Time-Varying Home Field Advantage in Football: Learning from a Non-Stationary Causal Process

Minhao Qi, Hengrui Cai, Guanyu Hu, Weining ShenarXiv 2025 · DOI · arXiv

Abstract

In sports analytics, home field advantage is a robust phenomenon where the home team wins more games than the away team. However, discovering the causal factors behind home field advantage presents unique challenges due to the non-stationary, time-varying environment of sports matches. In response, we propose a novel causal discovery method, DYnamic Non-stAtionary local M-estimatOrs (DYNAMO), to learn the time-varying causal structures of home field advantage. DYNAMO offers flexibility by integrating various loss functions, making it practical for learning linear and non-linear causal structures from a general class of non-stationary causal processes. By leveraging local information, we provide theoretical guarantees for the identifiability and estimation consistency of non-stationary causal structures without imposing additional assumptions. Simulation studies validate the efficacy of DYNAMO in recovering time-varying causal structures. We apply our method to high-resolution event data from the 2020-2021 and 2021-2022 English Premier League seasons, during which the former season had no audience presence. Our results reveal intriguing, time-varying, team-specific field advantages influenced by referee bias, which differ significantly with and without crowd support. Furthermore, the time-varying causal structures learned by our method improve goal prediction accuracy compared to existing methods.

Introduction

Home field advantage (HFA) is one of the most enduring observations in professional sports. While its existence is undisputed, the underlying causal mechanisms—ranging from crowd intimidation and referee bias to player familiarity—are difficult to isolate. Traditional sports analytics often treats a match as a single static data point or a stationary process where the "rules" of the game do not change. However, a football match is inherently non-stationary: tactics evolve, player fatigue sets in, and psychological momentum shifts minute-by-minute.

To capture this complexity, we introduce DYNAMO (DYnamic Non-stAtionary local M-estimatOrs), a causal discovery framework designed to uncover how these relationships evolve over the course of a 90-minute match. By applying DYNAMO to English Premier League (EPL) data, we demonstrate that HFA is not a static baseline but a fluid process heavily influenced by the presence of a crowd.

Methodology: The DYNAMO Framework

The primary challenge in modeling a football match is the violation of stationarity. Existing causal discovery algorithms typically assume that the causal graph remains constant over time. DYNAMO overcomes this by utilizing the concept of local stationarity. While the causal structure at the start of a match may differ significantly from the final minutes, the structure at any given minute $t$ is likely very similar to the structure at $t+1$.

DYNAMO employs a localizing kernel function (such as an Epanechnikov kernel) to weight observations near a specific time step. This allows the model to approximate a stationary process within a narrow temporal window. The framework is designed as a flexible wrapper for Directed Acyclic Graph (DAG) learning:

  • Linear DYNAMO: Utilizes the NOTEARS loss function for linear time-varying structures.
  • Non-linear DYNAMO: Employs NTS-NOTEARS (using Convolutional Neural Networks) to capture complex, non-linear interactions.

Crucially, DYNAMO provides mathematical guarantees for identifiability and estimation consistency without requiring the strict "faithfulness" or constant Gaussian noise assumptions that often limit traditional methods.

The "Ghost Game" Natural Experiment

To test the impact of crowds on HFA, we analyzed high-resolution event data from two distinct EPL seasons: 2020-2021 (the "ghost games" played in empty stadiums due to COVID-19) and 2021-2022 (the return of full crowds). The dataset includes over 1.6 million events across 760 games, tracking variables such as passes, shots, fouls, and yellow cards to predict Expected Goals (XG).

Key Insights

  • Dynamic Referee Bias: Our model reveals that referee bias is not a constant factor. For elite teams like Liverpool, the causal link between referee bias and home advantage grew steadily throughout the match—but only when fans were present. In empty stadiums, this effect was largely absent.
  • Tactical Interactions: Referee behavior appears to interact with team playstyles. For aggressive, high-possession teams like Manchester City, referees were more likely to issue early yellow cards to opponents when crowds were present, potentially as a response to the atmospheric pressure to control the game.
  • The Rank Divide: The return of crowds disproportionately benefited top-tier teams. Higher-ranked teams experienced significant early-match referee bias in full stadiums, whereas lower-ranked or relegated teams saw negligible differences in refereeing patterns between the two seasons.

Results and Performance

DYNAMO demonstrates superior performance in both synthetic and real-world applications. In simulation studies, DYNAMO consistently achieved higher F1 scores and lower Structural Hamming Distances (SHD) than baselines like DYNOTEARS and PCMCI+.

In real-world predictive tasks, accounting for time-varying causal structures led to more accurate goal predictions. For example, when predicting XG for Arsenal during the 2020-2021 season, Linear DYNAMO reduced the Mean Squared Error (MSE) to 0.00046, compared to 0.00061 for the stationary DYNOTEARS model.

Limitations and Future Directions

While DYNAMO offers a significant step forward, it faces certain bottlenecks. The non-linear version is computationally intensive, and the model currently requires a manual choice between linear and non-linear base learners. Furthermore, while team-level aggregates provide strong insights, future research could apply these techniques to player-level data to account for individual spatiotemporal contributions. Expanding the variable scope to include specific events like penalty kicks or VAR decisions could further refine our understanding of the causal drivers of home field advantage.

Summary updated

Abstract

In sports analytics, home field advantage is a robust phenomenon where the home team wins more games than the away team. However, discovering the causal factors behind home field advantage presents unique challenges due to the non-stationary, time-varying environment of sports matches. In response, we propose a novel causal discovery method, DYnamic Non-stAtionary local M-estimatOrs (DYNAMO), to learn the time-varying causal structures of home field advantage. DYNAMO offers flexibility by integrating various loss functions, making it practical for learning linear and non-linear causal structures from a general class of non-stationary causal processes. By leveraging local information, we provide theoretical guarantees for the identifiability and estimation consistency of non-stationary causal structures without imposing additional assumptions. Simulation studies validate the efficacy of DYNAMO in recovering time-varying causal structures. We apply our method to high-resolution event data from the 2020-2021 and 2021-2022 English Premier League seasons, during which the former season had no audience presence. Our results reveal intriguing, time-varying, team-specific field advantages influenced by referee bias, which differ significantly with and without crowd support. Furthermore, the time-varying causal structures learned by our method improve goal prediction accuracy compared to existing methods.
BibTeX
@misc{qi2025time,
  author = {Minhao Qi and Hengrui Cai and Guanyu Hu and Weining Shen},
  title = {{Time-Varying Home Field Advantage in Football: Learning from a Non-Stationary Causal Process}},
  year = {2025},
  eprint = {2506.11399},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.11399},
  howpublished = {arXiv},
}
Figure