Integrating Unsupervised and Supervised Learning for the Prediction of Defensive Schemes in American football

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.

💡 Research Summary

The paper tackles the problem of predicting whether an NFL defense will employ man‑to‑man or zone coverage before the snap, a task that is notoriously difficult because defensive assignments are deliberately concealed. Leveraging the rich player‑tracking data released as part of the NFL Big Data Bowl 2025, the authors construct a multi‑stage feature engineering pipeline and combine unsupervised and supervised learning to improve prediction accuracy and provide interpretable insights.

Data and preprocessing
The authors use 10 Hz tracking data and play‑by‑play metadata from the first nine weeks of the 2023 season, focusing on plays that contain pre‑snap offensive motion. After filtering out plays with multiple quarterbacks or “bunch” formations, they retain 3,963 offensive plays (2,980 zone, 983 man). For each play they keep the five offensive skill players (WR, TE, RB, etc.) and, after removing defensive linemen and pass‑rushers, select the five defensive players most likely to be guarding them. This yields 19,815 time series (5 defenders × 3,963 plays) that preserve the raw temporal structure of the data.

Feature construction
The authors build three successive feature sets:

1. Pre‑motion features – contextual variables (quarter, down, yards‑to‑go, score, time left) and spatial descriptors derived from the convex hulls of offensive and defensive players (area, width, length). They also include 20 standardized positional features for the ten retained players.

2. Naïve post‑motion features – six additional variables capturing the maximum x‑ and y‑displacements and total distance traveled by each team between the start of motion and the snap.

3. HMM‑derived features – a hidden Markov model (HMM) that treats each defender’s y‑coordinate as an observation generated from a mixture of five Gaussian state‑dependent distributions, each centered on the y‑coordinate of one offensive player. The hidden state indicates which offensive player the defender is currently guarding. The model incorporates a lag parameter l to approximate defender reaction time (e.g., l = 5 corresponds to a 0.5 s delay). Transition probabilities are modeled with a mixed‑effects multinomial logistic regression: they depend on the absolute distance between the two offensive players being considered, and include random intercepts for defensive role, team, and individual play. This non‑homogeneous HMM captures realistic switching behavior and accounts for systematic heterogeneity across positions, teams, and plays. After fitting the HMM via an EM algorithm with Bayesian priors, the authors decode each play’s most likely state sequence (Viterbi path) and summarize it into a set of statistics (e.g., proportion of time a defender is assigned to a particular offensive player, number of switches, average lag). These summaries become the HMM‑derived features.

Supervised learning models
Two predictive models are trained on each feature set:

• Elastic Net logistic regression – combines L1 and L2 penalties to perform variable selection while controlling over‑fitting.
• XGBoost (gradient‑boosted decision trees) – captures non‑linear interactions and complex patterns.

Model performance is evaluated using accuracy, AUC, log‑loss, and F1‑score with cross‑validation. Adding the HMM‑derived features yields consistent improvements: AUC increases by roughly 0.04–0.07, log‑loss drops by more than 10 %, and overall classification accuracy rises compared with models that only use pre‑motion or pre‑plus‑naïve post‑motion features. XGBoost benefits the most from the richer feature set, reflecting its ability to exploit the nuanced information encoded in the HMM summaries.

Statistical significance and interpretability
To assess whether the HMM features contribute information beyond the other variables, the authors apply the non‑parametric Generalized Covariance Measure (GCM) test, a conditional independence test. The GCM results reject the null hypothesis of conditional independence, confirming that the HMM features are significantly associated with the coverage outcome even after conditioning on all other covariates. Moreover, the mixed‑effects structure of the HMM provides interpretable random‑effect estimates: for example, cornerbacks exhibit positive random effects for man coverage, indicating a systematic propensity toward man assignments, while safeties show negative effects, reflecting a tendency toward zone schemes. Team‑level random effects reveal that certain defenses (e.g., Kansas City) are more adept at using pre‑snap motion to elicit diagnostic defensive reactions, thereby improving the predictability of their coverage.

Discussion and limitations
The study demonstrates the power of integrating unsupervised latent‑state inference with supervised classification in a sports analytics context. By extracting hidden defensive assignments from raw tracking data, the authors move beyond the traditional, subjective film‑based analysis of pre‑snap motion. The framework yields both higher predictive performance and actionable tactical insights for coaches and analysts. Limitations include the reliance on y‑coordinates only (ignoring x‑position and velocity), the use of a fixed reaction‑time lag, and the assumption of independence across defenders’ trajectories. Future work could extend the HMM to a multivariate formulation (incorporating x, y, speed, and acceleration), explore Bayesian non‑parametric state models, or integrate real‑time inference for in‑game decision support.

Conclusion
Overall, the paper presents a well‑designed statistical pipeline that fuses hidden Markov modeling of latent defensive assignments with elastic‑net and gradient‑boosted classifiers. The inclusion of HMM‑derived features substantially improves the accuracy of predicting man versus zone coverage and provides interpretable random‑effect estimates that illuminate team‑ and position‑specific defensive tendencies. This contribution advances the quantitative analysis of football tactics and offers a template for similar applications in other sports where hidden player interactions drive strategic outcomes.