Would this transfer actually work?
Real transfers, read with only what was known before the signing: goals, the risk his minutes collapse, what the coach's forwards did, and what the data can't see. Free public data.
What counts as a flop? You decide.
Every established player who changed club in the big 5, scored with only what was known at signing. Set your own definition of a flop and see whether the warnings held up - misses included.
Three transfers, read before they happened
One open case and two backtests where we know the outcome - including where the model was wrong. Each number says how far it can be trusted.
Score any forward
Any big-5 forward: fit to the Juventus profile, style match and projected Serie A output, computed live.
Rigor over decoration
The interesting part is not the score - it is knowing which numbers to trust, and saying so.
Signal vs noise
Finishing on 25 shots is variance, not skill - so it is shrunk toward expectation. Chance generation and shot volume stabilise fast, so they are trusted. The model says which is which instead of over-reading a goal drought.
Openda's 1-goal season = ~10th-percentile luck, not a forecastable failure.
A validated model, not a guess
Next-season scoring is learned from ~11k cross-league player-seasons and validated out-of-sample on a temporal split, with a player-clustered bootstrap CI.
Beats the naive baseline ~8% (MAE 0.075 -> 0.069), CI excludes zero.
Learned, not asserted
The league-transfer factor was once a hand-set 0.90. An adversarial review flagged it; now it is read from ~1.1k real transfers. Moving into the Premier League suppresses everyone - the data agrees with what analysts know.
Bundesliga -> Serie A retention ~1.0, learned from the transfers.
Honest by construction
The DNA score is transparent stats (robust-z + cosine), not a black box. ML is used only where real labels exist. Culture / mentality has no honest public model, so it is surfaced as evidence, never faked.
Every build was adversarially reviewed and hardened.