🏈 Exhibit: How the NFL Predictive Algorithm Works

Step-by-Step Summary

Step	Purpose	Example Output
1. Data Inputs	Pull 8-week rolling team stats, Vegas lines, injuries	Yards/play, turnover %, rest days
2. Feature Engineering	Compute matchup-adjusted strengths	Off – Def differential
3. Model Prediction	Predict continuous expected points per team	Home: 24.3, Away: 19.8
4. Post-Processing	Round to realistic NFL scores	→ 24–20
5. EV Computation	Compare model vs. book probabilities	+3.8% expected edge
6. Output	Display in report with odds & predicted score	"Lines Over the Next 7 Days" table

Feature Engineering — Deeper View

Sub-step	What It Does	Example
1. Normalize stats	Standardize team metrics so all inputs are on a comparable scale.	Yards/play → 0.65
2. Rolling averages	Use 8-week moving window to smooth randomness in performance.	24 → 22.7 adjusted
3. Matchup differentials	Compute offense vs. opponent’s defensive strength.	KC offense (0.8) – NYJ defense (0.5) = +0.3
4. Vegas baselines	Convert spread + total into implied team points.	Spread −3, Total 48 → 25.5 vs 22.5
5. Composite features	Combine all signals: efficiency, rest, travel, home-field, weather, etc.	X = [net_strength, rest_days, dome_flag]

Expected Value (EV) Computation — Breakdown

Component	Description	Formula / Logic
Model Probability	Chance of winning or covering derived from model simulations.	P_model = 67.2%
Implied Probability	Converted from sportsbook odds.	P_implied = 1 / (1 + (odds/100))
Expected Value	Model’s predicted edge over bookmaker.	EV = (P_model − P_implied) × 100
Filters	Retain bets only if EV ≥ threshold.	1% for spreads, 2% for ML/totals

Internal Mechanics

Regression trained on last 8 weeks (rolling window).
Uses both team efficiency and Vegas-implied baseline.
Produces continuous expected points for mathematical accuracy.
Displays quantized scores (0, 3, 7, 10, 14, etc.) for human readability.

Future Enhancement

Integrate realistic scoring directly into training:

Use a score bucket loss that penalizes predictions far from real NFL scoring clusters.
Alternatively, simulate outcomes from a Poisson-style scoring distribution aligned with actual NFL point structures.