Model Methodology

How BetIQ Picks Work

Every pick on BetIQ comes from a GradientBoosting ensemble trained on 556 NCAA Tournament games from 2016–2025. Here's exactly what goes into each prediction.

Algorithm
GradientBoosting Ensemble
CV Accuracy
89.6%
Training Games
556 Tournament
Year Range
2016–2025
Features
21 Total
Estimators
300 Trees
Calibration
Platt Scaling
Sanity Cap
Applied

How predictions are generated

1
Data collection
Each morning at 8 AM CST, the cron job pulls today's matchups from The Odds API (live DraftKings lines) and looks up team stats from our database (seeded from Sports Reference CBB). Stats include: SRS, OffRtg, DefRtg, NetRtg, pace, TOV%, ORB%, TS%, STL%, BLK%, win%, SOS, and more.
2
Feature engineering
For each game, 21 differential features are computed — e.g. SRS_A − SRS_B, NetRtg_A − NetRtg_B. Features are then standardized using a StandardScaler fit on the training data (exact mean and scale are embedded in the route for offline use).
3
Model prediction
The GradientBoosting ensemble outputs a raw probability for team A winning. This is then Platt-calibrated to correct for overconfidence at the extremes. Injury adjustments are applied multiplicatively after calibration.
4
Sanity cap
If team B's SRS is >5 points higher AND NetRtg is >8 points better AND performance-above-schedule is >5%, team A's win probability is capped at 30% regardless of model output. This prevents runaway predictions on extreme mismatches.
5
Edge calculation
Model probability minus market-implied probability (from DK moneyline) = edge. Only picks with positive edge are surfaced. Kelly fraction is computed using fractional Kelly (25% of full Kelly) for bet sizing.

Feature importance

Feature importances from the trained GradientBoostingClassifier, measured as mean decrease in impurity across all 300 trees.

SRS (Simple Rating System)
41.8%
Strength of schedule-adjusted rating. Single most predictive feature.
Net Rating Differential
9.5%
OffRtg − DefRtg per 100 possessions, normalized for pace.
Win % Differential
4.5%
Head-to-head win percentage gap between teams.
Seed Differential
4.0%
Tournament seeding gap, reflecting selection committee strength assessment.
Injury Adjustment
3.2% + 2.9%
Per-player impact multiplier applied to both teams separately.
Defensive Rating
2.8%
Points allowed per 100 possessions — tempo-free defensive efficiency.
Coach Tournament Edge
2.8%
Historical tournament W/L rate × experience weight.
Block %, True Shooting %
2.6% each
Shot suppression and scoring efficiency per possession.
TOV%, ORB%, STL%
2.0–2.4% each
Turnover rate, offensive rebound rate, and steal rate differentials.
Pace, 3PT Rate, FT Rate
1.5–2.2% each
Game speed and shot selection tendencies.
Pythagorean Luck
2.1%
Win% minus expected win% — corrects for close-game variance.
Strength of Schedule
2.0%
Average opponent quality across the full season.
Performance Above Schedule
2.3%
How much a team over/underperformed given their schedule.

Backtesting & Validation

5-Fold Cross-Validated Accuracy
89.6%
Held-out test set accuracy across 5 folds
Training Period
2016–2024
9 full NCAA Tournament seasons
Test Period
2025 Tournament
Completely unseen — 67 games
2025 Out-of-Sample Accuracy
82.1%
55/67 correct outright predictions
Important disclaimer

Past accuracy does not guarantee future results. Sports betting involves inherent randomness and no model can predict outcomes with certainty. BetIQ is an analytical tool — always gamble responsibly and within your means. Must be 21+ and in a state where sports betting is legal. If you have a gambling problem, call 1-800-GAMBLER.

See Today's Picks →Create Free Account