Methodology
A number you cannot interrogate is a number you should not pay for - so every decision behind these six models is published below, in enough detail that a skeptical reader can find the weak point and argue with it. The constants that implement those decisions are just as real; they open with an account on any plan, because keeping them from customers was never the point - keeping them from anonymous copying is.
Six models drive every projection on this site: Elo team ratings, win-probability blending, Monte Carlo game scoring, touchdown-rate regression, season-projection shrinkage, and draft-availability simulation. Each is explained below in plain language; the fitted constants behind each one open with an account, on any plan.
Ratings
You would expect a team rating to simply reset each year and reflect that season's record so far. Ours doesn't: it is Elo, carried forward from every regular- and post-season game since 2002, adjusted for home field - because playing on the road is genuinely harder - and weighted by how badly a team won or lost rather than merely whether it did. The challenge to that carry-forward is built in too: at each season boundary the rating regresses partway back toward the mean, so one hot year cannot calcify into an edge the following year has to disprove.
Win probability
You would think a proprietary rating should stand on its own, especially early in the season when it matters most. Early is exactly when we trust it least: in Week 1, the market's vig-free probability takes the larger share of the blend, because a roster that turned over across the offseason holds information a rating built on last year's games cannot possibly contain. Elo knows who won last year's games, not who changed teams since - and that gap is exactly why the market's read on the current roster is allowed to matter more than ours.
Scores
You would assume a game-score model earns its keep by out-guessing the market's total - calling a shootout the market missed, or a defensive slog nobody priced in. It doesn't try. The total is anchored to the market's number entirely: every simulated game draws its margin around a blended expectation, but the level of scoring is taken as given, not re-forecast. What the simulation actually contributes is the split - which team gets which share of that total - because pricing the total itself is treated as the market's job here, not something we claim to do better.
Touchdowns
The obvious way to project a player's touchdowns is to start from how many he scored last season. We start from where he touches the ball instead. A team's expected touchdowns come from a relationship fitted on the prior season's team-games and applied to what the market currently expects that team to score - not to last year's total, where one player ran hot and another ran cold for reasons that rarely repeat. Player rates are then built from expected touchdowns given field position, scaled for the game's environment and for whether the player is even expected to suit up, and converted to anytime and 2+ probabilities through the Poisson relation rather than backed into from a raw scoring count.
Season projections
The naive projection takes last year's per-game numbers and rolls them forward untouched. Ours starts by asking how much of that season was actually earned: opportunity is shrunk toward a role-appropriate prior according to that statistic's own year-over-year stability, and efficiency - where a good season is often a good season and also somewhat lucky - is shrunk considerably harder than volume, which repeats far more reliably. Games played comes from positional base rates rather than a healthy-season assumption, overridden by hand only where an injury is actually documented. Even the spread around the projection is calibrated to strip out the luck that inflated or sank last season's finish.
Draft availability
You would expect a 'will he be there' tool built by a company with its own rankings to quietly lean on those rankings. It doesn't touch them. Draft availability comes from simulating 12-team snake drafts in which every other drafter's private valuation is consensus ADP plus noise, with that noise widening later in the draft, where real opinion is genuinely less converged. The result is driven entirely by ADP rather than by our own board, so the answer describes the room you are drafting into, not whether we personally rate the player.
You would expect a model's weak points to be exactly what a vendor keeps quiet. Not here: the precision notes and stated limits - where each source rounds, and where this model is actually weak - are published in full. An account opens it, at every tier and the trial included - there is no paid tier for this, and nothing here is withheld from customers. It is withheld from anonymous copying, which is a different thing entirely.
New accounts get full Active access free for 30 days - no card. Choose a plan before it ends and access continues without interruption; do nothing and access locks, not charges and not deletes - purchasing any time afterward turns it back on.
SEE THE CONSTANTS →Team Elo
Every active franchise's rating going into the season.
You would expect the actual ratings to be the one thing this page keeps back. They are - because a rating is model output, not method, and there's nothing in output to argue with, only a number. All 32 team ratings carried into the season are here in full. An account opens it, at every tier and the trial included - there is no paid tier for this, and nothing here is withheld from customers. It is withheld from anonymous copying, which is a different thing entirely.
New accounts get full Active access free for 30 days - no card. Choose a plan before it ends and access continues without interruption; do nothing and access locks, not charges and not deletes - purchasing any time afterward turns it back on.
SEE THE CONSTANTS →Data from nflverse (https://nflverse.com), CC-BY-4.0
Data from the Hockey Databank project
What this is not
Forecast and research output. Market lines appear as model inputs and as a calibration check. No wagering edge, stake, or recommendation is expressed or implied. Where model and market diverge, the honest prior - stated throughout and applied in the blend - is that the market is more often right.