How this works, and where it is weak

FFT previously shipped a projected-points column and a rank column, and built its headline feature — comparing our rank against draft-market ADP to surface value — on top of them. That premise was tested directly against the shipped dataset of 335 players. It failed.

And the rank was not independent either: it tracked the draft market at 93 to 97 percent concordance, because it was in substance the row order of the ADP table the dataset had been parsed from. The consensus engine found the same thing from the other direction — across the 216 players every ADP source ranked, 23,191 pairs, the model ordered the pool exactly as the market priced it, with zero inversions.

This model exists to fix that at the root: to build a projection out of measured on-field inputs, so that when it disagrees with ADP the disagreement carries actual information.

The model must not consume ADP, or anything derived from it, anywhere. The forbidden inputs, by exact field name:

The prohibition is total, and it covers every stage that could smuggle the market back in. Not as a feature. Not as a tie-break, because ordering two equal projections by ADP re-imports the market ranking through the back door. Not as a filter, because projecting only “the top 200 by ADP” makes the player pool itself a market artifact. Not as a calibration target, because tuning the output until it looks right against ADP is precisely the failure being fixed. And not as a blend — a 70/30 model-market blend is 30 percent market, and the result cannot honestly be called independent.

Partial independence is not a weaker form of independence. It is a projection that agrees with the market by construction and therefore cannot tell you when the market is wrong. The engine module has no value imports at all; its only import is a type-only one, erased at compile time.

Every input is a measurement of something that physically happened on a football field, or a currently-true roster fact.

• nflverse season statistics

  Games, pass attempts and completions, passing yards and touchdowns, interceptions, carries, rushing yards and touchdowns, targets, receptions, receiving yards and touchdowns. This is the volume spine. Fantasy points appear only as backtest ground truth, never as an input.

• nflverse snap counts

  Offensive snap share. It does not change a projected rate — it changes how far that rate is regressed, so a player at 90 percent of snaps is treated as a slightly larger sample than one at 25 percent.

• nflverse player bio

  Birth date for age at the projected season, and draft round and pick as the market-independent proxy for expected rookie opportunity. Draft position is a team's own valuation, made before any fantasy ADP existed.

• nflverse weekly rosters

  A second, earlier source for the same draft capital. players.csv fills a class in late — for 2026 it carried a pick for 1 rookie of 233 — while the weekly roster file carries an overall pick for 80 of them, and the round is inferred from the measured round boundaries of the 2015-2026 classes. A rookie no source can place is labelled unknown and flagged as a positional floor, not quietly called undrafted.

• Sleeper live overlay

  Current team, depth-chart position and order, injury designation. Last season's stats describe last season's situation; this is how the model learns the situation changed.

• Red-zone opportunity

  Red-zone targets and carries, derived by FFT from openly-licensed play-by-play. Kept because it is correctly joined and gives a claim its context — not because it was shown to help. The measurement is below.

The method itself is the standard defensible one: fantasy production is mostly opportunity, and volume is far more stable year over year than efficiency. So the model projects availability, then per-game opportunity, then applies an efficiency rate pulled toward the positional mean in proportion to how little evidence supports it. As of v2.2.0-beta, quarterback and tight-end availability defaults to a healthy-starter games baseline rather than carrying last season’s games count forward: games played is close to unpredictable year over year, and at those two positions letting it carry forward buried durable starters who happened to miss time. A current injury designation still discounts a player separately. Touchdowns are projected from a per-opportunity conversion rate rather than carried forward as a count, and that rate is regressed harder than anything else in the model, because finishing rate is the least persistent thing a player does.

Train on data available through 2024. Predict 2025. Score against what actually happened in 2025. The evaluation season appears on exactly one side of the comparison — as ground truth — and never as an input.

Leakage control is explicit. Nothing from 2025 enters the projection path; the positional priors are recomputed from 2024 inside the model itself. The shrinkage constants were fitted on 2023 to 2024 and the age curve and rookie priors on 2021 to 2024, with 2025 held out. The Sleeper live overlay is deliberately excluded, because it describes today's depth charts and injuries rather than the 2025 preseason, and feeding it in would be time travel.

The model is scored against two baselines, and the first one is the hard one:

• Last season repeated — a player’s 2024 PPR total as his 2025 projection. This is what most casual ranking effectively is, and it is genuinely hard to beat.
• Positional average — every player at a position gets that position’s mean. The no-information floor.

The scored population is every player with at least six games in 2025: 442 of them. The model produced a projection for 422, which is 95.5 percent coverage. The 20 it missed entirely are scored as 0.0 rather than excluded, because a coverage failure is a failure and hiding it would flatter the average.

The backtest is one command with no arguments. It needs no ML library, no new dependencies, and no paid data — it is arithmetic and documented heuristics, which is what makes it auditable.

Model version

fft-projection-2.3.0-beta

Figures on this page measured against

fft-projection-2.3.0-beta

Training window

nflverse seasons through 2024

Held-out season

2025

Run transcribed here

2026-08-21

Scoring

PPR season totals, plus a separate per-game cut

It prints the full report to stdout and writes a copy to node_modules/.cache/fft-backtest/backtest-latest.txt, which is the file the model API serves as its backtest summary. Output is deterministic: two runs on the same inputs produce identical numbers, and if they do not, that is a defect. The figures on this page moved from the previous transcription because the model changed, not because the harness did — the same harness was run against both model versions on the same day and the same inputs, which is what makes the before-and-after comparisons below legitimate.

The full engineering write-up, including the parts too detailed for this page, lives in the repository at docs/methodology.md alongside docs/projection-model.md. The harness itself is scripts/backtest-projections.mjs and the engine is src/utils/projectionModel.ts.

Spearman is rank correlation against actual finish, which is usually what a drafter cares about. MAE is mean absolute error in PPR points. Higher Spearman is better; lower MAE is better.

Collected in one place, at the same prominence as the wins, because this is the section that makes the rest of the page worth reading.

The projection is one half. The other half is the board of disagreements: what the model says about a player, set against what the draft market is charging for him, with the measured football reason for the difference attached. Four signals feed it, each weighted by how well it held up rather than by how interesting it sounds.

A verdict requires both halves. The signals must sit in the top or bottom quarter of the scored population and the market must be mispricing the player in the same direction, by more than eight picks. Strong signals with the market already agreeing come back ALIGNED — which is most of the board, and is the honest answer. The draft market is largely efficient, and a tool claiming an edge everywhere is not being straight with you.

Things the model cannot see, stated here rather than buried, because someone deciding whether to trust a number needs these more than they need the method.

• Coaching and scheme changes. A new coordinator can reallocate targets in ways no prior-season statistic anticipates. Not modelled at all.
• Offensive line quality, and its effect on rushing efficiency. Not modelled.
• Quarterback changes. A receiver's projection does not move when his quarterback does. A real and known gap.
• In-season injuries. The model projects a healthy-ish games total from history and cannot know about a Week 3 hamstring.
• Holdouts, suspensions and late roster moves beyond what the live overlay reflects at build time.
• Team changes are handled crudely — by widening the uncertainty, which captures 'we are less sure' but not 'this specific offense throws 90 more times a year'.
• Age curves are positional averages applied uniformly. Individual aging varies far more than the curve implies.
• Point estimates only. The model outputs an expected value, not a distribution — no ceiling, no floor, no bust probability. Twelve points a game from a volatile player is not the same asset as twelve from a stable one.
• Name matching across sources is fuzzy. Suffixes, apostrophes and legal-name changes cause mismatches, and a mismatch usually drops a player's history and makes him look like a rookie.
• No kicker or team-defense projections. Neither is meaningfully opportunity-modellable with these inputs, and producing a confident-looking number anyway is exactly what this work removed.