HeatRadar started as a baseball site. The NFL section runs on the same idea: build a number from things you can actually measure, show the work, and say plainly where it is weak.
This page is the method. If you want the numbers themselves, start with the weekly rankings, the start/sit tool or the heat index.
Every player, every week, is the same three things multiplied together:
Touchdowns are built separately, from goal-line work rather than total volume. Those two are not the same thing: across 161 back seasons the correlation between a running back's total carries and his carries inside the five is only 0.56. One back took 243 carries and got four of them inside the five. Another took 59 fewer and got nineteen.
Every efficiency stat is blended against what the defense on the other side actually allows to that position, weighted by how much of a sample the player has. The weights are measured, not chosen:
Volume moves with the game environment too. A high total means more passing for both sides, and a big favorite runs more. Both effects are small and both are fitted rather than assumed, on 7,051 team games. The spread barely touches passing volume at all, which is worth knowing if you have ever benched a receiver for being on a favorite.
The published weekly number is half our model and half the betting market, blended one stat at a time. A player with a touchdown price but no rushing line gets a market-informed touchdown number and a pure-model rushing number. A missing line is not information about the player, it is information about how widely he is bet, so it never drags a projection down.
There is also a market-only view that takes our model out entirely. That one is strict: a player appears only if the books have priced every stat his fantasy points are made of. A running back with a touchdown price and no rushing line would otherwise show up as a market opinion the market never expressed.
On the prop tables you will see three numbers. The Vegas line is what the book posted. Fair value is that line with the book's margin removed and converted from a median to an average. Our projection is ours. The middle column is the one worth understanding: a posted line is a median and a projection is an average, and because fantasy outcomes are right skewed the average sits above the median. Compare our number to the raw line and you will call four out of five players an over. Compare it to fair value and the split is even, which is what a real disagreement looks like.
Heat rates a game as a place to have fantasy players. It comes from four things: the points each team is expected to score, how fast both offenses play, the combined total, and how close the game projects to be. How much each one counts was measured across 2,716 team games rather than picked, and the first version of this was wrong in a way worth admitting: it left the team's own implied total out entirely, which put the highest scoring offense on the Week 1 board near the bottom of the ratings.
Team volume, recency weighting, per-touch efficiency, games played, goal-line touchdowns and the market blend all run automatically, the same way every time. Role changes, rookie arrivals, camp reports, a back inheriting a departed teammate's goal-line work, the effect of a receiver changing teams: those are judgments, made by hand, one player at a time.
The model cannot read a beat report. It also cannot know that a historical rate was earned in an offense we are projecting to change, which is the standing test applied to any number that looks too good.
Stated plainly, because a methodology page that only lists strengths is marketing.
The weekly rankings, start/sit tool and heat index rebuild every morning from Wednesday through Sunday. Fresh prop lines and injury designations are pulled first, then the projections, then the pages. Every step has to succeed before anything goes live, so a failed odds pull leaves the previous board up rather than publishing half a new one.
The rest-of-season rankings and the full-season boards move when the projections behind them change, which is less often and always deliberate.
One person: Jeromey Rome, a former sports journalist turned software engineer. Same as the baseball side of the site. Nothing here is syndicated and nothing is generated from a feed.