MCM 2026: Modelling Audience Votes

Estimated the unpublished audience votes behind 34 seasons of Dancing with the Stars eliminations, then used the estimates to compare how different scoring rules would change the results.

Period
2026
Context
Mathematical Contest in Modeling (MCM/ICM), COMAP
Team
Team of 3
My role
Team leader
Status
Paper submitted within the contest; judged Meritorious Winner
Recognition
Meritorious Winner, MCM/ICM 2026 (Team 2615257)
  • Ensemble regression
  • Constrained Monte Carlo
  • XGBoost
  • SHAP
  • Sensitivity analysis

Demo & artifacts

A method explainer on an invented week of six couples: three elimination rules, estimating hidden audience votes from an elimination by constrained Monte Carlo, and replaying other rules on the estimates. It was written for this site and does not reproduce our paper's data, model or results.

My contribution

  • Led a three-member team through the contest. I divided the work and organised how we checked each other's assumptions against the deadline.
  • Worked with my teammates on the team's approach, which combined ensemble regression with constrained Monte Carlo estimation, replays of alternative scoring rules, and SHAP-based interpretation.

The methods, the paper and the award below belong to the team as a whole.

The problem

In each season of the show, contestants are eliminated by combining judges’ scores with audience votes. The scores and the eliminations are public, but the votes are not. The task was to estimate the hidden votes from 34 seasons of records and use them to examine how different rules for combining scores and votes would change the results.

How it works

Audience-vote estimation pipelineJudges' scores and elimination records from 34 seasons go into ensemble regression and constrained Monte Carlo sampling, constrained so the estimates reproduce observed eliminations. The resulting vote estimates feed three analyses: replaying ranking, percentage and judges' save rules; interpreting an XGBoost model with SHAP; and sensitivity tests under perturbation.Observed data34 seasons of judges' scoresand elimination recordsEstimate hidden votesensemble regression +constrained Monte CarloVote estimatesconsistent with outcomes;not the true votesconstraint: estimates mustreproduce each eliminationRule replayranking, percentage, judges' saveInterpretationXGBoost explained with SHAPSensitivitystability under perturbationWHAT THE ESTIMATES WERE USED FOR
Modelling pipeline. Audience votes were never published, so the model searches for vote estimates that are consistent with what was observed: judges' scores and who was eliminated each week. The estimates were then used to replay alternative scoring rules and were checked with SHAP and perturbation tests.

The central idea is to accept only estimates that agree with what actually happened. Under the rule used in a given season, a set of estimated votes must, combined with the judges’ scores, reproduce that week’s elimination. We combined ensemble regression with constrained Monte Carlo methods to produce estimates that meet this condition across the seasons.

With estimates in hand, we replayed three scoring rules (ranking, percentage and judges’ save) to see how outcomes would differ under each. We then used SHAP to interpret an XGBoost model and perturbed the inputs to check whether the conclusions held.

What the results can and cannot claim

Because the true votes are never observed, no model can be checked against them directly. Two indirect checks are available: whether the estimates reproduce the observed eliminations, which the constraint enforces, and whether the conclusions stay stable when the inputs change, which the sensitivity analysis tests. A statement about how a scoring rule changes outcomes is only worth making if it survives reasonable perturbations of the estimates behind it.

Results & validation

  • Meritorious Winner, MCM/ICM 2026 (COMAP), Team 2615257. This is a team award.
  • Produced audience-vote estimates for 34 seasons that are consistent with the observed judges' scores and elimination records.
  • Replayed the ranking, percentage and judges' save rules on those estimates to compare how each rule would change outcomes.
  • Interpreted the XGBoost model with SHAP and tested how the results respond to perturbations of the inputs.

Estimated votes are a model output, not recovered ground truth. Many vote patterns can produce the same eliminations, so the estimates describe what is consistent with the public record.

Limitations & next steps

  • Observed eliminations only partly constrain the votes, so different vote patterns fit the same record. Conclusions about the scoring rules are only as strong as the estimates behind them.
  • A rule replay assumes viewers would have voted the same way under a different rule, which real audiences might not do.
  • SHAP explains how the fitted model uses its inputs. It describes the model, not the causes of real voting behaviour.