How it works

Every projection starts with the team, not the player. The model estimates how many plays a team runs, how often it throws, and how many points it scores. Only then are those totals divided among players, so a receiver’s targets are always a share of a real number rather than a guess made in isolation. That is why a good player on a slow, run-heavy offense projects lower than his talent alone suggests.

Shares come from usage, weighted toward recent seasons and regressed toward what players in that role normally do. A single big year moves a projection less than the year itself did, because most single big years are partly luck. Draft capital, depth-chart position and beat reporting about who is taking first-team reps all adjust the share.

Efficiency is applied last — yards per target, catch rate, touchdown rate — and is regressed harder than volume, because it is far less stable year to year. Volume is the signal; efficiency is mostly noise wearing a signal’s clothes.

The cross-position order is value over replacement, not points. VOR is games multiplied by how much a player beats a freely available replacement per game. It exists because 280 points from a quarterback and 280 from a running back are not worth the same: you can stream a quarterback and you cannot stream a back. On raw points, quarterbacks would occupy most of the first round, which is not how anybody drafts.

Your league changes the ranking, not just the labels. League size and starting lineup set replacement level, and replacement level is what the board is ranked on. A 14-team league drains every position far deeper than an 8-team one — running back replacement runs 175 points at 8 teams against 98 at 14 — so a shallow league leaves startable backs on the wire and beating replacement there is worth much less. Change the size or the starters and the order changes with it.

Tiers are cut where the projections genuinely separate, not into equal buckets. A tier is a set of players you should be close to indifferent between; the gap to the next tier is the thing worth acting on. They grow as they go down, because by the middle rounds you really are choosing between near-identical players.

It barely forecasts injuries. Games played this year explain only about 6–16% of games played next year. The model projects a spread anyway because ranking needs one, but a difference of one or two games between two players is noise, not a finding. Where a player is currently hurt or suspended, that is real and is priced; where he simply missed time before, treat it lightly.

It cannot price a handcuff. A backup running back is worth something because he inherits a job if the starter goes down. That is option value, and a point projection has no way to express it — so the board will always look too low on players whose case is “if”. The market prices it and we do not. That is a known gap rather than a disagreement.

Some rankings are hand-set. Where the model and a human read of a situation genuinely differ, a player’s projection is overridden by hand, with the reason recorded. Roughly a third of the top of the board carries one. They are views, not corrections, and they are the first place to look if a ranking seems wrong to you. Every one carries its reasoning, written out in full on the player’s own page.

The market is good. Measured against realised finishes on held-out seasons, average draft position beats this model at every position. That is not an argument for ignoring the board — its job is to disagree in specific places for stated reasons, and the biggest disagreements are where the thinking is. It is an argument against assuming a gap means the model has found something.