One Metric, Different Jobs: Target-Specific Feature Admission for Major League Baseball Player Projection

Baseball metrics are often discussed as though they possess one universal level of predictive usefulness. In practice, a measurement may improve one forecast while serving primar- ily as a diagnostic or contextual variable for another. This study develops and applies a target-specific framework for determining when a candidate baseball metric should alter a projection of future player performance. Using Major League Baseball player-season data from 2015–2025, component-specific candidate features are evaluated against next-season weighted on-base average (wOBA), Fielding Independent Pitching (FIP), baserunning run-value rate, fielding run-value rate, and Baseball Reference WAR rates for hitters and pitchers. A global registry of 402 fields yields 208 feature–target admission evaluations under component-specific eligibility; this is not a full Cartesian product of every field against every target. All headline results use seven expanding-window temporal folds with outcomes from 2019–2025; no headline result uses a random row split. Candidates are assessed for incremental out-of-time predictive value, temporal stability, redundancy, coverage, subgroup robustness, and target dependence. Results show that feature usefulness depends on the outcome being projected. Two-year expected wOBA and exit velocity improve next-season wOBA forecasts beyond a baseline already containing two-year wOBA, whereas Barrel%, Hard-Hit%, chase rate, and sprint speed do not earn Projection status for that target. Stuff+ improves next-season FIP pre- diction beyond a two-year FIP baseline. Sprint speed provides essentially no incremental value for next-season wOBA but improves next-season baserunning projections in all seven temporal folds. Selective admitted-feature models outperform broader all-feature models for wOBA and FIP, but not universally: the all-feature model performs better for hitter WAR rate. These results suggest that baseball feature selection should be framed around a specific forecasting target and judged by incremental temporally validated information rather than contemporaneous association or mechanistic appeal alone