Home/Visitor score margins use current field-player and goalkeeper ratings separately. Shared keepers are time-weighted. Total goals use a recent-weighted average from 123 completed games, not a fixed 21-goal target. Positions are not required.

We also test opponent- and teammate-adjusted player effects. In 83 chronological test games, that candidate missed each team’s score by 2.82 goals on average, versus 2.76 for a simple scoring-average baseline. It has not earned a place in live picks yet and remains in shadow testing. Pair chemistry is not used without stronger evidence.

These are retrospective tests using only earlier dates, not a claim about future accuracy. Historical rating snapshots are unavailable, so the live rating adjustment has not been backtested. New published model predictions save their ratings, lineups, keeper shares, and model version. Older/manual predictions remain unchanged.