F1 |
|---|---|---|
| monitor macro-F1 directly | — | |
| monitor error rate directly | — | |
| Regime | supercritical | published | |
|---|---|---|---|
| mnist/hard | |||
| cifar10/hard | |||
| cifar10/noisy | |||
| cifar10/subset | |||
| emnist_balanced/full |
Proper Scoring Rules
Proper scoring rules evaluate probabilistic forecasts and incentivize honest probability reports by rewarding distributions that match observed outcomes.
See all 34 papers