This page explains how the Snapcount projection engine actually works — the season-regime architecture, the volume-times-efficiency decomposition, the walk-forward training that never sees the future, and the discipline that keeps the numbers honest. It uses real terms (Spearman, calibration, conformal) without the formula walls. When you want the exact math, the white paper has every definition and every number.
Every projection site hands you a single number — "17.4 points" — with no range, no confidence, no track record you can audit, and no way to know whether that number came from a model or a hunch. When the projection misses, it's quietly replaced next week. The incentive is to look precise, not to be accurate.
Our answer is not "we're smarter." It's a measurement discipline: one un-game-able yardstick, training that structurally cannot peek at the future, and a public record of hits and misses — including the head-to-heads we currently lose.
Most projection systems run one model from September to January. But the information available in Week 1 (nothing current — only last season plus an offseason of churn) is a different world from Week 10 (nine weeks of role and usage data). Our serving layer switches estimators by week-of-season, and each regime's estimator is the one that actually won its validation — even when the winner is embarrassingly simple.
At Week 1 every within-season signal our models feed on is null — there is no current season yet. We spent five research iterations trying to beat a "model-free book" (each player's prior-season average, playoffs included, plus recent form for QBs) with real models fed offseason information: betting-market priors, roster churn, coaching changes, contract news. Every model candidate failed the gate; the simple book won. So the book is what we serve at Week 1 — it scores 0.571 weekly rank correlation on 2021–2024 backtests vs 0.550 for the recency family we'd otherwise use (+0.021). Serving the humble winner instead of the fancy loser is the whole philosophy in one decision.
The core engine doesn't predict yards directly. It splits every player-week into opportunity (how many targets, carries, or attempts will this player get?) and efficiency (what does he do with each one?). Opportunity is driven by role — snap share, target share, depth chart — and is far more predictable than efficiency, which mean-reverts hard. Our headroom audits show most of the achievable ranking edge lives in projecting volume.
Every fit is walk-forward: to predict season S, week W, the model trains only on rows from before that point — earlier seasons, or earlier weeks of the same season. Random k-fold cross-validation is structurally invalid for this domain (it leaks future weeks into training), so we never use it. For Weeks 1–4 the rule gets stricter still: the model may train only on fully prior seasons, because the normal in-season window would overlap the weeks being predicted.
Regularized walk-forward models project each player's this-week targets and carries from strictly lagged role signals; a gradient-boosted median (quantile loss) produces the displayed yards. Beats the recency baseline in 4/4 backtest seasons at WR and TE, 3/4 at RB.
Every QB model we've built failed validation (0/4 seasons vs the baseline), so QB serves a tuned recency average — and we say so instead of shipping a model that grades worse than the naive answer.
Touchdowns are predicted as probability distributions (P of 0, 1, 2+ TDs), not counts — boosted-tree classifiers with per-position probability calibration. Measured any-TD calibration error: 0.031.
Every projection ships as a floor–ceiling band targeting 70% coverage, width scaled per player by trailing volatility. Measured coverage: 0.699 RB / 0.712 WR / 0.708 TE — the target, hit, not asserted.
Week 1 serves the validated prior-season book; Weeks 2–4 serve the strict prior-seasons-only model. Promoted 2026-07-12 after five research iterations failed to beat the book at W1.
A calibrated probability that a player beats or misses his own expectation — a confidence layer, not an accuracy moat, and we label it that way.
The architecture is table stakes. What actually separates the system is the loop every change must survive before it can touch a served projection.
time_to_throw is blocked, the weekly aggregate
avg_time_to_throw is allowed — because one is charted after the game and one is
published pregame-usable.We built the prior — do some teams reliably start seasons slow under the same coaching staff? Across 127 same-staff consecutive-season pairs, early-vs-late form anti-persists (correlation −0.217): last year's slow start predicts, if anything, the opposite. Verdict: NULL — the feature is refuted and retired, on the record.
We hand-curated every coordinator change 2015–2026 (with citations) and fed new-coach and scheme-import features to the Week 1 model. It failed the gate in 3 of 4 seasons and lost to the simple book. The offseason's coaching digest is team-level news; Week 1 player ranking stays owned by last season's volume.
Pregame spreads, totals, and season win totals — the market's own digest of the offseason — failed the Week 1 gate both inside the model and as a re-ranking tilt. (They do show a real Week 4 hint, now cross-confirmed by four independent feature families — that's a roadmap item.)
Five iterations of Week 1 candidates, every one graded against a dumb prior-season book. The book won every time. What we serve at Week 1 is the book. Discipline means the incumbent keeps its seat until something honestly beats it.
Every claim above traces to a published artifact you can open right now.
We froze the model on 2016–2024 data, then replayed the entire held-out 2025 season through it — 5,227 player-week projections, generated after the season ended, scored against what actually happened. A holdout scorecard, not a live track record — and labeled as such.
The board, graded week by week for 2024 — including the biggest misses, because a replay that only shows hits is marketing, not measurement.
Per-week graded scorecards in machine-readable form. Same fixed grading rule every week.
Head-to-head vs RotoWire and ESPN on identical players, weeks, and yardstick. It currently shows us losing — see below — and it ships anyway.
| Position | Snapcount | Best provider | Gap |
|---|---|---|---|
| QB | .416 | .546 (RotoWire) | −.130 |
| RB | .730 | .771 (RotoWire) | −.042 |
| WR | .715 | .745 (ESPN) | −.030 |
| TE | .643 | .682 (ESPN) | −.038 |
Mean weekly rank correlation (Spearman), TD-free score, 2024 W9–18, all-common population — every row graded on players covered by truth, our board, and every provider. Source: market_scoreboard.json.
The white paper documents every formula, mask, hyperparameter, and validation gate on this page, with the same numbers — nothing rounded up for marketing.