Model cards Opportunity Finder

Opportunity Scores

Outside players ranked by the wins each adds to the Winter Plan’s roster; skill signals explain why he’s listed, computed checks test them, and neither sets the rank.

Supports: Which outside players deserve a scout’s time and a call this winter. Refresh: Weekly in offseason Open Opportunity Finder

What it predicts

Wins added to the plan’s recommended 40-man through the depth-chart engine, the same measurement as the Free-Agent Board.

Inputs

Keystone Stuff and its gap to results
Pitch Lab flags
Hitter decision, contact, and bat-tracking metrics
Contract status, surplus value, and estimated free-agent cost

Method

Listed: outside players with enough 2026 time and at least one scored signal (results lagging Stuff or contact quality, an underused elite pitch, a dead-weight pitch, a plus bat, an approach fix, a lefty-killer edge that holds up once regressed)
Ranked by wins added: the candidate on the plan’s roster with his package out and the 40-man trimmed by the plan’s cut order; players their clubs are unlikely to trade, or trades no package of three prospects can pay for, rank after every other row
Price in two currencies never added together: the prospect package his club would want (the least value meeting its ask, top 3 protected) and his salary
Results lagging Stuff: a time-ordered joint regression (Stuff graded by a model trained on 2024 only, 2025 → 2026) of next season’s results on this season’s results and the gap together; a pitcher’s upside is the joint forecast minus the results-only one, with an 80% range
Every flag comes with computed checks (where he chases, whether a split held up, luck vs. profile); scouts keep what numbers can’t answer. A Stuff-over-results gap reads as “results lag stuff” only when its check forecasts a run or more; otherwise “no forecast gain”, which doesn’t list a pitcher on its own
Salary under the owners’ proposed floor is shown at both ends of its shadow price: free if the club has no other use for the floor money, face value if another use would take it

How well it works

Computed by the pipeline from the data on this site

Backtest: pitchers flagged on 2025 data (poor results, Stuff ≥ 105) vs. peers with poor results and Stuff < 100, Stuff graded by a model trained on 2024 only: Flagged +0.37 runs/100 pitches in 2026 (n = 36), vs. +0.26 for controls (n = 111): a gap of 0.8 standard errors. Both groups regress upward. The flag’s extra rebound is inside the noise. On the old cross-fit grades, whose 2025 model had seen 2026, the gap looked like +0.59 vs. +0.24.
What a Stuff-over-results gap adds, holding results equal (357 pitchers, 2025 → 2026, Stuff from a model trained on 2024): Each run per 100 pitches of gap came with 0.81 ± 0.12 runs of better results the next season This joint regression is what the Opportunity Finder uses to size a pitcher’s upside. The old slope on the gap alone (0.71) was mostly results regressing to the mean.
Backtest: hitters whose 2025 xwOBA beat wOBA by .025+ vs. hitters within ±.010: +0.024 vs. −0.002 wOBA; the gap is 2.3 standard errors (95% interval +0.002 to +0.049, Welch t) Clears the noise on a small flagged sample (15 hitters).

Data: Keystone Stuff (Statcast 2024–26); hitter Statcast metrics; RosterResource contracts.

What it can't see

How we'd measure impact

Hit rate of flags in backtests
Surplus of acquisitions the tool sourced
Scout confirmation rate