Ball Ranks does fantasy football now. Here's the model — and the backtests, including the ones we lost.

Ball Ranks started as a fantasy basketball board with an unusual rule: no ranking change ships until it has been replayed against completed seasons, and the experiments that failed stay on the public record next to the ones that worked.

Today the same rule comes to fantasy football. The 2026 football draft board is live, with a page for every NFL player, and this post is the receipt drawer: what the models are, how they scored, and where they lost.

Football is a different game, so the math is different

Our basketball board ranks players across nine scoring categories. Football leagues don't work that way — almost everyone plays points leagues, where your team's week is one number. So the football board doesn't borrow the basketball machinery. It does three things instead:

1. Project season fantasy points for every player, under your scoring format (standard, half PPR, or full PPR — the board switches between all three).
2. Convert points into value over a replacement-level starter at the same position. A quarterback's 300 points and a tight end's 220 are not the same asset, because the 12th-best quarterback is much closer to the top than the 12th-best tight end. The "VAL" column is the number that makes cross-position comparisons honest.
3. Show the room's opinion next to ours. Every player carries a live average draft position from thousands of real drafts this week. Where the model and the room disagree is exactly where your draft is won or lost.

The two models

Model Zero is the deliberately simple baseline: last season's actual fantasy points, blended with consensus ADP. It is honest about what it is — a sanity-checked starting point, not a crystal ball.

Model One goes further, using ideas that already earned their keep in our basketball research — each one re-proven on fifteen seasons of NFL history before it shipped:

• Three seasons of per-game production, weighted toward the most recent — one hot or injury-ruined season doesn't own a player's projection.
• Games played leaning on actual recent availability — durability matters, but last year's fluke injury isn't destiny.
• Age adjustments at position-specific cliffs — running backs decline earlier and harder than quarterbacks. Our first football research cycle found the launch cliffs were too gentle; the backtests wanted aging priced harsher, exactly as they did in basketball.
• A new-team haircut — a veteran changing teams gets trimmed 10% per game. Movers were over-projected in every one of the 38 replays we checked. Basketball found the same thing a year earlier; football agrees.
• A backfield that adds up — a team's carries are a roughly fixed pot, so work its current running backs did not personally produce last season has to land on the backs who are there now. Vacated carries lift the backs who remain; a big arrival, including a high-drafted rookie, costs the incumbents. We tested this on receivers and quarterbacks too and it did nothing — pass volume is elastic in a way carries are not.
• A restrained trained-model blend — 75% of the projection is the formula above, 25% is a ridge regression trained on every NFL season since 2008, forward-only so it never sees its own answers. Basketball ships the identical recipe.
• Rookies inherit expectations from where the room drafts them, lightly discounted — because rookies have no NFL stat line, and pretending otherwise would just be noise.

The backtests — all thirty-eight of them

When this board launched we had replayed three seasons. That was never going to be enough — our basketball model is graded on fifteen — so we went and got the rest: every draft year from 2011 through 2025 in standard and PPR, and 2018 through 2025 in half PPR (the public ADP archive starts later there). Every replay uses only information available before that season's opening kickoff: prior-year stats plus that summer's archived ADP, with three of the fifteen seasons locked away as untouched holdout years while the model was being tuned. We scored the top 150 of each board against what actually happened, measuring the average size of a points miss — lower is better.

The honest surprise: the fifteen-season view overturned parts of the three-season story twice. First, the launch claim that Model One beat the market's ordering in eight of nine replays collapsed to roughly half on thirty-eight replays — three lucky seasons had flattered the model — and we recalibrated how much the board listens to the room. Then a second audit caught the board over-ranking quarterbacks and kickers, because it ordered by raw projected points while drafts are played in value over replacement; we moved the board to draft currency and regraded everything in it. Today's Model One board places players closer to their actual value-over-replacement finish than a pure-ADP board in 32 of 38 shared-pool replays, including seasons the model was never tuned on.

On points accuracy, Model One now beats Model Zero in all 38 season-format replays — a 15% aggregate reduction in the average miss — with every one of those wins earned through the same hypothesis-test-ship discipline as basketball: each change swept on twelve search seasons, confirmed on three untouched holdout seasons, and the experiments that failed (a per-position ridge, a rookie draft-capital boost, pooling a team's entire output) rejected and published in the research ledger alongside the winners.

We publish the losses for the same reason we publish them on the basketball side: a model whose failures are hidden is a model you can't trust. The full numbers live in the repository's backtest artifacts.

What this board is not (yet)

Honesty about the ceiling: the models don't yet watch film, parse coaching changes, or model target shares play-by-play. But football's research program is no longer "just beginning" — it now runs on the same fifteen-season harness, holdout discipline, and public hypothesis ledger as basketball, and its first research cycle already shipped three model versions and rejected four ideas that looked good until the holdout seasons said otherwise. What you get today is a disciplined, backtested, replacement-aware board with the room's live opinion attached — and a paper trail for every claim in this post.

In-season is where the roadmap points next: weekly rest-of-season ranks and waiver tools that stay useful long after draft night.

Where the data comes from

NFL statistics come from the open nflverse data project. ADP comes from FantasyFootballCalculator's public API of real drafts. Depth charts are the league's published charts, not a stat-derived guess. Ball Ranks is independent and unaffiliated with the NFL.

Draft with receipts: open the 2026 board.