How the board is built
A model trained on thirteen seasons of what actually happened, blended with the dynasty market, and graded on seasons it never saw. Here is what goes in, how we test it, and how often it is right.
The record
Most rankings never tell you how they did. Ours are graded the same way every time: rebuild the world as it looked in August of a past season, rank it, wait, and compare with what those players actually scored over the next five seasons. The model is never allowed to see the season it is graded on. Then we put the same test to the dynasty consensus.
Held-out seasons 2021–2024, graded against points above replacement over the following seasons (through 2025). Market = FantasyPros dynasty expert consensus nearest 15 August, plus dynasty draft data in 1QB. 5,587 pairwise disagreements on the 1QB board; 702 of them 20+ places apart. Ahead-of-market count uses rank correlation over the players the consensus ranks. Numbers are re-run every time the backtest runs and updated here.
Format by format
Rank correlation between a board and what actually happened: 1.0 is a perfect order, 0 is random. Every dynasty ranking you have ever read would score somewhere in the 0.6s on this test; the gaps below are the whole game.
| League format | Consensus | Our board | Seasons ahead | Right when we disagree |
|---|---|---|---|---|
| 1QB, full PPR | 0.658 | 0.678 | 4 of 4 | 63% |
| Superflex | 0.688 | 0.707 | 4 of 4 | 64% |
| 1QB, 3 WR | 0.674 | 0.690 | 4 of 4 | 59% |
| Half PPR | 0.654 | 0.674 | 4 of 4 | 63% |
| TE premium | 0.649 | 0.682 | 4 of 4 | 65% |
Of the players who went on to finish in the top 50, our 1QB board had 30 of them in its top 50 before the season; the consensus had 27. Rookies are the hardest call in dynasty, and there the model beats the simplest rival, draft order, in 10 of the 13 rookie classes since 2012 — on classes it had never seen.
What goes in
Around seventy facts about each player, all public, all refreshed automatically every week in season and every offseason as players enter and leave the league. Nothing is typed in by hand. The model decides how much each one matters by looking at which ones actually predicted the next five seasons — not by anyone's opinion, including ours.
Usage and role
Target share, snap share, share of the team’s air yards, red-zone targets and carries, touches per game this season and over three, and where he sits on the depth chart. Volume is the most durable signal in football.
Production
Points per game last season, his three-season average, his best season, the trend across them, and how many top-tier seasons he has already put up. A track record counts for more than a hot month.
Age and mileage
Age, years in the league, career touches, games missed over three seasons and the kind of injuries behind them. Every position ages on its own curve, and the model learned those curves from history rather than being told them.
Draft capital
Round and overall pick. For a rookie it is most of what we know; for a veteran it fades as real seasons replace it. The model learned how fast to let it fade.
Team situation
How good his offense is at passing and running, how often it throws, the quality of his quarterback, and his contract: years left and salary as a share of the cap, which is what tells you whether the team is committed to him.
The Dynasty Value Engine
Our own projection system’s read on every player — projected points, value over a replacement-level starter, age-adjusted career value — goes in as an input alongside the raw facts, so the model can use it where it helps and overrule it where history says it should.
One more thing goes in: where dynasty drafters are actually taking him in startup drafts. Not expert rankings — those are what we grade ourselves against, and they never touch the model. Draft position is different: it is thousands of people putting real picks on the line, and it turned out to be one of the most predictive facts we have.
How it is tested
This is the part that makes the numbers above mean something. It is also the part most rankings skip.
- Rebuild the past. For every season from 2012 to 2024, we rebuild the player table exactly as it stood before week one of that year: last season’s stats, that April’s draft, that summer’s depth charts and contracts. Nothing from later is allowed in.
- Rank it blind. The model is trained on the other seasons and produces a board for the one it is being tested on. It never sees that season’s results, and it never sees how those particular players turned out in any later year either — a live board cannot know that, so the test must not.
- Grade it on what happened. Each player’s value is what he actually scored above a replacement-level starter over the next five seasons, later seasons counting a little less — because that is what a dynasty roster actually collects. A season not played scores zero, so durability is graded, not assumed.
- Grade the market the same way. The expert consensus from the same August, on the same players, against the same outcomes. If we cannot beat it, the change does not ship.
Every candidate improvement goes through this before it touches the site, on its own, off against on. Most of them fail. That is what the strike-through list below is.
Why the market gets a vote
The board you see is 60% our model and 40% the dynasty market. Not because we lack conviction: the model alone beats the consensus too. But the market knows things no dataset does — a coach’s comment, a camp report, an injury nobody has filed yet — and history says a board that keeps 40% of that beats one that ignores it. We tested every split from 0% market to 100%. Sixty-forty won.
What the market never gets is a say in the model itself. Our read of a player is formed without it, then blended. That is why the two can disagree, and why the disagreements are where we make our calls.
What we tested and threw out
Being smart about rankings is mostly refusing to add things that sound like they should matter but do not. Each of these was built, added to the model, and run through the full test. The ones crossed out made the board no better or worse over thirteen seasons. The rest stayed.
Average separation, for instance, has almost no relationship with fantasy production five years out (correlation −0.02). Target share, which we already had, comes in at +0.43. Most tracking stats are either a weaker copy of volume or noise.
What this does not mean
Sixty-three percent is not ninety. Dynasty football is one of the noisiest things anyone tries to forecast: a torn ligament in week two rewrites five years. Our edge over the consensus is real, repeatable and, so far, present in every season and format we have tested. It is also a few points of correlation, not a crystal ball. Use the board the way we do: as the best available starting point for a decision, and as a map of where the market is most likely to be wrong.
The test is only as honest as its rules, so here are ours: the model never trains on the season it is graded on or on those players’ later results; the market is never an input to the model; nothing is hand-weighted; and every number on this page is recomputed from the same test each time it runs. When we found a flaw in an earlier version of the test in September 2026, we pulled the old numbers from this site and published the corrected ones. These are the corrected ones.
The model is retrained after each season and re-tested across every league format. The Dynasty Value Engine is the projection system built by Liam; the learned model and its backtest are the Digest’s. Questions about the method: ask us on X.