Methods

Glossary

Every term, model output, or metric that appears across the platform, in plain language — the same definitions behind the ⓘ hover hints on every page.

Start here

71 terms

P(elite): Legacy research output for the four-tier model's top career outcome. Public tier probabilities are currently withheld because the served prediction pipeline is not bound to a clean, maturity-safe evaluation artifact.
P(starter): Legacy research output for the four-tier model's starter outcome. It is not a publication-ready probability; maturity-safe validation currently has too few folds.
P(role): Legacy research output for the four-tier model's role-player outcome. It remains withheld until the exact served model and its maturity-safe evaluation are cryptographically bound.
Bottom-tier probabilityP(bust): Legacy API name for the four-tier career model's Bottom-tier probability. The tiers are cut by career value within position, so a guard is judged against guards rather than receivers. This is not the separate Bust% model shown on risk surfaces.
Career Score: A historical outcome index combining participation, durability, awards, and recorded contract outcomes. It can describe completed careers; it is not current proof that the prospect model predicts future careers. Component-level source and permission lineage is under audit.
Predicted tier: The largest of four legacy research probabilities. Public tier output is withheld until a clean, maturity-safe model artifact and its exact evaluation are promoted together.
Tier: A four-level research career-outcome label. The current tier predictor is not publication-ready and should not be read as a validated forecast.
NFL grade: Synthetic NFL-grade proxy. A calibrated probability (0-100%) that the player's NFL seasons looked like Pro Bowl / All-Pro play, built from public nflverse stats. Two numbers: recent-3-year weighted / career peak.
NFL grade (recent3): A weighted average of the player's last three NFL seasons, with the most recent season counting three times as much as the oldest (3x / 2x / 1x weighting). It answers the question a single-season grade can't: is this player currently good, or was he good once? A high peak with a fading recent-3 is the signature of a career in decline; the reverse is a player still climbing.
NFL grade (peak): The player's best single NFL season by our grade — the ceiling he actually reached, not the ceiling scouts projected. Useful next to the recent-3 grade: a big gap between peak and recent form usually means injury, age, or a role change; a small gap means what you saw is what he is.
NFL grade proxy (single season)pff_proxy_score: The model's probability that a given NFL season was played at Pro Bowl or All-Pro level, calibrated against actual award outcomes rather than anyone's subjective grade. We don't watch film for this number — it's built from playing time, production, and market signals, then checked against who actually made those teams. That calibration is why we call it a proxy and not a grade.
NFL grade proxy (recent three years)pff_proxy_recent3: The recent-form version of our NFL grade proxy: the last three seasons weighted 3x / 2x / 1x with the newest counting most. It captures whether a player is currently performing at a high level rather than coasting on an old reputation — the difference between a name and a player.
NFL grade proxy (peak season)pff_proxy_max: The best single-season score of our NFL grade proxy across the player's career — the year everything clicked. Peak season is the honest measure of ceiling: it actually happened, unlike a draft-day projection. Compare it with the career average to see how repeatable that peak was.
NFL grade proxy (career average)pff_proxy_mean: The career-long average of our NFL grade proxy across every season the player has played. Averages flatter compilers and punish late bloomers, so read it beside the peak and recent-3 numbers: a modest mean with a high recent-3 is a player who figured it out; the reverse is one the league figured out.
OVR AUC: One-vs-rest AUC: the probability a model ranks a randomly selected member of one outcome above a randomly selected non-member. 0.50 is chance and 1.0 is perfect. A value is publishable only when it comes from the exact served artifact under the stated temporal protocol.
Macro OVR AUC: The four one-vs-rest AUC values averaged with equal weight. Draftanomics currently withholds a public tier-model value because only one five-season-available fold can be evaluated.
Bust AUC: How well the bust model separates eventual busts from non-busts, scored on draft classes it never saw during training — so the number can't be flattered by memorization. 0.5 = coin flip; higher = sharper.
Brier: Grades the model's probability calls against what actually happened (technically, the mean squared error of its predictions). Lower = the odds it quotes track reality more closely.
Confidence: How certain the model is about its own prediction — a measure of profile familiarity, not of the player's quality. Low confidence means this prospect's combination of production, athleticism, and pedigree is unusual in the training data, so the grade rests on thinner precedent. A weird profile with low confidence is exactly where scouts and models most often disagree — and where either one can be right.
Draftanomics Pick: Where the model believes this prospect should be drafted, given everything it can see: college production, athletic testing, age, program context, and where the public boards have him. It's a value statement, not a prediction of draft night — teams reach and teams wait. When our pick and the consensus split hard, one of us is wrong, and that disagreement is the most useful thing on the page.
Predicted pick: The model's experimental projected overall pick number. Treat the displayed range as a research estimate unless the page links it to a promoted, out-of-fold evaluation artifact.
Actual pick: Where the player was really drafted — the outcome, straight from the league's own card. Every backtest and re-grade on this site is scored against this number and the career that followed it, never against anyone's opinion.
Consensus rank: The prospect's aggregate rank across tracked public boards. It is a market-opinion feature, not ground truth. Historical snapshots remain research-only where immutable pre-draft capture time and reuse permission are not verified.
Bust probabilityBust Probability: The separate bust model's estimate that the player's rookie-deal window meets Draftanomics's bust definition: bottom-end career value for his position, contribution below an absolute floor, or — for a top-40 pick — return below the expectation attached to that draft cost. The displayed number is pulled toward the historical position rate when held-out evidence is weak and withheld where the model cannot support a claim. It is not the four-tier model's Bottom-tier odds.
Scout grade: The average grade, out of 10, across the team-internal scout reports filed for this prospect on this platform. It's human judgment, deliberately kept separate from the model — the point of collecting both is to see where they split.
Scout Δ: The model's grade minus the scouts' grade. Positive means the model is higher on the player than the humans; negative means the scouts see something the data doesn't — or think they do. The biggest deltas in either direction are the prospects worth arguing about, because someone is provably wrong.
Scouts: The average team-internal scout grade for this prospect, out of 10, alongside how many scouts filed a report. One glowing report is an anecdote; five independent grades that agree are a signal. Weight the count as much as the average.
Consensus strengthf_cons: How high the public draft boards rank this prospect, normalized against his draft class. This is the market's opinion distilled to one input — and deliberately just one input among many, so a player the boards love but the data doesn't gets a real argument instead of an echo.
Athletic compositef_athl: Athletic composite (RAS or position-specific equivalent), scored against the prospect's position group — 0 is average for the position, positive is better than average (a z-score).
Production compositef_prod: The college production composite: yards, touchdowns, and opportunity adjusted by position and context. It is one model input; no current public evidence supports calling it the strongest predictor of NFL success.
Program strengthf_school: Program strength — the quality of the college program the prospect played in, measured by SP+ rating. It works both ways: production at a powerhouse came against practice-squad depth charts full of future pros, while a small-school stat line faced lighter competition. The model prices the context so neither gets a free ride.
Age at draftf_age: Age on draft day, normalized. Younger is usually better — a 21-year-old and a 23-year-old with identical production are not equal prospects, because the younger one did it with less physical development and has more runway left. The penalty scales by position; running backs age on a different clock than offensive linemen.
Recruiting pedigreef_recruit: High-school recruiting pedigree — the star rating and composite ranking the prospect carried into college. It's a prior, not a verdict: five-stars bust and walk-ons make Pro Bowls. But at the population level, recruiting rankings retain real signal years later, so the model keeps a thumb on the scale.
Per-play impactf_ppa: Per-play impact from College Football Data's PPA metric — how much each snap the player was involved in moved his team's scoring expectation. Volume stats reward opportunity; per-play efficiency rewards the player. The distinction matters most for prospects with modest totals in crowded offenses.
Senior Bowl invitef_senior_bowl: Whether the prospect was invited to the Senior Bowl — the all-star game NFL front offices treat as a job interview. The invitation itself is the signal: it means real scouts wanted a longer look. A binary flag, but a meaningful one for Day 2 and Day 3 prospects.
Strength of schedulef_sos: Strength of schedule across the prospect's college career — who the production actually came against. A thousand-yard season built on cupcakes and one built on ranked defenses are different resumes, and the model refuses to price them the same.
Breakout agef_breakout_age: The age at which the player first produced at a high level. Early breakouts — dominating as a true freshman or sophomore — predict NFL success far better than a lone senior-year explosion, which is often a fifth-year player beating up on younger competition. One of the strongest signals in the model for receivers especially.
Production trajectoryf_production_trajectory: The shape of the player's college career: still improving in his final season, or peaked early and plateaued? A rising trajectory suggests untapped development; a player whose best football came two seasons ago is asking teams to bet on a version of him that stopped showing up.
Workload sharef_workload_share: How much of his team's production the player personally carried — target share for receivers, carry share for backs, and the equivalent for other positions. A player who WAS the offense is a different bet than one who feasted on the attention defenses paid to teammates.
Athleticism × productionf_ras_x_production: An interaction term asking how rare the combination is: this level of athleticism AND this level of production in one player. Great testers who never produced and great producers who can't move are both common; players who are genuinely both are the shortlist this feature exists to find.
Schedule-weighted productionf_sos_weighted_production: College production re-weighted by strength of schedule — the stat line priced by who it came against. This is the feature that separates real producers from stat-padders: numbers against ranked defenses count for more, numbers against overmatched opponents count for less.
Age-adjusted productionf_age_adj_production: Production adjusted for how old the player was when he did it. A 19-year-old and a 22-year-old posting the same line are not the same prospect — the teenager did it against older, stronger competition relative to himself. Younger production is worth more, and the model prices it that way.
Conference strengthf_conference_strength: A conference-level quality adjustment — Power 5 production and Group of 5 production are priced differently, because the weekly competition is different. Cruder than the per-game strength-of-schedule features, but it catches the baseline gap those can miss.
Position purityf_position_role_purity: How consistently the prospect actually played his listed position. A 'cornerback' who spent half his snaps at safety is harder to project at either spot, and position-switchers carry extra translation risk the model prices in. Purity isn't virtue — it's just measurability.
Board agreementf_board_consistency: How much the public boards agree about this prospect. Tight agreement means a clean, legible profile; wide disagreement means the analysts are seeing different players in the same tape. Historically, that disagreement itself carries signal — scout-divided prospects bust more often than their average rank suggests.
Coverage: How many public draft boards rank this prospect at all. Coverage is the confidence footing under every consensus number on the page: a consensus rank built from twenty boards means something; one built from two is barely a rumor. We show the count so you can weight the consensus accordingly.
Board coverageconsensus_coverage_count: How many public draft boards rank this prospect at all. Coverage is the confidence footing under every consensus number on the page: a consensus rank built from twenty boards means something; one built from two is barely a rumor. We show the count so you can weight the consensus accordingly.
Predicted AAV: An experimental estimate of second-deal dollars per year. The current serving run is not yet bound to a promoted out-of-fold evaluation artifact, so no accuracy claim is published for it.
Actual AAV: A recorded historical contract's average annual value where a qualifying deal and source row exist. Contract-source completeness and first-versus-largest-deal semantics are under audit, so absence must not be treated as proof of no deal.
AAV: Average annual value — a contract's total worth divided by its years, in millions per season. It's the standard unit for comparing contracts of different lengths, and the reason we project second-deal AAV specifically: the second contract is the first one a player earns on performance rather than draft slot.
Predicted contract valuepredicted_aav_M: The model's experimental projected average annual value for this player's second NFL contract, in millions per year, conditional on earning a second deal. No out-of-sample accuracy or uncertainty claim is published until this serving run is tied to a promoted evaluation artifact.
Board: The default research board: model features blended with a public-consensus signal. Its prior predictive-accuracy claim is withdrawn while maturity-safe validation and historical consensus provenance are rebuilt.
Sober: Same blend as the default Board, but caps how much public-board hype can move a prospect. Anthony Richardson 2023 falls from #13 → #41 here.
Pure model: What the model says in a vacuum: college production, athletic testing, age, and program strength, with every public board stripped out. It disagrees with the default board on purpose — this view exists to show you exactly where the data and the draft industry split, which is where the interesting arguments live.
Consensus-aware: The former name of the default board, renamed simply 'Board' in May 2026. Model signal blended with the weighted analyst consensus — kept in the glossary because older links and saved views still use the name.
Measurables only: The former name of the 'Pure model' view, renamed in May 2026. The model with no analyst-board input — production, testing, age, and program strength only. Kept here because older links still use the name.
Risk scoreRisk Score: Composite 0–100. Combines bust probability, injury flags, character flags, and scheme-fit risk. >=50 = material risk to flag in meetings. >=20 = worth noting. <20 = clean profile.
Feature contribution: Pre-rank score added to the prospect by this feature, in standard-deviation units. +1.0 ≈ pushes the player ~25 spots earlier in the draft order; −1.0 ≈ ~25 spots later. Magnitude over sign — small contributions ride along; only the top 3–5 are decision-relevant.
Raw scoreraw_score: Draftanomics's overall prospect score, on the same 0-to-20 scale as the Rookie score: 0 is the weakest profile we've ever graded, ~20 is an elite prospect's expected NFL production in years 1-3. Appears on /class-strength and /movers.
Rookie scoredraftiq_rookie_score: Predicted NFL years-1-3 production grade on a 0-20 scale. Built mostly from projected playing time — six parts snap share, two parts share of games started — plus a charting-grade component for positions the box score can't see, or a box-score component for positions it can.
Elite probabilityp_elite: Legacy four-tier top-outcome research score. It is currently withheld as a validated probability.
STEAL: A hindsight verdict on a real draft pick: the model had this player going ten or more spots later than where he was actually taken, and the career that followed says the drafting team was right to jump. Steals are scored against outcomes, not projections — a team only banks the credit when the player delivers.
FAIR: A hindsight verdict: the actual pick landed within ten spots of where the model's board had the player. No edge claimed in either direction — the team paid roughly what the player was worth. Most picks, most years, land here, which is exactly what an efficient draft market should look like.
REACH: A hindsight verdict: the player was drafted ten or more spots earlier than the model's board had him. A reach isn't automatically a mistake — teams hold private information the model can't see — but at the population level, reaches underperform their draft slot, and the re-grade pages keep the record.
Capital: Total draft-capital remaining or spent, expressed as a percent of the team's pre-draft capital. Capital is the Jimmy Johnson chart value of all picks owned. 100% = full slate, 0% = exhausted.
Score: The position-normalized version of the model's raw grade, used on the movers page: each prospect measured against the median prospect at his own position in this class. Raw grades aren't comparable across positions — quarterbacks and guards live on different scales — so this version answers the only fair question: how good is he for a player at his spot?
Biggest fall: Distance between the player's pre-draft consensus rank and the pick where they ended up — when they slid by 10+ spots. Surfaces during live draft as picks come off the board.
Consensus dispersion: How much the analyst boards disagree about this prospect, measured as the standard deviation of his rank across every board we track. Low dispersion is a clean, agreed-upon profile. High dispersion means the boards are split — and historically, that split itself is information: scout-divided prospects carry more bust risk than their average rank implies, because one of those boards is badly wrong.
Snap-weighted charting grade: Player's career charting grade weighted by snaps played each season. Punishes high-grade-low-snap one-game wonders relative to a player who graded out at the same level on triple the volume.
RAS: Relative Athletic Score (Kent Lee Platte). A 0-10 score that grades an NFL Combine performance against every other prospect at that position dating back to 1987. RAS sub-scores (size / speed / explosion / agility) decompose it.
Scout synthesis: Draftanomics's machine-composed read of the public scouting reports and consensus-board rows on file for a prospect. It cites sources, surfaces analyst disagreement, and produces a verdict, case-for, and case-against. It refreshes when new sources come in.