Bellcow GM / Methodology
How Bellcow thinks
Last updated 25 August 2026 · re-measured every season
Three sentences hold this whole page together. The data provides evidence. Bellcow's own models do the calculating. The AI only explains the result in a sentence. Nothing on the number side of the product is generated by a language model, and nothing on the language side is allowed to invent a number.
Why that ordering matters
A chatbot asked for a projection will produce one, confidently, every time — because producing text is all it does. Bellcow is built the other way round: the projection, the value, the bid and the odds are computed by deterministic code that returns the same answer every time from the same inputs. The model is handed those results and given one job: write them down clearly. Ask it something the engine has no answer for and it says it does not know.
The chain, end to end
What he actually did
Snaps, carries, targets, routes, red-zone touches, depth-chart rank — from play-by-play for every game, not a weekly box score.
What that was worth
Expected points: what a normal player scores on that exact workload. It separates a real role from a lucky touchdown.
What he does next
Projected carries, targets, yards and touchdowns — then scored under your league's rules, not a generic PPR default.
What he's worth to you
Priced against the players actually sitting on your waiver wire. Scarcity in a 14-team superflex is not scarcity in an 8-teamer.
What it does to your odds
Your roster with and without the move, each run through 1,500 simulated seasons of your remaining schedule.
What you should do
One ranked list of three moves, each with the exact player, the exact drop and the exact bid. You execute it on your platform.
1. The data: what counts as evidence
Bellcow computes its own football numbers. It does not license anyone's consensus rankings and it does not resell a fantasy platform's projections. Every source is registered with its licence and checked in code before its data is allowed anywhere near a recommendation.
| Category | What it is | Where it comes from |
|---|---|---|
| Play-by-play | Every snap of every game: carries, targets, air yards, red-zone usage, game script. | nflverse (CC BY 4.0) |
| Snap counts & participation | How much of his offence a player was actually on the field for, week by week. | nflverse |
| Rosters & depth charts | Who is on which team and where he sits in the pecking order. | nflverse |
| Official injury reports | League-published designations — Out, Doubtful, Questionable, IR, PUP. | nflverse |
| Schedule & byes | The real 2026 fixture list, used for bye detection and season simulation. | nflverse |
| Game lines | Spread, total and moneyline, as a read on expected game environment. | nflverse games data |
| Combine & draft | Athletic and draft-capital priors for players without much NFL sample. | nflverse |
| Weather | Public forecast data for outdoor stadiums. | National Weather Service |
| Trade-market values | What managers across thousands of leagues actually pay — an independent signal, never the brain. | Stats Guy Fantasy |
| Your league | Rosters, scoring, lineup slots, waiver system and budget, schedule, transactions. | Your platform, read-only |
What is deliberately not in that list. FantasyPros, PFF, RotoWire, Footballguys, 4for4, Establish The Run, Fantasy Points, RotoViz and the rest are benchmarks only — never ingested. We compare our output to theirs when we evaluate a change; their numbers never enter the product. Pro Football Reference is blocked outright for production use. The full source register, with licence status, is on the Data Coverage page →
2. The models: what does the calculating
Expected fantasy points
Given a player's workload — carries, targets, air yards, red-zone opportunities, snap share — what would a normal player have scored on it? Comparing that to what he actually scored is how the engine separates a real role from touchdown luck.
Trained on 15,751 player-weeks. Training error 3.24 PPR points — that is a training figure, not a holdout result. The holdout results are in section 5.
Projections
Component projections — carries, targets, yards, touchdowns — for the coming week and for the rest of the season. They ship as raw stat lines and are scored in your browser under your league's own rules: PPR setting, TE premium, passing-TD value, custom bonuses, D/ST points-allowed tiers. A superflex PPR league and a standard league therefore never see the same rankings.
Role change before the box score
Rising snap share, target share, route participation and red-zone work, scored before production shows up. In the 2025 holdout, quiet players the breakout engine ranked highest averaged 5.8 points a game over the following three weeks against 3.6 for the ones it ranked lowest. Players it flagged as scoring above their real workload lost 3.7 points a game over the next three weeks, against −1.1 for everyone else.
Replacement level and scarcity, for your league
Every player is repriced against the pool actually available to you: the best free agent at his position in your league, your own starters at that position, and the number of starting slots your lineup uses. This is why the same player can be a must-add in one league and correctly ignored in another.
FAAB sizing and claim order
The dollar figure is scaled to your remaining budget and to your league's observed winning bids — the median and the high end of what has actually been paid this season. Every figure, including the conservative one, is capped by what you have left. In priority leagues no dollar figure is produced at all; you get claim-order advice and a straight answer on whether the move is worth your top priority.
Two-sided by construction
A trade is only proposed when it improves both starting lineups — yours by a real margin and theirs by more than zero. Bellcow also estimates acceptance from the other manager's positional needs, depth and trade history. A deal that only helps you isn't a trade idea, it's a wish.
1,500 seasons, paired
Championship, bye and playoff odds come from simulating the rest of your season 1,500 times, using your real remaining schedule where your platform exposes one. Weekly matchup win probability is a separate 2,500-run simulation of the two starting lineups.
The important part is how changes are measured. The simulator is seeded, and the base roster and the changed roster share the same random draws (common random numbers). The 1,500 runs are split into 15 paired batches of 100, which gives each difference its own standard error, and nothing is printed unless it clears twice that error. Otherwise the product says “too close to call”. A no-op change now measures exactly 0.00 points; before this was fixed, the same no-op could read anywhere inside roughly ±1.9 points of pure noise.
Where a platform does not expose a matchup schedule — Yahoo, War Room and pasted leagues — Bellcow builds one deterministic balanced round-robin at import and says so on the Data Health page, rather than quietly redrawing random pairings on every run.
3. The AI: what it is and isn't allowed to do
Bellcow uses a large language model (Anthropic's Claude) for exactly one thing: turning engine output into readable English — the sentence under a recommendation, and the in-app assistant's answers.
- It cannot produce a projection, a player's injury status, a value, a bid or an odds figure. Those arrive as computed inputs.
- It cannot invent a fact about your league. It is given a compact summary of your league and told to answer from it.
- It says when it doesn't know. Asked something the engine has no answer for, the correct output is an admission, not a guess.
- It is optional. With no AI key configured the assistant falls back to built-in rule-based answers, and the rest of the product is unaffected.
Language models still make mistakes, and a wrongly-phrased sentence around a correct number is a real failure mode. Read the honest version of that in the AI Disclaimer, and exactly what is sent to Anthropic in the Privacy Policy.
4. Ranking: one currency for three different decisions
A waiver claim, a trade and a lineup change are normally argued about separately because there is no shared unit. Bellcow prices all three in the same one — the modeled change in your championship probability — so the list ranks itself. One exception is deliberate: a time-critical lineup fix, where a starter is out or on bye and doing nothing forfeits points today, is pinned above the odds ranking. Everything else sorts by the measured delta, with Bellcow's own 0–99 priority score breaking ties.
The plan is three moves. Everything else Bellcow found stays visible as “also considered” rather than padding the board. See a full sample plan →
5. The backtest, and what it actually measures
The models learned from the 2022–2024 seasons and never saw 2025. We then replayed 2025 week by week — weeks 2 through 17, sixteen weeks in all; week 1 has no in-season form to read yet and the final week is left out — showing the engine only what existed at that moment. No hindsight, no peeking at the result it was being asked to predict. Every player who was on a roster that week is graded, including the ones who got hurt later in the year. Training seasons are pinned in code so a retrain cannot ingest the 2025 holdout.
| Position | Avg error / week | Rank correlation | Hit rate |
|---|---|---|---|
| Running backs | 4.5 pts | 0.73 | 63% top 24 |
| Wide receivers | 4.3 pts | 0.66 | 50% top 24 |
| Tight ends | 3.5 pts | 0.65 | 49% top 12 |
| Quarterbacks | 6.6 pts | 0.43 | 50% top 12 |
| Kickers | 3.0 pts | 0.15 | 46% top 12 |
What the table means
- Avg error / week — the average absolute gap between the projected score and the actual score, in fantasy points, per player per week.
- Rank correlation — how closely our weekly ordering of players matched the order they actually finished in that week.
- Hit rate — how often a player we projected inside the top 24 (top 12 for TE, QB and K) actually finished there.
- Kickers are in the table on purpose. At 0.15 we have essentially no ability to tell one kicker from another week to week — and neither does anybody else. Leaving the row out would be cherry-picking. Play whoever has the best game script and don't spend a pick on it.
What the table does not mean
- It is not a claim that following Bellcow's recommendations wins leagues. It measures projection accuracy, not championships. Nobody can guarantee you a title, and we never will.
- It is not a head-to-head win over a paid feed. Published weekly-projection error for backs and receivers generally lands between 4.5 and 5.5 points a week, which is the neighbourhood we are in. We are showing our homework, which is more than most will do; we are not claiming a trophy for it.
- It is not permanent. Every figure is re-measured at the end of each season, and a change that makes the score worse does not ship.
One correction worth naming: an earlier version of this table read RB 67% / WR 53% / TE 56% / QB 56%, because the replay was scoring against end-of-season roster status and silently dropping players who got injured later in the year — an easier test than the one we said we were running. That was fixed, the numbers above are the honest ones, and they are lower than the numbers they replaced.
6. Where the model is weak
- Quarterbacks and kickers. A 0.43 rank correlation at QB is modest, and 0.15 at kicker is nothing. Treat those recommendations as low-conviction, and Bellcow labels them that way.
- No live gameday inactives. There is no commercial inactives feed behind the product, so Bellcow cannot see a 90-minutes-to-kickoff scratch. Re-check your lineup before kickoff.
- Acceptance estimates are modeled, not measured against real accept/reject outcomes. Treat them as a rough ordering, not a probability you can bank.
- Simulated schedules. Yahoo, War Room and pasted leagues get a modelled balanced round-robin, which makes their odds directionally useful rather than exact. The app says so where it matters.
- Formats we don't model. IDP, best ball, dynasty and guillotine are out of scope today. The supported-format list is on the homepage →
Related
Data coverage — sources, refresh cadence and what isn't live
A full sample GM plan
A guided walkthrough of the app
The legal data-sources statement
AI disclaimer