Model Center

v1.0-transparent — a weighted blend of named signals. No black box, no guarantees.

Model
Active version
v1.0-transparent
Every projection stores its version
Signals
9
All weights sum to 1.00
Projections this week
147
Week 1
Bet-grade calls
62
26 leans · 59 passes
Versions in data
Loading

Signals and weights

The whole engine, in one table.

SignalWeight
Season baseline
Per-game production across the full season sample. The anchor for everything else.
30%
Recent form
Last 3-5 games, weighted toward the most recent, to catch a changed role.
20%
Expected usage
Snap share, route participation, target share and carry share — opportunity before efficiency.
15%
Opponent matchup
Defensive efficiency, pressure rate and positional funnel of the opponent.
12%
Game environment
Expected pace, plays, total and projected game script.
8%
Injury / role change
Teammate absences that shift volume, plus the player's own designation.
6%
Home / away
Small venue effect. Deliberately kept small — it is a weak signal.
3%
Weather
Wind and precipitation for outdoor games; ignored in domes.
3%
Market anchor
A light pull toward the posted line, because the market knows things we do not.
3%

Weights are fixed for v1 and applied identically to every player. Nothing is fitted after the fact.

Projection inspector

Pick a projection to see how the number was assembled.

Projection
134.3
Passing Yards
80% range
72.4–196.1
σ 48.2
Market line
182.5
Diff -48.2
Model prob
84.1%
under side
Implied prob
50.0%
From the posted price
Edge
-34.1 pp
+0.683u
Passlow confsparse dataGenerated 9/13/2026, 10:42:59 AM · v1.0-transparent
Loading feature contributions

How a decision is reached

Edge, confidence and data quality are separate gates.

1. Edge

Model probability minus the probability implied by the price, in percentage points. Under 3 points is noise. Projection difference is reported separately because it answers a different question.

2. Confidence

Set by sample size and input completeness, never by the size of the edge. A large edge on thin data stays low confidence.

3. Decision

Bet needs 6+ points, high confidence and complete inputs. Lean is a real but modest edge. Pass is used freely and is not a failure.