How the model works

The methodology
behind every number.

Six data sources, a gradient-boosted baseline, and a Claude-powered reasoning layer. Every projection is back-tested weekly so claimed confidence actually means something.

Inputs

Six data sources

Every projection pulls from play-by-play, market, weather, injury, and athletic data — refreshed daily during the season.

01

nflfastR

Play-by-play, EPA, CPOE, success rate, 2006–present

02

Next Gen Stats

Air yards, separation, time to throw, coverage grades

03

Pro-Football-Reference

Depth charts, injury reports, snap counts

04

OpenWeather

Game-day weather: wind, precipitation, temperature, indoor/outdoor

05

Vegas market

Implied totals, spreads, line movement across sharp books

06

Combine / pro days

Athletic testing, 40-yard, vertical, three-cone percentiles

Pipeline

Six stages, end-to-end

01

Ingestion & cleaning

Daily pulls from six data sources, normalized into a unified Postgres schema. De-duplication, missing-value handling, and outlier flags before any modeling touches it.

02

Feature engineering

Weighted rolling splits (3-game, 5-game, 10-game) for every usage metric. Defensive matchup adjustments by coverage scheme. Game-script priors from Vegas implied totals.

03

Baseline projection

Gradient-boosted regression trained on 2006–2024 play-level data. Position-specific models (QB, RB, WR, TE) with cross-validation by season to prevent look-ahead bias.

04

AI reasoning layer

Claude 4.7 interprets the projection against the prop line, matchup context, and injury data. Outputs a confidence score, direction, and written reasoning trace.

05

Market edge computation

Compares our projection against book lines. Flags mismatches where the model differs from the market by 2+ standard deviations after weather and injury adjustments.

06

Back-test & recalibrate

Weekly back-test vs. actual outcomes. Recalibrate confidence bands so that claimed 70% picks actually hit 70% of the time. Miss rate is public on the pricing page.

Back-test · 30-day rolling

Calibration over claims

Updated weekly · 2025 season
Prop accuracy vs. market72.1%
Projection precision (RMSE)±2.3 pts
Edge vs. book lines+14%
Confidence calibration±1.8%

Confidence calibration measures how closely our claimed confidence (e.g. 70% of picks hit) matches actual hit rate over a rolling 30-day window. Lower drift = better calibrated.

Philosophy

What we won't do

No lock-picks or sure things

Confidence maxes out at ~85% for a reason. Anyone claiming 95%+ is either miscalibrated or lying.

No cherry-picked accuracy stats

We publish the full 30-day window and disclose when we regress. Streak-hunting is marketing, not methodology.

No model-washing bad inputs

If the data is missing or stale, we flag the projection as low-confidence — we don't fill gaps with priors and pretend.

See the model in action

Every tool is free during beta — AI prop analyzer, player database, 2026 draft prospects, WNBA hub. No signup required.