How comparisons work

MatrixScore 2.0

An estimated comparison of actual starting-lineup production across league sizes and scoring rules. Wins, schedules, drafting decisions, lineup efficiency and waiver activity do not enter the headline score. Draft-position adjustments are zero.

What 100 means

100 represents the model’s estimate of ordinary starter production for your league’s rules, size and starting slots in that completed NFL week. It is learned from external redraft leagues, not your league’s average or the website’s participants. It is not an optimal-lineup target.

Estimated percentile: a weekly score of 100 + 15z is shown as the share of teams in equivalent formats it beats, Φ(z). A season average of n weeks spreads less than one week, but a team’s weeks are correlated, so the estimate uses z × √(n ÷ (1 + 0.35 × (n − 1))). It is a model estimate, not a rank among FantasyMatrix members. Public season rankings require at least 3 completed weeks; single weeks still rank as weekly views.

Each week: 100 + 15 × (actual starter points − estimated reference mean) ÷ estimated reference spread. Season scores average the same complete weeks. A score of 120 is 20 index points above the reference—not 20% better, a percentile or proven manager skill.

How the reference is built

We score recorded NFL production under your exact native rules and calculate the player supply available for the legal starting slots across the league. A frozen statistical model translates that production environment and league structure into an estimated ordinary mean and spread. The optimal allocation describes available production; it is not the 100 standard. No simulated drafting, lineup decisions or waivers are used.

The coefficients stay fixed; each completed week uses that week’s NFL results. Your roster, score and website rank never fit your reference. Native rules, model version and NFL source revisions must match. Missing or contradictory source facts leave a score unavailable.

Who is in the average?

Every team in a Sleeper league, or a publicly readable ESPN league, imported by a participating member enters public team comparisons anonymously, by team name only. Private ESPN leagues uploaded from your browser stay in your own dashboard. Co-owners do not count a team twice. Filters change the comparison group, never the underlying score. “Versus others” excludes the selected team. Only leagues with at least eight teams enter public rankings: a very small league leaves far more of the available NFL production per roster, so its scores do not compare fairly with ordinary leagues. A league is also held out of a week when fewer than three quarters of its teams scored that week. Smaller or quiet leagues keep their Matrix Scores on Dashboard and in their own league view. Public averages and ranks require at least five teams from three leagues. Percentiles describe the displayed sample, not all NFL fantasy players.

One ranking, additional context

MatrixScore 2.0 displays estimates to two decimal places. Equal rounded scores share competition ranks (1, 1, 3). League matchup views use authorized league results; unjoined members remain anonymous. Only individual teams enter leaderboards.

Advanced metrics

All-play asks how often your provider points would have beaten every opponent that week. Schedule luck is actual wins minus expected all-play wins for recorded matchups. Starting efficiency compares your actual starters with the best legal lineup from your roster for that week; missing bench data leaves it unavailable. Weekly variation measures consistency, not quality. Player consistency shows sample standard deviation across at least three observed starts, alongside the average and number of starts. Lower means steadier. Benched weeks are excluded; missing records are never filled with zero.

Availability and corrections

Missing stats, unknown player IDs and missing weekly settings are not zeros. Completed-week corrections produce new immutable results; old live results are hidden until recalculation succeeds. Only the current NFL season is available; prior-season browsing and imports are excluded.

Validation and limits

The frozen model used 51,690 reconciled team-weeks from 254 owner-connected groups. An evaluation of 7,268 rows from 220 other groups was inspected before this release; it is not a new untouched test. The reference inherits the selection and behavior of those external leagues.

Earlier strict uncertainty and subgroup ranking tests failed. The overall 95th-percentile parameter interval width was about 9.75 Matrix points; small-league comparisons were less stable. Those failures have not become passes. Scores display two decimal places for consistent comparisons and shared ranks; those digits do not establish statistical precision, and even several points may be uncertain. Sample league-size groups are not forced to average 100.

Initial scoring supports 4–16-team formats inside the model’s numerical range, with exact supported ESPN or Sleeper rules and complete legal starter inputs. Evidence is weaker for small and unusual formats. Leagues with 18–20 teams, unsupported event categories, incomplete NFL inputs or out-of-range structures remain unavailable. Yahoo and CBS MatrixScores are unavailable. Existing ESPN helper distribution limitations still apply.

Public rankings include every team in Sleeper leagues and publicly readable ESPN leagues imported by participating members, shown anonymously by team name. Leagues imported only by system-test accounts are excluded. The Benchmark is the average of other eligible teams; it is distinct from the model’s 100 standard.

Sources: the documented Sleeper API, with discovery seeded from publicly listed Megalabowl participants. This convenience sample is concentrated in 12-team leagues and is not representative of all fantasy players. Some managers participate in multiple leagues. Historical data is used internally for research; the product displays the current season only.

Numeric data attribution

NFL numeric data: nflverse contributors, licensed under CC BY 4.0. Numeric fields are mapped into our event dictionary; editorial text and images are excluded. Internal research datasets are separate from user results and never appear as imported teams.