BandStats analysisLocal, read-only model reports

OFFLINE REPORT VIEWER

Brass Band Rankings

A statistically defensible set of international brass band rankings.

RANKING RUN

dynamic-bayesian-plackett-luce as of 2026-11-13

How to read this: annual PL estimates one strength for each band-year. Higher values mean a greater predicted chance of finishing ahead of another band in the same observed field. After a band’s final observed year, its as-of strength decays toward the population mean according to the configured half-life. Uncertainty is a diagonal Laplace approximation around the MAP estimate; it also propagates through inactive years, but does not yet represent posterior correlations between bands and years.

RankBandMean strengthApprox. standard deviationConservative strengthRank change

ANNUAL STRENGTH HISTORY

Fitted strength history

The chart follows the ten highest-ranked bands at the selected year. Hover over a point to see the band, observed year, and fitted strength. Points exist only in years with an included result; lines connect consecutive observed states and do not mean a band competed every intervening year. Historical points are retrospective estimates and may be informed by later results through the ranking snapshot’s as-of date. The separate uncertainty chart shows approximate marginal distributions; it omits correlations between band-year states.

Approximate uncertainty at the selected year

The box plot shows the ten highest mean-strength bands. The centre line is the estimated mean, the box spans the middle half of the approximate marginal distribution, and whiskers span an approximate 95% interval. These are not joint uncertainty intervals.

ROLLING OUT-OF-SAMPLE BACKTEST

How do the fitted models predict later contests?

Each score comes from a contest that was excluded from that fold’s training fit. Higher log likelihood is better. The comparison is paired: both models score the same contest occurrence.

equal-strength

-10.716mean full-order log likelihood1941 held-out contests · winner log loss 1.857

static-plackett-luce

-10.244mean full-order log likelihood1941 held-out contests · winner log loss 1.702

annual-dynamic-plackett-luce

-11.784mean full-order log likelihood1941 held-out contests · winner log loss 1.847

annual-mean-reverting-plackett-luce

-10.487mean full-order log likelihood1941 held-out contests · winner log loss 1.673

Static PL advantage

+0.472mean log-likelihood difference95% contest bootstrap interval 0.42 to 0.527

Dynamic PL advantage

-1.541over static PL95% contest bootstrap interval -1.716 to -1.393

A bootstrap interval above zero supports the named first model for this retrospective set. It resamples whole contests, so a large field does not count as many independent contests.

PAIRWISE CALIBRATION

Do stated pairwise probabilities match observed results?

Calibration shown for annual-dynamic-plackett-luce.

For each strict pair of bands, the model predicts a probability that a fixed band-ID orientation finishes ahead. A calibrated model’s average predicted probability should be close to the observed win rate. Tied pairs are omitted; pairs from the same contest are shown for diagnostics, not treated as independent evidence.

Predicted rangePairsMean predictionObserved win rateVolume
0.0-0.199094.4 %22.2 %
0.1-0.2609114.7 %34.0 %
0.2-0.3498425.0 %40.3 %
0.3-0.4480534.9 %42.8 %
0.4-0.5530445.0 %47.0 %
0.5-0.6875953.1 %53.0 %
0.6-0.7473064.9 %56.3 %
0.7-0.8537875.0 %60.8 %
0.8-0.9633785.3 %67.1 %
0.9-1.01155595.8 %79.3 %

WHERE THE MODEL HELPS

Performance slices

These break the same held-out scoring down by reviewed contest geography, observed field size, and whether a field contains a band absent from its training fold. Sparse buckets need cautious interpretation; use the interval and occurrence count together.

DimensionBucketContestsStatic − equalDynamic − static95% interval (dynamic)
contest-geographyAU629+0.088-1.043-1.264 to -0.811
contest-geographyBE39+1.177+0.167-0.292 to 0.64
contest-geographyCH80+1.38-0.284-0.821 to 0.212
contest-geographyDE7+0.165+0.6070.282 to 0.896
contest-geographyGB463+0.836-2.571-3.034 to -2.139
contest-geographyGB-SCT11+0.368-0.574-1.961 to 0.787
contest-geographyGB-WLS80+0.579-0.986-1.583 to -0.438
contest-geographyNL52+0.349-1.339-2.033 to -0.729
contest-geographyNO64+1.605-0.483-1.007 to 0.048
contest-geographyNZ347+0.03-2.294-2.731 to -1.855
contest-geographySE16+0.567+0.188-0.113 to 0.519
contest-geographyZZ67+0.192-1.5-2.255 to -0.892
contest-geographyinternational86+1.319-0.611-1.077 to -0.169
field-size11+372+1.234-3.231-3.785 to -2.657
field-size2124+0.007-0.078-0.195 to 0.028
field-size3-5612+0.068-0.67-0.824 to -0.524
field-size6-10833+0.499-1.643-1.888 to -1.423
new-band-statuscontains-new-band1046+0.464-1.717-1.963 to -1.475
new-band-statusknown-bands-only895+0.482-1.334-1.547 to -1.128

DATA COVERAGE

What entered the model?

1991 eligible partial rankings from 2545 imported occurrences, containing 15025 ranked rows. An eligible observation is an ordering of its reported subset, not a complete contest field.

Contest geographyEligible occurrences
AU629
GB503
NZ357
international86
GB-WLS80
CH80
ZZ67
NO64
NL52
BE39
SE16
GB-SCT11
DE7