Trained with a dataset of over 1,000 penalty kicks, the deep learning model demonstrates superior predictive capabilities compared to actual goalkeepers.

“Penalty kicks are often decisive moments in football that can influence the results of important tournaments,” states David Freire-Obregón from the University of Las Palmas de Gran Canaria, Spain. “Yet, real-time support for goalkeepers mostly relies on intuition. We sought to determine if machine learning could effectively predict the direction of a shot based on the kicker’s movements.”

Freire-Obregón and his team analyzed 1,010 penalty kicks from televised matches in Spain. Out of these, 640 were deemed usable by the AI, while the remainder were excluded due to blurriness, being too brief, or other obstructions.

Each video clip was processed through 22 different deep learning models. The goal was to predict whether the penalty kick would go left, right, or center left, or lower right, depending solely on the player’s body posture and footedness.

The top-performing model was able to accurately identify the direction of the ball—right, left, or center—52% of the time. Excluding the middle option improved the model’s accuracy to 64%.

The researchers were astonished, stating, “We can uncover the intent behind subtle movements, which serve as clues before the ball is kicked,” says Freire-Obregón. He hopes this insight will aid training for goalkeepers, although he admits utilizing AI predictions in actual matches presents additional challenges.

“Our next goal is to determine if we can predict the outcome of a penalty kick solely based on the kicker’s pre-shot movements,” he adds. “If feasible, how quickly can such predictions be made while retaining acceptable accuracy?”