Retrospectively validated transfer analysis — tested on real transfer records; AUC 0.83 on the fully unseen 2025 season (95% CI 0.80–0.85, n=1,747), calibration ±3. Three independent axes, never collapsed into one score: Retention (will he play/stay — calibrated probability) · Style-Role (is he the profile you want) · Level (does the move make career sense). + Demonstrated-level (DL) divergence warning and league-translation layer.
Only permanent transfers are modeled: loan returns (retention 26.5%) and winter moves (49.9% vs summer 76.0%) were label noise — removed. Extra signals with validated retention contribution: FM adaptability + professionalism. Rejected (leaked or turned out to be CA in disguise; not in the model): transfer fee, market value, injury history, position competition, tactical fit. Robustness: tested out-of-year across 5 seasons (2021-2025, 6,229 transfers); probabilities within ±3 in every band.