Models, equations, and assumptions.

Every franchise begins at 1500. A pregame rating difference becomes an expected result; the winner gains exactly what the loser gives up.

S is 1 for a win, .5 for a tie, and 0 for a loss. Ratings retain 75% of their distance from 1500 between seasons. Completed head-to-head games count; median games and byes do not. The validation harness compares K, margin treatment, offseason regression, and postseason inclusion with Brier score, log loss, and calibration before any production parameter changes.

Replacement is the dynasty roster fringe

The model counts QB, RB, WR, and TE ownership across Sleeper matchup rosters for 2022–25 and daily roster snapshots from 2026 onward. Starter, bench, taxi, and reserve memberships all count. Completeness is tested independently by position: a positional count must be at least 90% of that source’s largest valid count. Accepted matchup and daily evidence is then pooled.

For each completed season, players are ranked by Weeks 1–17 points. Replacement production is the median season total in a band around the cutoff, divided by 17. The band radius is the larger of three players or 5% of the rank.

One additive points-to-wins conversion

A legal all-replacement lineup is compared with every valid completed team score in Weeks 1–17, including playoff and consolation weeks. A team-week is excluded when its score or complete ten-player core lineup is unavailable. The numerator uses those exact same team-weeks, preventing regular-season and postseason horizons from being mixed. The average actual team has .500 all-play performance by construction; that gap creates the wins-above-replacement budget.

The denominator is identical for every position, so WAR is linear and additive. Replacement performance is zero; below replacement is negative; exceptional games are not artificially capped. Season and career WAR are exact weekly sums.

Known zero versus missing

Sleeper’s league score overrides provider data for rostered players. A submitted starter with zero receives negative FPAR. A nonzero provider performance is known. A provider zero for a non-starter, or no provider row, is unknown unless independent participation evidence exists; unknown bench weeks are excluded from hindsight optimization rather than inserted as zero. Historical active-zero, bye, and inactive distinctions remain limited by provider coverage.

V42 uses separate models for incoming rookies and players with NFL experience. Both routes estimate the player’s WAR in each of the next three seasons on the same 201.15 fantasy-points-per-win scale used for historical WAR. The models were trained only on information that would have been available before the season being predicted.

Missing prior seasons remain explicit inputs rather than disqualifying a player, which lets the model handle second- and third-year players. Experience is the number of completed NFL seasons. A rookie enters the primary route only when the draft and college records produce one unambiguous player match; otherwise the draft-only fallback is used.

Only Year 1 is prorated as the current season progresses; Year 2 and Year 3 remain full-season forecasts. Three-year WAR is projected production, so it is not reduced by a time discount. FantasyCalc and FantasyPros are displayed elsewhere as independent market and consensus views, but neither is an input to projected WAR. This separation prevents the site from merely restating market price as production.

Performance is measured by repeatedly training on earlier seasons and predicting a later season that the model has not seen. This expanding-window design is stricter than randomly mixing old and new player-seasons, and every feature is frozen at the point when the forecast would have been made.

Selected model results

“Next 3 seasons” means the sum of the next three annual WAR outcomes, so its error is naturally larger than one-season error. The rookie model was selected for its balance of error, ordering, and top-player identification rather than for winning every individual metric; the draft-only fallback had slightly better one-year MAE but weaker ordering and top-player results. In paired tests against the same incumbent efficiency model without true-experience features, the selected incumbent model reduced one-year MAE by 0.0017 and three-year MAE by 0.0114. The respective 95% bootstrap intervals were −0.0024 to −0.0010 and −0.0138 to −0.0086. These are modest improvements, not a claim of precise foresight.

WAR Inventory sums positive projected WAR held on the roster. It measures owned production inventory, not a fully realizable lineup. Starter WAR solves the highest-WAR legal QB/RB/WR/TE/FLEX/SUPER_FLEX lineup, fills every required slot, never duplicates a player, and permits a negative starter when required.

Weekly ROS optimization is used when weekly coverage exists. Otherwise the page visibly uses the uniform-availability approximation. Dynasty mode solves ROS, Year 2, and Year 3 separately so the lineup can change by period.

This is realized starter production and intentionally hindsight-based. Historical Roster Opportunity is the best legal lineup supported by known weekly outcomes. Unknown bench performances cannot enter as zero.

If the submitted lineup is malformed, incomplete, or contains missing starter evidence, opportunity is set to actual Manager WAR, lineup loss is zero, and the week receives a data-quality flag. This prevents incomplete data from manufacturing hindsight opportunity.

Historical decisions stay open and accumulate through the current snapshot. They measure outcomes, not what a manager reasonably knew on the transaction date.

The production benchmark uses nearby outcomes from the same startup or rookie draft, omits the target from its local neighborhood, and applies a shared non-increasing isotonic curve. A strict target-specific leave-one-out curve is reported as sensitivity because strict cross-fitting and one globally monotonic production curve are not generally identical.

A resolved traded pick contributes its same-draft benchmark to Trade Surplus and its selected player’s difference from that benchmark to Draft Surplus. An unresolved pick has zero realized WAR, remains visibly open, and is not described as having zero expected value. Historical transaction weeks currently use Sleeper’s leg boundary unless a reliable timestamp-and-kickoff boundary is available; each ledger row exposes the method and confidence.

The calculator shows the canonical projected WAR, FantasyCalc market price, and roster-specific Starter WAR change. Its current-state engine preserves 20 active, 5 taxi, and 3 reserve slots. Taxi and reserve players remain owned but are not directly startable. Incoming players default to active.

For uneven deals, plausible cut or waiver-add combinations are evaluated under the roster limit. The primary objective is legal Starter WAR; WAR Inventory breaks ties. Incoming picks do not consume a current spot but receive a future roster-pressure warning.

Future picks

FantasyCalc pick prices map to similarly priced young players using the FantasyCalc-independent three-year comparable. Mature league rookie outcomes are pooled with the market estimate when available. Raw player and pick production is unfloored. Non-increasing isotonic curves enforce pick order and permit ties; no artificial per-slot decrement is added.

Aggregate pick WAR is allocated evenly across the drafted player’s first three seasons as a disclosed temporary allocation. A current-year pick has all three seasons inside the window, next year has two, two years away has one, and later picks have zero inside today’s three-year window. Until the position-probability model is complete, lineup impact is labeled Best-Case Positional Fit: an upper bound that tests QB, RB, WR, and TE and keeps the best legal result.

Win-now requires positive tilt and at least 100 redraft points; rebuilding requires positive future tilt. Qualification happens first, then each view keeps the top 20 by raw tilt. User sorting reorders that fixed candidate pool.

Each regular-season team score is compared with every other club that week: 11 comparisons per club-week in a 12-team league. All-play is a standings column and does not enter Manager WAR.

Historical trade cards show WAR Gap = |A − B| for two sides and Total WAR Involved = |A| + |B|. Total WAR Involved is gross absolute career-to-date outcome, not net league value.