A six-market GBM stack (total bases, runs, RBI, singles, hits, HR) with isotonic calibration + market-edge weighting on the AU-quotable OVER side. Walk-forward out-of-sample: the lead Top-3 pool holds +26.4% TEST / +53.9% on the 2026 holdout; the served Premium pool runs +5.4% / +8.6% across ~11-12 picks a slate. The genuine under edge stays geo-fenced US, out of scope here.
A two-pass gradient-boosted model with a learning-ledger error-attribution layer feeding the second pass, plus a per-market meta head that fades the biggest raw-edge-vs-book picks (the far-alt traps) and keeps the high-confidence ones.
The edge is structural, not a placebo: books over-price the OVER on counting props, so the UNDER is systematically underpriced. The full player-prop book holds +5.78% out of sample on the realizable-price-corrected close, with runs and RBI the strongest meta markets.
The model is promotion-ready, judged on walk-forward plus a held-out late window. It comes online as its own tier on the go-ahead — until then no MLB pick or record is published here. Fail-closed, same as every channel.
Walk-forward out-of-sample — model trained only on prior folds. TEST + 2026 holdout are the deployable numbers; last-30 is the recent-window check. ROI on staked units.
AU over-markets (total_bases, home_runs, runs, rbi, singles, hits) · walk-forward TEST + 2026 holdout
MLB player-prop research is built on the same per-market spine — runs, RBI, hits, total bases, strikeouts — and comes online with the channel.
How the model worksThe finder and alt ladders extend to MLB markets the moment the book is promoted — every rung priced, best-EV flagged, NO LINE where a book is missing.
About the finderFree through the NBA Playoffs — evening delivery, public record, cancel anytime.
