The model
How our projections are built, tested, and wrong.
Every projection site says it’s accurate. Almost none show their grading. This page is the full accounting for the FantasyHQ 2026 model: the method, the rules it’s scored under, its results against real baselines including the expert consensus, and the specific players it got badly wrong. It’s written by the person who built it, who forecasts consumer demand for a living and holds fantasy football to the same standard.
Top-down, not player-by-player
Most projections build each player in isolation, which quietly lets a team’s parts add up to more football than the team will play. Ours starts with the whole offense and divides down, so a quarterback can never throw for fewer yards than his receivers catch.
Team budget
Project the whole offense first: how many pass attempts, carries, and targets the team will produce over 17 games, from its pace and pass-rate tendencies, with an adjustment when the coaching staff changes.
Player shares
Divide that budget among the roster. Each player gets a share of targets or carries from his prior usage, his depth-chart role, and (for rookies) draft capital. When a player leaves, his opportunity is explicitly redistributed to whoever remains.
Efficiency
Convert opportunities to production with per-touch rates (catch rate, yards per carry, TD rate), each regressed toward position norms by how predictable that stat actually is. Touchdowns get their own chain: team TDs, pass/rush split, then each player's share of red-zone opportunity.
Stat line
Team volume × player share × efficiency = a full 17-game stat line for roughly 900 players, haircut to expected games played (15 for QB/WR/TE, 14 for RB). Because of step 1, everyone's numbers add up to a coherent team total.
We do not project injuries
A deliberate stance, adopted in July 2026: established starters are projected at essentially a full healthy season, not at their own injury history. Games played is one of the weakest year-to-year signals in football, and a draft board that bakes last season’s injury into this season’s line quietly caps the upside of exactly the players who win leagues. In-season, the waiver wire replaces a hurt starter’s missed weeks; nothing replaces a healthy elite season you didn’t draft.
The stance had to pay rent like everything else: re-tested over all four seasons, projecting proven starters healthy reduced points error at every position. The honest cost is that when the injury does recur, the model misses high. It backtests as missing by less than the consensus in those seasons anyway (2025 Joe Burrow: model 307, consensus 341, actual 139; 2025 Kyler Murray: model 293, consensus 317, actual 81), and a small residual stays in each starter’s line because durability history isn’t zero information, just weak.
We tested one exception at quarterback: pricing each QB’s own expected games into his draft rank measurably improved the ordering. We shipped it for one day and took it out. It quietly moved a healthy-lined Joe Burrow from QB10 to QB20, which is exactly the hedging this section promises not to do. The ordering gain is real and stays documented in our ablation records; the product stance won.
The grading rules
A projection model is only worth what its test is worth. Ours is scored under four rules:
- No peeking. To grade a season, the model may only use information that existed before that season started (our cutoff is June 1). Depth charts are reconstructed from prior-season rosters plus documented transactions, so nothing from hindsight leaks in.
- Four full seasons. Re-run and scored on 2022, 2023, 2024, and 2025, at every position, 600+ graded players a season.
- Two scorecards. Rank ordering (did it put players in the right order?) and points error (how far off was each stat line?). A model can ace one and fail the other, so we report both.
- Same basis as the baseline. Every comparison scores both models on identical players under identical scoring. No cherry-picked pools.
Results against a real baseline
The baseline is a three-year weighted average of each player’s production: tougher than carrying last season forward, and roughly how a “set and forget” ranking behaves. The model beat it at every position in each of the four seasons, with about 42% less points error overall. The gap is widest at quarterback, where the error is more than halved.
Accuracy = 100% minus the average projection miss per player, measured against a front-line starter’s season (the average top-12 QB/TE and top-24 RB/WR finish, full PPR, 2022-2025). Both models scored on the same basis. Bigger bar = more accurate.
Against the expert consensus
The hardest baseline in this market is the FantasyPros preseason consensus, an average over dozens of professional analysts. Very few models beat it, and anyone claiming to crush it deserves your suspicion. Scored on identical players across 2022-2025, ours lands at parity with it on points error (a fraction ahead over the four seasons), ahead of it on both scorecards at WR, even with it on points error at TE, and just behind it at QB and RB:
| Pos | Rank orderingmodel vs consensus | Points errormodel vs consensus, pts/season | Edge |
|---|---|---|---|
| QB | 0.750 vs 0.757 | 52.8 vs 52.4 | consensus |
| RB | 0.770 vs 0.774 | 43.6 vs 42.9 | consensus |
| WR | 0.782 vs 0.767 | 39.2 vs 41.4 | model |
| TE | 0.747 vs 0.745 | 28.4 vs 28.5 | model |
Rank ordering = Spearman correlation between projected and actual finish (higher is better). Points error = mean absolute error of season fantasy points (lower is better). 4-year averages weighted by player count, archived preseason consensus, identical player pools.
Why that matters even where it’s a tie: the consensus is the market. Everyone at your draft is looking at some flavor of it. Our model arrives at its numbers independently, so the players where it disagrees with the room are exactly where an edge can live.
Where the model has been wrong
The model’s personality is skepticism: it trusts recent usage and discounts stories. That cuts both ways, and the clearest illustration is that it has been on both sides of Christian McCaffrey:
- 2022, badly wrong. Coming off two wrecked seasons, McCaffrey’s recent usage justified 144 points by the model’s math. The consensus said 293. He scored 356.
- 2024, exactly right. The same skepticism about a 28-year-old back coming off a monster workload projected 193 while the consensus said 350. He scored 48.
- 2025, wrong again. Model 164, consensus 295, actual 417.
The systematic blind spots, named plainly:
- No recent data reads as no talent. Calvin Ridley returned from a season-long suspension in 2023; with no recent usage to trust, the model projected 13 points. He scored 230. The consensus, which knew the story, said 218.
- It can’t see a situation change coming. Sam Darnold’s history justified 7 points in 2024. Minnesota made it 320. When a player’s context transforms, the crowd adapts faster than the data does.
- RB is its weakest position. Committee backfields and volatile roles are where human judgment still beats usage history; the consensus out-orders the model at RB (0.774 vs 0.770 over four years).
- Weeks are noise; seasons are signal. Week-level fantasy outcomes are close to unpredictable for anyone, which is why this is a season-total model. Don’t trust anyone selling weekly precision.
For balance, the same skepticism has hit big where the crowd priced in a repeat: it was out on 2024 McCaffrey (consensus 350, actual 48) and 2025 Justin Jefferson (consensus 312, actual 202; model 226, far closer to the number).
Failed ideas stay documented
Every feature has to pay rent in backtest error, and the ones that didn’t are recorded rather than quietly kept:
- Vegas win totals: rejected. Folding betting-market win totals into team volume improved error by a rounding digit. It stays off.
- A red-zone touchdown model: rejected. A dedicated model for red-zone conversion didn’t beat the simpler explicit TD chain in testing.
- Coaching-change flags: kept. Removing them measurably hurts team-volume accuracy. They stay.
From points to draft value
Projected points aren’t draft value. On the projections board, every stat line is scored under your league’s exact settings, priced on four seasons of realized scoring at each rank, and measured against a replacement level built by filling your league’s actual starting slots. Change the scoring or the roster shape and the whole value column rebuilds live.
The honest caveat, and the 2026 test
The model’s settings were tuned on the same four seasons it was graded on, so expect live 2026 accuracy a notch below the backtest. That is true of every model you’ll ever read, whether they say it or not. The real out-of-sample test is the 2026 season itself, and we’ll grade it in public: dated calls during draft season, and accuracy reports against the consensus once real games exist.