ahanjain/mls-wage-estimator

MLS Wage Estimator

This export contains the current v4_career_ridge MLS wage model artifact and the exact linked training cohort used to fit it.

Cohort

• Foreign MLS arrival wage rows from 2020+
• Verified Transfermarkt ID required
• Current training rows: 922
• Career rows flagged missing before defaulting: 0
• DP tail target compression above $1,803,125 with weight 0.35

Files

• wage_model_v2.joblib: trained wage model artifact
• wage_model_v2_training_data.csv: linked training cohort used for the current artifact
• tm_arrival_lookup_cache.parquet: TM lookup cache used to resolve direct-arrival wage rows
• tm_career_cache.parquet: cached Transfermarkt career histories used to compute career features
• metrics.json: model metadata and current performance snapshot

Current Metrics

• Train log-MAE: 0.501
• Train within 50%: 64.0%
• Train rank corr: 0.781
• Train spread ratio: 1.028
• OOS log-MAE: 0.489
• OOS within 50%: 66.3%
• OOS rank corr: 0.808
• OOS spread ratio: 1.096

Notes

The model is a Ridge regressor over age, season, minutes, position, and TM-derived career features. It uses the cleaned 922-row foreign-arrival cohort with only usable TM career profiles, live-TM-only app lookups, and post-calibrated predictions with a monotone quantile map on the log-wage target. The live default was moved from XGBoost to Ridge because the smoother response surface produced more credible player-by-player wage estimates while keeping the overall backtest competitive.