[Paper Review] Integrating Unsupervised and Supervised Learning for the Prediction of Defensive Schemes in American football

Rouven Michels, Robert Bajons+1 more|arXiv (Cornell University)|Feb 11, 2026

Sports Performance and Training0 citations

TL;DR

The paper presents a hybrid framework that combines a non-homogeneous hidden Markov model with supervised learning (elastic net logistic regression and XGBoost) to predict NFL defensive schemes (man vs. zone) using pre-snap and motion-phase tracking data. HMM-derived features improve predictive performance and offer interpretable insights via random effects and GCM tests.

ABSTRACT

Anticipating defensive coverage schemes is a crucial yet challenging task for offenses in American football. Because defenders' assignments are intentionally disguised before the snap, they remain difficult to recognize in real time. To address this challenge, we develop a statistical framework that integrates supervised and unsupervised learning using player tracking data. Our goal is to forecast the defensive coverage scheme -- man or zone -- through elastic net logistic regression and gradient-boosted decision trees with incrementally derived features. We first use features from the pre-motion situation, then incorporate players' trajectories during motion in a naive way, and finally include features derived from a hidden Markov model (HMM). Based on player movements, the non-homogeneous HMM infers latent defensive assignments between offensive and defensive players during motion and transforms decoded state sequences into informative features for the supervised models. These HMM-based features enhance predictive performance and are significantly associated with coverage outcomes. Moreover, estimated random effects offer interpretable insights into how different defenses and positions adjust their coverage responsibilities.

Motivation & Objective

• Forecast defensive coverage (man vs. zone) before the snap using tracking data from the first nine weeks of the 2024 NFL season.
• Integrate unsupervised latent guarding assignments via a non-homogeneous HMM with supervised models.
• Assess the predictive benefit and interpretability of HMM-derived features in defense scheme classification.
• Explore team- and motion-pattern differences in defensive schemes.
• Provide statistical inference on feature associations using modern conditional independence tests.

Proposed method

• Combine unsupervised and supervised learning in a two-layer framework (HMM inner model, supervised outer models).
• Use a non-homogeneous HMM to infer latent defender–offensive player guardings from vertical defender trajectories, with lagged means to model reaction time.
• Incorporate HMM-derived features into elastic net logistic regression and XGBoost models to predict man- versus zone-coverage.
• Derive HMM features from local decoding probabilities and summarize them into metrics such as most-likely guarded player, switch counts, entropy, and play-specific random effects.
• Employ forward algorithm for efficient likelihood computation and Laplace approximation for integrating random effects in the HMM.
• Evaluate models on metrics like accuracy, AUC, and logloss, and use Generalized Covariance Measure (GCM) tests to assess the significance of HMM features.

Figure 1: This figure illustrates the step-by-step pre-processing for a game between the Arizona Cardinals and the Kansas City Chiefs. From top left to bottom right, we first display all players, then exclude offensive non-skill players, next omit defensive linemen, and finally remove pass-rushing l

Experimental results

Research questions

• RQ1 Can pre-snap and motion-phase tracking features reliably predict NFL defensive coverage as man or zone?
• RQ2 Do HMM-derived features capturing latent guarding assignments improve predictive performance beyond naive motion features?
• RQ3 What is the interpretability and statistical significance of the HMM features in predicting coverage?
• RQ4 How do random effects (play, team, position) inform defense scheme tendencies across games?
• RQ5 Can conditional independence tests reveal significant associations between HMM features and coverage outcomes after conditioning on other variables?

Key findings

• The unsupervised HMM identifies latent guarding assignments and reaction dynamics during motion with a lag parameter (l≈4 tenths of a second).
• HMM-derived features substantially enhance predictive performance of man vs. zone classification over models using only pre- and post-motion features.
• GCM tests show that HMM features are significantly associated with coverage outcomes after conditioning on other features, enabling interpretable inferences about feature influence.
• Random effects capture meaningful heterogeneity across positions, teams, and plays, offering insights into defense-specific tendencies.
• The framework enables analysis of team behavior and motion patterns in exploiting motion to determine defensive schemes.

Figure 4: This figure illustrates how guarding assignments evolve during an example play based on decoded state sequences from our HMM analysis. Lines between offensive and defensive players represent inferred matchups over time. The top-left panel shows initial formations at the line of scrimmage;