TDSIGNAL
Top plays

How it works

Every number on TD Signal comes from a chain of simple, visible steps. Here's the math behind both models.

Touchdown model

Expected TDs = team TDs × player share × defense × weather × usage trend
Chance to score = 1 − e−expected TDs

Team touchdowns

Projected points from the spread and total, times about 0.105 touchdowns per point.

Player share

His portion of team touchdown chances, weighted toward red-zone carries and targets.

Adjustments

The opponent's touchdowns allowed to his position, wind and rain, and whether his role is growing.

Model accuracy

The touchdown model is tested against every past game: it re-runs each week using only what was known beforehand, then checks who actually scored. A correction layer is refit to those results every morning, so the model keeps learning from its misses.

0%0%20%20%40%40%60%60%80%80% Predicted chance to score Actually scored
4,205Past player-games tested
17Weeks of history
+1%Accuracy gain on held-out weeks

Points on the dashed diagonal are perfectly calibrated. Gray is the raw formula, orange is after the correction layer. Accuracy gain compares prediction error (Brier score) on the most recent quarter of weeks, which the correction layer never saw while fitting.

Prop model

Volume = plays × pass or run rate × target or carry share
Projection = volume × efficiency × defense × weather × venue × role
Final = 75% model projection + 25% season average
Early season: average = this season × (games + 1) ÷ 10 + last season × the rest

Pace

Mostly the team's own plays per game, partly the opponent's, nudged for high or low totals.

Game script

Underdogs throw more and favorites run more: each point of spread shifts pass rate about a third of a percent.

Volume

Team pass attempts or carries times the player's target or carry share, adjusted for his usage trend.

Efficiency

Catch rate and yards per catch, carry, or attempt from his season, the core of each projection.

Defense

The opponent's yardage rank (up to about 8% either way) plus how it defends his position.

Small samples

A few big games can skew an early-season average, so until about Week 10 each player's numbers blend in last season's (same team only). Four games this season counts for 50%, nine or more for 100%.

Injuries

When a player is ruled out, his teammates inherit his workload: the next running back up takes most of his carries, and his targets spread across the other receivers. Each player keeps his own efficiency.

Conditions

Wind and rain for passing, a small home-field bump, a turf speed bump, and a dome boost for passers.

Chance to beat the line = P(outcome > line)

Receptions use a Poisson distribution and yardage uses a normal distribution with a spread that fits each stat (passing yards are steadier than receiving yards). The floor and ceiling cover the middle 80% of outcomes. For now the line is the half-point nearest each player's season average; once sportsbook lines are connected, they'll replace it so you can see true value.