Full-History Bradley-Terry (FH-BT)

• Full-History Bradley-Terry (FH-BT) is a statistical model that unifies explanatory analysis and online predictive updates for paired comparisons.
• It leverages full-history maximum-likelihood estimation and stochastic gradient ascent to refine latent strength parameters in competitive settings.
• The framework extends classical methods by incorporating draws, covariates, and low-rank constraints to improve predictive performance.

The Full-History Bradley-Terry (FH-BT) framework is a generalization of the classical Bradley-Terry model for paired comparison, designed to unify the strengths of explanatory statistical modeling and predictive, supervised learning for outcomes such as competitive sports results. FH-BT introduces maximum-likelihood learning over the entire history of pairwise outcomes, supports batch and online (Élő-style) updates, integrates covariates and features in a generalized linear model (GLM) framework, and extends naturally to incorporate draws, low-rank structural constraints, and matrix completion interpretations. Its validity and efficacy have been experimentally confirmed on synthetic data and English Premier League outcomes, demonstrating state-of-the-art predictive performance (Király et al., 2017).

1. Relationship to Classical Bradley-Terry and Élő Systems

The FH-BT model extends the paradigms of both the classical Bradley-Terry (BT) model and the Élő rating system. In the BT model, each contestant or team $i$ is described by a latent "strength" parameter $\theta_i$. The probability that team $i$ defeats team $j$ is modeled as

$p_{ij} = \sigma(\theta_i - \theta_j + h)$

where $\sigma(x) = (1 + e^{-x})^{-1}$ is the logistic sigmoid and $h$ is an intercept accounting for home advantage. However, while the BT model is typically used for explanatory analysis, the Élő rating system is a widely used heuristic baseline for online predictive updating.

FH-BT exploits the mathematical closeness between these models, as well as similarities to logistic regression, to create a unified statistical approach that supports supervised probabilistic prediction and efficient online updates. Unlike the original Élő algorithm—which is not a likelihood-based statistical model—FH-BT provides a principled probabilistic interpretation and a direct connection to GLM and modern matrix completion schemes (Király et al., 2017).

2. Maximum Likelihood Fitting Over Full History

The core of FH-BT is the batch ("full-history") maximum-likelihood estimation of the latent strengths $\theta$. For a historical dataset $D = \{(i_k, j_k, Y^{(k)})\}_{k=1}^N$, the model optimizes

$$\ell(\theta|D) = \sum_{k=1}^N \left[ Y_{i_k j_k}^{(k)} \log \sigma(\theta_{i_k} - \theta_{j_k} + h) + (1 - Y_{i_k j_k}^{(k)}) \log (1 - \sigma(\theta_{i_k} - \theta_{j_k} + h)) \right]$$

This formulation allows one to compute gradients analytically and apply gradient ascent (or minorization–maximization) to obtain a global optimum $\theta_i$0, constituting the "full-history" fit. The batch fitting procedure ensures that latent strengths are informed by the entire sequence of observed pairwise outcomes and not just the most recent matches or updates (Király et al., 2017).

3. Online Updating and the K-Factor

FH-BT supports online learning via stochastic gradient ascent. Upon observing a new outcome $\theta_i$1, a single-step update with step-size $\theta_i$2 ("K-factor") is given by

$\theta_i$3

where $\theta_i$4 is the current predicted probability of $\theta_i$5 defeating $\theta_i$6. This update mechanism is exactly stochastic gradient ascent on the FH-BT likelihood; the one-sample gradient is $\theta_i$7. FH-BT thus generalizes the Élő update, marrying it with the statistical grounding of maximum-likelihood estimation, and enables flexible adaptation as new data arrives (Király et al., 2017).

4. Model Extensions: Draws, Features, and Low-Rank Structure

Draws: Proportional Odds Extension

For outcomes comprising win, draw, and loss (i.e., $\theta_i$8), FH-BT introduces a proportional-odds structure with a shared draw parameter $\theta_i$9: $i$0 where $i$1.

Incorporation of Features

FH-BT naturally accommodates additional information by generalizing $i$2: $i$3 or, more generally,

$i$4

Structural constraints such as low-rankness, anti-symmetry of $i$5, and shared intercepts are used for parsimony and tractability (Király et al., 2017).

Low-Rank Matrix Structure

The matrix of pairwise log-odds $i$6 is anti-symmetric ($i$7) and, in the classical model, of rank two. Extensions permit factorizations such as

$i$8

and nuclear norm regularization for matrix completion perspectives (Király et al., 2017).

5. Empirical Validation and Predictive Performance

Empirical evaluation of FH-BT has been conducted on both synthetic and real-world datasets. On synthetic data generated from rank-2 or rank-4 log-odds, FH-BT extensions with appropriately matched rank systematically outperform the rank-2 (classical) BT model in accuracy and log-likelihood (Wilcoxon $i$9). In English Premier League data, a two-stage strategy comprising initial full-history batch fitting followed by K-factor updates outperforms pure online or naive batch re-fit approaches (paired t-test $j$0). Incorporating a single covariate for team promotion significantly improves log-loss over the vanilla FH-BT ($j$1).

The best FH-BT variant achieves approximately 52.8% win/draw/lose accuracy (95% CI [50.6%, 55.0%]) and mean log-loss of approximately $j$2, comparing closely to performance based on Bet365 betting odds (54.1%/$j$3) (Király et al., 2017).

6. Significance and Interpretive Summary

The Full-History Bradley-Terry model provides a statistically principled, flexible, and computationally efficient GLM-style probabilistic predictor for pairwise outcomes. Its latent strengths are learned from the entire match history, supports efficient online adaptations, generalizes naturally to multi-outcome and covariate settings, and offers more expressive structural modelling through low-rank embeddings. This synthesis of explanatory and predictive modeling yields highly competitive results and validates the importance of full-history batch fitting augmented by lightweight online adaptation for practical prediction tasks such as sports outcome modeling (Király et al., 2017).