How FORGE produces a projection
An overview of the pipeline from play-level input to graded projection. Specific model weights and formulas are proprietary and are not published; what follows is the process and what each output means.
Sportsbook lines, betting spreads, totals, odds, implied probabilities, consensus prices, and line movement are not inputs to the FORGE v6.1.9 engine. When displayed elsewhere in FORGE, market data is external comparison and audit context shown only after FORGE has produced its independent projection.
- 01
Play-level inputs
Every play is recorded with its situation — down, distance, field position, time, and score state — so that outcomes are evaluated against what the situation was worth, not raw yardage totals.
- 02
Opponent and context adjustment
Unit performance is adjusted for the quality of the opponent faced, venue, rest, and playing conditions. A strong result against a strong defense counts for more than the same result against a weak one.
- 03
Team and unit ratings
Adjusted performance is expressed through the FORGE Rating, alongside offensive, defensive, and quarterback rating outputs. Recent games carry more weight than early-season games.
- 04
Game simulation
Each matchup is simulated repeatedly from FORGE team, unit, quarterback, and context ratings. The projected score, FORGE projected margin, margin distribution, and win probability are generated without sportsbook spreads, totals, odds, consensus prices, or market movement.
- 05
Calibration
Raw probabilities are calibrated against historical outcomes so that a stated 70% corresponds to a roughly 70% observed win rate. Calibration is re-checked every week.
- 06
Grading
Once a game is final, the projection is scored on winner accuracy and margin error and added to the public performance record. Nothing is removed retroactively.
Reading the outputs
- FORGE Rating
- Overall team-strength output adjusted for opponent quality.
- FORGE projected margin
- FORGE's own simulated point differential, expressed toward one team. It is not a sportsbook spread or betting line.
- Win probability
- The calibrated share of simulations each team wins.
- Confidence
- How strongly the model separates the two teams in the current run.
Principles
Outputs, not opinions
FORGE publishes what the model produces. Pages do not add narrative predictions on top of the numbers.
Uncertainty stays visible
Probabilities, distributions, and confidence levels are shown alongside point projections, never in place of them.
Market data stays separate
Sportsbook and market data does not influence the FORGE engine. When shown in Betting or Market Comparison, it provides context for comparing the independently generated FORGE projection with the market and for postgame audit.
Accountable by default
Every projection is graded and the record is public, including the weeks the model gets wrong.
Independent
FORGE is not affiliated with any league or club and uses no league or club marks. Teams are identified by city, name, and standard abbreviation.
External data sources
Play-by-play data provided by nflverse and nflfastR under CC BY 4.0. nflverse-data · nflfastR · CC BY 4.0
See the process applied to real games on the NFL predictions page, or view team strength on the NFL power rankings.