Machine Learning for Football Match-Point Prediction. Algorithms Performance on Small Datasets

Match-point prediction is an operationally relevant task in football analytics, supporting tactical preparation, squad management, and performance monitoring. Based on a dataset provided by InStat on the results of a struggling Romanian football team, this study proposes a reproducible and auditable decision-support workflow framed as a supervised binary classification problem that estimates the probability of securing at least one point by match end (90 minutes plus stoppage time) using match data. Predictors are engineered to represent three interpretable constructs: average team age, tactical deployment, and key-player exposure. The workflow was implemented in R using two intertwined frameworks for data processing and exploration (tidyverse) and Machine Learning (tidymodels). Five classification algorithms were benchmarked. Results provide some insights on the classification performance when applied to small datasets. Also, the average team age and key-player minutes emerge as the most important predictors in explaining the variability of point attainment for the reference team.