DraftFM Predicts Magic: The Gathering Card Picks.

This research introduces DraftFM, a specialized foundation model aimed at predicting optimal card selections in new Magic: The Gathering (MTG) expansions immediately upon release, a period referred to as "day-zero" when no human draft data is available for supervised learning. The model operates solely on publicly available information about cards, including structured features and fixed text embeddings, without relying on card identities, set identities, or usage statistics. This design allows it to score unseen cards using the same mechanism as familiar ones. DraftFM is a discrete-choice policy that evaluates cards available in a current pack, considering the already drafted pool and the overall draft state. Trained on 149 million human picks from 29 expansions, the 1.6-million-parameter network demonstrated strong predictive capabilities on three entirely withheld expansions, achieving top-1 agreement rates of 50.8%, 60.4%, and 56.7%. This significantly outperforms uniform chance, which is around 7% for an opening pick. Furthermore, the architecture, after being refitted on all 32 observed expansions, generated a card ranking for the then-unreleased set "The Hobbit." This ranking, cryptographically sealed and published before the set's release, showed agreement with six independent expert reviewers comparable to the agreement among the reviewers themselves. The study highlights the model's ability to generalize to entirely new game content based on its inherent features.