ValuCast

1. Start with the current eligible prospect universe and MLBAM identity.
2. Build each player's factual baseball profile: role, age, level, current MiLB stat line, sample size, draft/signing facts, ETA, and availability context.
3. Score the profile through ValuCast's prospect model and universal outcome index.
4. Apply sample, availability, and bucket calibration rules by group, not by name.
5. Rank the final scores into one universal prospect board.

Category Fit is a separate league-settings view. It can help a user understand roster fit, but it does not generate the public prospect rank.

• Can affect score: Prospect Model v0.6, universal outcome context, MiLB performance rates and sample reliability, draft/signing facts, age/level context, and factual availability status.
• Cannot affect score: outside dynasty rankings, outside values, value history, public prospect rankings, and market signals. Those are comparison context only.
• Bucket calibration: applied by rule, not by name. Current rules cover lower-minors pedigree compression, thin upper-level pitcher samples, and upper-level hitters with full samples but limited game impact.

Hitter prospect cards show Closest MLB Shapes: the real MLB seasons (2000–2025, age 26 or younger, 400+ PA) nearest the prospect's MLB-translated K% / BB% / ISO. Rates are compared era-relative — z-scored within each season's population of regulars — so a strikeout rate means what it meant in that run environment. Alongside the named matches we show how the nearest matches with a complete five-season follow-up window actually aged (match seasons 2000–2020 — the ones with a complete five-season follow-up by 2025), deduped by player so no career votes twice. Outcome tiers are measured, with published cut points: 450+ PA/yr with a 110+ era-relative OPS ("regular with an above-average bat"), 400+ PA/yr ("everyday regular"), 150+ ("limited playing time"), under 150 ("faded out") — playing time is the yardstick, so an injury year or a light-workload position (catchers) counts against a player, and 2020 is pro-rated to 162 games.

The same translated axes also produce three transparent component matches: Power (ISO), Contact (K%), and Approach (BB%). Each card names the closest historical season and its measured distance on that single era-normalized axis. We do not turn distance into a match percentage.

Pitcher shape comps use translated K-BB%, K/9, and BB/9. Historical starter and reliever seasons are normalized and matched in separate role pools; mixed historical seasons and prospects with ambiguous current usage are suppressed. Pitcher cards show names, rates, and measured distance only — no outcome odds or future-role claim.

The bias we can't remove, named: comps can only be made to players who actually earned a qualifying young-MLB season. Players who reached that bar and then declined are counted (the "faded out" tier), but a prospect with this exact shape who never got a real MLB look is absent from the pool by construction. That structural survivorship is exactly why a shape comp is a descriptive lens, never an outcome forecast. Comps never feed a ValuCast score, know nothing about defense or speed, and only render for prospects whose translation is high-confidence with a real sample. Rebuilt nightly by the public data pipeline.

A Marcel-style method: recent seasons weighted 5.0,4.0,3.0 (most recent first), regressed to the league mean, with per-component rate (events ÷ opportunities) reconstructed into counts. Hitters are age-adjusted and add Statcast input de-noising (blend actual contact/power toward Savant xBA/xSLG, redistributing into 1B/2B/3B/HR by the player's own extra-base mix). Pitchers use a continuous starter-probability blend so swingmen and converted arms aren't miscategorized; the v1 pitching model has no age curve (a planned improvement) and consumes no third-party projection.

Season weighting is done per component as weighted events ÷ weighted opportunities (not a weighted rate). For a component with event count E and opportunities PA, weights 5.0,4.0,3.0: rate = (5·E₋₁ + 4·E₋₂ + 3·E₋₃) / (5·PA₋₁ + 4·PA₋₂ + 3·PA₋₃).

Regression to the mean over N opportunities with a regression constant n_reg (hitters 1200, pitchers 300): projected_rate = (rate·N + league_rate·n_reg) / (N + n_reg), then the projected count is rebuilt as projected_rate × projected_PA.

Age adjustment (hitters only) nudges hitter projections along an empirical aging curve. The v1 pitching model applies no aging curve.

Hitter de-noising (knobs α): blend the actual rate toward the Savant expected rate, rate* = (1−α)·rate + α·x (x = xBA/xSLG), then redistribute the extra hits/bases preserving the player's hit-type mix.

Pitcher role blend: with p_sp the starter probability, each component is shifted continuously by f[c]^(h_sp − p_sp) — no hard SP/RP split.

Worked example (HR rate). An age-29 hitter (near peak, age factor ≈ 1.0) with no Statcast movement: 30 HR / 600 PA, 26 / 580, 20 / 520 over the last three seasons (weights 5/4/3). Weighted events ÷ weighted opportunities = (5·30 + 4·26 + 3·20) / (5·600 + 4·580 + 3·520) = 314 / 6880 ≈ 0.046 HR per PA. Regressing toward a 0.033 league HR rate with n_reg = 1200 gives (314 + 0.033·1200) / (6880 + 1200) ≈ 0.0438. Projected PA = 0.5·600 + 0.1·580 + 200 = 558, so projected HR ≈ 0.0438 × 558 ≈ 24.4 HR.

We hold out future seasons (2020–2025), project forward from prior seasons only (no peeking), and score against simple baselines. Reported as a mean-absolute-error (MAE) ratio — below 1.0 means lower error than the baseline.

• Hitting vs classic Marcel, n = 1770 qualified hitter-seasons, correlation-win rate 0.519 — essentially even on ranking (a near coin-flip); the edge is in error magnitude on rate stats, not in re-ordering players. qualified hitters (>= 200 actual PA, projectable, has prior season).
• Pitching vs persistence (last-year carry-forward), n = 2234 qualified pitcher-seasons, skill correlation-win rate 0.694 — a modest ranking edge; most of the improvement is smaller error on the rate stats below, not re-ordering pitchers. qualified pitchers (role-specific IP floor, projectable, has prior season).

Per-stat MAE ratios:

Hitting (vs classic): AVG 0.942 · OBP 0.959 · SLG 0.96 · OPS 0.951 — counting stats ≈ 1.0 (de-noising doesn't touch them).

Pitching (vs persistence): ERA 0.632 · WHIP 0.751 · K_9 0.834 · BB_9 0.715 · IP 0.981 · K 0.932.

W, SV and QS are reported separately and treated as lower-confidence — they depend heavily on team decisions and opportunity, not pitcher skill.

Scoring notes — 2026-07-12, the night before first publish

Before the gate matured we ran an independent adversarial audit of the scorecard math (a different AI model family from the ones that built and reviewed it, then every finding re-verified against the raw archive). Everything it found is disclosed here, dated, before the first number published — including what cuts against us.

• Compliance fix, disclosed: the headline decided-rate and control lift had been computed on the all-calls cohort, contradicting the registered maturity rule above ("headline cohort = calls at least 14 days old"). Pinned to the matured cohort before first publish. On that day's data the decided rate moved up (24.6% → 28.7%) and the lift moved down (1.36x → 1.31x); both still miss our pre-registered targets. Artifact version 0.2.0 → 0.2.1; decided calls younger than 14 days now publish alongside as immature_decided, never hidden.
• Known contaminant, published as registered: "we backed off" is attributed mechanically — our rank moved more than the field's. Model retraining nights move our whole board at once, and 50 of the current 65 retreats happened without the field moving away (in 20 of them the field actually moved toward us). The registered rule counts every one against us, so that is how we publish it. A re-baseline-aware retreat rule will be pre-registered as a dated v0.3 and applied forward only — rewriting the rule after seeing the score is exactly what pre-registration exists to prevent.
• Statuses recompute; nothing is hand-frozen: every build re-derives every call from the full archive, so a "caught up" call whose gap re-opens later re-opens with it and the headline moves both ways. Ledger copy that previously described those outcomes as "final" has been corrected.
• Conservative choices we are keeping: controls are drawn from players never flagged as of the latest build (an as-of-call-date pool would read ~1.40x on all calls instead of 1.36x), and controls are matched on rank band, role, and date but not development level (a same-level-only sensitivity reads ~1.65x matured instead of 1.31x — above target). Both alternatives would raise our self-grade, so neither ships as a post-hoc change; they live here as published sensitivities.

Stabilized read — added 2026-07-13 (the day after first publish)

The first day the headline went live it read one thing; the next build it read another, off a single day of new board data. That is the metric working as designed — it re-scores every open call against the newest field snapshot each night — but it means a single build is a noisy reading, not a track record. Trumpeting one good day would be cherry-picking; trumpeting one bad day would be its own kind of dishonesty. So from 2026-07-13 forward we also publish a trailing 14-build average of the exact same daily numbers, replayed from the committed archive so anyone can reproduce it, alongside the min and max so the day-to-day noise is visible rather than hidden. The pre-registered daily number is unchanged — this is a summary laid on top of it, not a rewrite of it. We report the average, we show the range, and we let the number be as modest and as noisy as it honestly is.

Our rankings depend on a bunch of judgment-call dials — how much to trust a hot three-week stretch with no track record behind it, how much a bad current season should outweigh a great last year, how much a prospect's odds of reaching the majors count versus his fantasy ceiling once he's there. None of these has a single right setting. So instead of handing you one ranking and asking you to trust it, we grabbed each dial, nudged it to a reasonable alternative, re-ran all 2,796 ranked prospects, and counted how many actually changed spots. It's us showing our work — which assumptions move the board a lot, and which barely matter. This is a real re-scoring of the live model, not a simulation; every figure below regenerates nightly from the public data pipeline.

The pattern that holds across every dial: the top 100 barely moves. Your blue-chip prospects are locked in no matter how we set things — almost all the churn is deep in the board, where prospects are nearly tied and any nudge just reshuffles a crowd.

Two things this analysis does not tell you: whether any of these alternate settings would have predicted real outcomes better — that's a separate, harder question we validate with held-out backtests before changing anything live — and it isn't recomputed on every page load, since a full board rebuild takes real compute. This is a dated study (run 2026-07-07 against that day's committed board) that we re-run and republish as the model changes.