← MCM 2026: Modelling Audience Votes
Scoring-rule explainer
Made-up week · method illustration
On the show, judges' scores are public and audience votes are not, yet both decide who goes home. This page walks through the three ideas our team's paper was built on, using one invented week of six couples.
Nothing here is from our paper's data or results: the scores, votes and couples are invented, and the code was written for this page. It shows the logic of the approach, not our estimates for the 34 real seasons. Where the rules leave a detail open, the choice made here is stated.
One week, three rules
- Rank sum: add each couple's judges' rank and audience rank; the worst total goes.
- Percentage sum: add each couple's share of all judges' points and its share of the votes; the lowest goes.
- Judges' save: the bottom two by rank sum face the judges, who keep one (here, the one with the higher judges' score).
- Ties go against the couple with fewer votes.
Your week
| Couple | Judges' total (of 30) | Audience votes |
|---|
From the elimination back to the votes
In the real data only the judges' scores and the eliminated couple are known. Constrained Monte Carlo draws many possible vote splits and keeps the ones under which the rule really eliminates that couple. What survives is the range of votes the public record allows. The audience votes from section 1 are not used here.
Vote splits are drawn from a symmetric Dirichlet distribution: α = 1 treats every split as equally likely, larger α favours even splits. The prior matters most where the record says least, which is usually the couples near the top.
Replaying another rule
With the accepted vote splits, each rule can be replayed on the same week. The bars show, over all accepted samples, who each rule would have eliminated. This is how a rule comparison can carry the uncertainty of the estimates with it instead of relying on one guess.
A replay assumes viewers would have voted the same way under another rule, which real audiences might not do.
What this page is and is not
- It illustrates the three steps named on the project page (rule replay, constrained Monte Carlo, comparison of rules) on invented numbers.
- It does not reproduce our paper's model, which also used ensemble regression, XGBoost with SHAP and sensitivity tests across 34 seasons. The paper, data and code are not published here.
- Choices made only for this page: six couples, judges' totals out of 30, a symmetric Dirichlet prior, 20,000 samples per run, ties against the couple with fewer votes, and judges who save the higher-scoring couple.