How projections are made
What the numbers mean, how we make them, how sure we are and what they leave out.
In plain English
We project how much each player will be used next game, turn that into stats and fantasy points, and adjust for his team's expected score and his injury status. Everything is built from public data, and we publish how every projection did on the track record page, so you can judge whether to trust it.
A projection is a best guess, not a promise: touchdowns and big plays swing single games. The rest of this page has the details.
Our approach
Football outcomes are noisy. A touchdown, a long play or a change in game script can swing a single game far more than any model can foresee, so we do not promise exact fantasy points.
Instead, each player page combines the signals that are predictable — how much the player is expected to be used, what that usage usually produces and how uncertain the result is — and shows what changed since the last projection and what it means for the next game.
Every number on a page comes from our models and data, and the text is checked against them when each page is built.
1. Opportunity: how much will the player be used?
The Opportunity Model projects football usage for the next game: targets for wide receivers and tight ends, designed carries and targets for running backs, dropbacks and designed rushes for quarterbacks.
Designed rushes are planned quarterback runs. They exclude scrambles and kneel-downs, which official rushing attempts (the box-score 'carries') include, so the two numbers can differ for quarterbacks.
It is a gradient-boosted count model trained on past seasons using only information available before each game: recent and season-to-date usage, the previous season, team context, the depth chart, injury reports and teammates who are out. It is tested on whole seasons it never saw (walk-forward validation).
2. Production: what does that usage produce?
The Fantasy Production Model turns projected usage into expected football production: receptions, yards, touchdowns, pass attempts, completions, interceptions, rushing attempts and fumbles lost.
Each stat is projected usage multiplied by the player's efficiency per opportunity. A player's own history is blended with the league average, and the less history a player has (or the noisier the stat, such as touchdowns), the more weight goes to the league average. Recent games count more where the data shows they should.
We then scale every stat by the team's expected points for the game, which come from public game lines (spread and total). We don't show the lines or give betting advice.
3. Fantasy points: a translation, not a separate guess
Fantasy points are computed from the projected stats with standard scoring rules (PPR, half PPR and standard). The models never train on fantasy points. Scored points use the official stats, including 2-point conversions and return touchdowns. Projections leave those two out: they are rare and not modeled.
4. Ranges: how certain is the projection?
The likely range is where the actual value falls about half the time (a 50% range), if the player is active. Floor and ceiling (an 80% range, 8 games in 10) are kept only where a week has no likely range. Production and fantasy-point ranges carry the uncertainty in the player's usage forward into the stats.
A range is shown only if, in back-tests on four full past seasons, ranges of that kind contained the actual result close to 80% of the time in every season, including for the hardest-to-predict players. Otherwise the page says the range is not shown. Today that applies to touchdowns, interceptions and fumbles (single-game counts are too small for an honest 80% range) and to quarterback dropbacks, passing stats and quarterback fantasy points (starters and backups make the outcome two-sided).
5. Timing and history
Every projection states when it was made (for example 24 hours before kickoff) and the data snapshot it used. A projection only uses information available before its own timestamp. Projections are stored permanently and never rewritten after the game; a newer projection is shown next to the previous one with the change.
Injury reports are built into the numbers. A Questionable player's projection is discounted by his last practice: to about 92% after a full practice, 71% after a limited one and 54% if he did not practice. Players listed Out or Doubtful get no projection, and they come off every pick list at each refresh.
6. Known limitations
Single-game fantasy outcomes vary a lot: for fantasy-relevant players a typical miss is several points even when usage is projected well, mostly because of touchdowns and big plays.
Quarterback changes that are not known before the projection time cause the largest usage misses.
Late injury news can arrive after the 24-hour projection, especially for Thursday and Friday games.
Route data and weather are not used; in our tests the available versions added little.
Model versions
Model versions on this site: opportunity_hgb_poisson_v1, opportunity_hgb_poisson_v1_r81, prod_eb_v1. Every projection is stored with its model and data snapshot, and each player page shows when it was made.