In this thesis, we analyzed the problem of predicting coaches’ decisions in
fourth-down situations (4th down) in American football using machine learning methods and play-by-play data. Using NFL games from the 2018–2024
seasons, we constructed a dataset of fourth-down situations and designed a
feature space that included characteristics of the current game state, weather
conditions, betting predictions, and team statistics from previous games. To
predict the three possible coaching decisions, we applied multiclass classification models, with XGBoost achieving the best performance, reaching a
classification accuracy of 89.8 % and an F1 score of 0.857. Feature importance
analysis showed that the model relied primarily on field position, expected
points, distance to the first down, and remaining game time. The results
demonstrated consistency between the model predictions and commonly observed strategic decision-making patterns in American football.
|
