ScoutGPT · ECML PKDD 2026

Son left Tottenham in 2025.
Who should they
have signed?

A season on, ScoutGPT replays every possession he touched in his last year at the club, with someone else in his place.

Modeling Matches as Language: A Generative Transformer Approach for Counterfactual Player Valuation in Football

Loading…

How the seasons were split
Train 2020/21, 2021/22, 2022/23, and the first half of 2024/25 1,329
Valid 2023/24, kept whole for early stopping 380
Test Second half of 2024/25 191

Son's counterfactual runs over his 2024/25 possessions, so part of it sits inside the training half. That is why the model replays Son as himself. On the held-out half his self-replay matches his real record within 1%, which is the control for the rest.