2026 Olympic Mixed Doubles Projections

Methodology

 

Training Data: WMDCC2022 & 2022 Olympics

I developed the model using training data from the World Mixed Doubles Curling Championships from 2016-2025, and from the 2022 Olympic Winter Games Mixed Doubles.

 

Core rating system: Elo

I use an Elo framework where the probability pair A beats pair B depends on the difference in their ratings. Ratings are updated after each game.

 

Margin-of-victory scaling

To reflect stronger evidence from larger wins, I multiply the update size by a soft margin-of-victory term:

 

                   MOV = log(margin + 1)

 

​This increases responsiveness without letting blowouts completely dominate.

​

Country priors + dynamic shrinkage (τ = 20)

Pairs are not independent across time: countries select pairs, and pair composition can change. To handle pair turnover (selection effects), I maintain both:

• a pair rating (pair-specific strength)

• a country rating (baseline strength for that nation)

When predicting a game, I use an effective rating for each pair:

​

​
 

where n is the number of games the pair has played so far and

​

​
 

With τ = 20, new pairs are treated mostly as “their country” until enough evidence accumulates to trust the pair rating. More specifically, when the model has 20 games of data from a pair,  their performance in those games and their country rating both have an equal impact on the pair's effective rating.

​​

Tournament-level fast adaptation

Teams often exhibit tournament-specific deviations (e.g. because of ice conditions, health, strategy, “form”). To capture higher correlation within an event, I allow faster learning during the tournament by multiplying the Elo K-factor by an event weight of 2 when forecasting sequentially within that tournament. In other words, when forecasting the 2026 Olympics, a game played during the 2026 Olympics will impact a teams Elo rating twice as much as a game played in an earlier tournament. This leaves long-run historical strength intact while letting the model adapt quickly to what is happening within a tournament. 

 

No time decay

​Explicit time decay underperformed when backtesting in my data. In other words, mixed doubles pair/country strength appears stable enough that older results remain predictive of current performance. The model instead handles “recency-like” effects through:

• within-tournament fast adaptation (event_weight)

• selection-aware country shrinkage (τ)