Bayesian Imputation for Unplayed Games in Round-Robin Chess Tournaments: Application to Grand Chess Tour, Bucharest 2026

Technical Summary: Bayesian Imputation for Unplayed Games in Round-Robin Chess Tournaments

Problem Statement
Round-robin chess tournaments, particularly at the elite level (e.g., Candidates, Grand Chess Tour), face a critical administrative and statistical challenge when a player withdraws mid-event. Current FIDE regulations (Article 6.6) employ a dichotomous rule based on a 50% completion threshold:

• < 50% games played: All results involving the withdrawn player are annulled.
• ≥ 50% games played: Played results stand, and unplayed opponents are awarded a full point (forfeit).

The authors argue this rule creates arbitrary discontinuities and distorts final standings. Annulment erases legitimately earned victories, while forfeit awards unearned points that ignore the withdrawn player's actual form and the strength differentials of their remaining opponents. In elite fields where standings often separate by half a point, these distortions can alter prize distributions and World Championship qualification outcomes.

Methodology
The paper proposes a Bayesian framework based on Best Linear Unbiased Prediction (BLUP) to impute scores for unplayed games. The approach treats the problem as a missing data issue, optimally combining two information sources:

1. Prior Information: Pre-tournament Elo ratings, which provide an expected score ($E_{i,W}$) based on historical strength differentials.
2. Observed Data: The withdrawn player's actual performance ($\bar{s}_W$) in the games played prior to withdrawal.

The core model assumes a random effects structure where a player's score against an opponent is the sum of the Elo-based expectation, a common random effect representing the withdrawn player's current tournament form ($\beta$), and idiosyncratic noise.

The resulting imputation formula for an unplayed opponent $i$ is:
$$I_{i,W} = E_{i,W} + \frac{n}{n+k} (1 - \bar{s}_W - \bar{E})$$

Where:

• $n$: Number of games played by the withdrawn player.
• $k$: Prior strength parameter (recommended $k=3$ for classical time controls), representing the confidence in Elo predictions relative to observed variance.
• $\bar{E}$: Average Elo expectation of the opponents actually faced.
• The term $\frac{n}{n+k}$ acts as a shrinkage weight, balancing the prior (Elo) against the observed data. As $n$ increases, the estimate relies more on observed form; as $n$ is small, it relies more on ratings.

Key Contributions

1. Formal Statistical Framework: The paper mathematically formalizes the fairness problem and derives the Bayesian BLUP estimator, proving it is consistent, point-conserving ($I_{i,W} + I_{W,i} = 1$), and minimizes mean squared error among linear unbiased predictors.
2. Simulation Validation: A Monte Carlo study involving 180,000 simulated tournaments (18 scenarios varying by withdrawal timing, player form, and rating spread) was conducted using leave-one-out cross-validation (LOOCV) to evaluate predictive accuracy against ground truth.
3. Case Study Application: The methodology was applied to the Grand Chess Tour Bucharest 2026, where GM Alireza Firouzja withdrew after 5/9 rounds with a poor score (1.0/5).
4. Open-Source Tooling: An R Shiny application was developed to allow organizers to implement the method transparently, visualize sensitivity to the prior parameter $k$, and compare outcomes against traditional rules.

Results

• Performance vs. FIDE Rules: The Bayesian BLUP method reduced prediction error (RMSE) by 26% overall compared to the current FIDE dichotomous system.
  • It outperformed the Forfeit rule by 42.8% (largest gains when the withdrawn player was underperforming).
  • It outperformed the Annulment rule by 12.2%.
  • Bayesian imputation achieved the lowest RMSE in all 18 experimental scenarios.
• Bias-Variance Tradeoff: Pure Elo-based imputation had low bias when form matched ratings but failed when form deviated. Pure performance-based imputation was unbiased but had high variance. Bayesian imputation achieved the optimal balance, maintaining near-zero bias while significantly reducing variance.
• Policy Implication (Annulment vs. Forfeit): Even without adopting Bayesian imputation, the simulations indicate that uniform annulment is statistically superior to the current dichotomous system, achieving 15–45% lower RMSE than forfeit across all scenarios. The 50% threshold is shown to be an arbitrary discontinuity.
• Case Study Outcome: In the Bucharest 2026 scenario, Bayesian imputation awarded unplayed opponents 0.55–0.70 points, whereas the forfeit rule awarded 1.0. This preserved the competitive advantage of players who had actually defeated Firouzja, whereas the forfeit rule effectively neutralized those victories by giving equal points to those who had not yet played.

Significance and Claims
The paper claims that the proposed methodology provides a principled, transparent, and mathematically defensible solution to a problem currently handled by arbitrary administrative rules.

• Competitive Integrity: By preserving the relative advantage of players who defeated a struggling withdrawn player, the method safeguards the legitimacy of tournament standings, which is critical for high-stakes events like the Candidates tournament.
• Adaptability: The method adapts smoothly to the amount of available data, avoiding the "cliff" effect of the 50% threshold.
• Recommendations: The authors recommend that FIDE adopt Bayesian BLUP imputation (with $k=3$) for all World Championship cycle events. As a minimum reform, they propose replacing the current dichotomous rule with uniform annulment, which the data shows is superior to the current forfeit-based system.

The paper concludes that while fractional scores (e.g., 0.69) may require communication adjustments (e.g., reporting to one decimal place), the statistical rigor of the method is essential for maintaining fairness in modern elite chess.