How CHD Scout Works
CHD Scout is a multi-factor prediction model that runs as a SQL function inside our database. For every NCAA Division I basketball game, it produces a predicted final score, win probability, and confidence rating. The system fires automatically when a game's status changes to live, capturing the prediction with all contextual factors frozen at tip-off time. No look-ahead bias. No retroactive adjustments.
The model — currently on version 21 — was backtested against 2,275 games and validated through 21 iterations of development. V21 addressed the accuracy collapse that hit conference play, where familiarity between opponents and reduced home court advantage caused earlier versions to overpredict favorites. The result: a 7.14 percentage point improvement in conference game accuracy, from 66.6% to 73.7%.
Prediction Pipeline
Team efficiency, NET rankings, player stats, and schedule data sync from ESPN every 5 minutes
Nine prediction factors computed for each upcoming matchup: efficiency, HCA, form, rest, and more
When the game goes live, the prediction is locked in — predicted scores, margin, confidence, model version
After the final buzzer, every prediction is graded against the actual result and logged permanently
Nine Factors That Drive Every Prediction
Every CHD Scout prediction is built from nine components. The model starts with efficiency margin as the foundation, then layers on contextual factors that academic research and our own backtesting have shown to improve accuracy. Each factor has been individually validated — ideas that hurt accuracy were rejected, including Pythagorean ratings, defensive stat overlays, and NET trajectory signals.
Efficiency Margin
The foundation. Compares offensive and defensive points per 100 possessions, adjusted for opponent quality. The single most predictive metric in college basketball.
NET Rankings
The official NCAA ranking integrated as a quality signal. Especially important in toss-up games where NET rank is more predictive than raw efficiency.
Player Form
5-game rolling PRA (Points + Rebounds + Assists) with recency weighting. Detects hot and cold streaks before they show up in season averages.
Home Court Advantage
Venue-specific HCA calculation based on actual home/away performance. Not all home courts are equal — Cameron Indoor is worth more than an empty arena.
Conference Dampener
Reduces HCA by 30% for conference games. Home teams win ~52% in conference play vs ~65% in non-conference — familiarity erases the home edge.
Margin Compression
Applies a 0.90 factor to predicted margins above 8 points in conference games. Conference opponents rarely get blown out — the model accounts for this.
Rest & Travel
Adjusts for days between games and travel distance. Back-to-back road games and cross-country flights measurably impact performance.
Competitive Boost
Venue-scaled boost for close predicted games. When the margin is tight, the home team historically outperforms — this captures that edge.
Nudge System
Reweights signals in toss-ups: 20% efficiency, 50% NET, 30% form. In coin-flip games, overall quality (NET) matters more than raw efficiency.
V22 investigation tested 9 additional ideas from academic papers. Four were adopted (dynamic nudge weights, conference-specific HCA, variance dampener, Elo). Five were rejected (Pythagorean, margin quality, defensive stats, 3PT variance, variance-only).
Accuracy by Confidence Level
Not all predictions are created equal. The model assigns a confidence level based on the predicted margin — and accuracy scales directly with confidence. Strong picks are right nearly 9 times out of 10. Toss-ups are barely better than a coin flip, which is exactly what you should expect from games that are genuinely too close to call.
Games with a clear favorite. Large efficiency gaps, lopsided NET rankings, or dominant home court. The model's bread and butter.
Solid favorites but the underdog has a realistic path. Conference games between tournament-caliber teams often land here.
Slight edge to one side. The nudge system begins adjusting signal weights, giving more influence to NET rankings and recent form.
Essentially coin flips. A late free throw, a momentum run, or a cold shooting stretch decides these games. 55% is actually strong for this tier.
Why toss-ups matter: True toss-ups are inherently unpredictable — 55% is actually good for games that are essentially coin flips. Any model claiming 70%+ accuracy on games with a sub-3 margin is likely overfitting. The honest approach is to acknowledge uncertainty, which is why CHD Scout displays confidence levels prominently on every prediction.
How CHD Scout Compares
The college basketball analytics landscape includes several well-known models. Here is how CHD Scout stacks up against KenPom, BartTorvik, and ESPN BPI across the metrics that matter most.
Accuracy estimates for KenPom and BartTorvik are based on public analyses and community reports. ESPN BPI does not publish accuracy metrics. CHD Scout accuracy is verified in real-time on our Accuracy page.
How to Read a CHD Scout Prediction
Every game on College Hoops Data displays a prediction card with several key elements. Here is what each piece means and how to use it.
The team the model favors and by how many points. A prediction of "Duke -7.5" means the model expects Duke to win by 7.5 points.
Strong, Moderate, Lean, or Toss-Up based on the predicted margin. Higher confidence means higher historical accuracy for similar picks.
The top contributing factors for this specific prediction — which inputs moved the needle most. Look for hot/cold player indicators and HCA flags.
Displayed on every prediction so you can track which model version generated it. Currently V21 for all new predictions.
Model Evolution: V17 to V21
CHD Scout is not a static formula. Each version is backtested against thousands of games before replacing the previous version in production. Ideas that do not improve accuracy are rejected — V20 (NET trajectory) was tested extensively and shelved because directional signals under 60% accuracy in toss-ups always hurt overall performance.
Introduced the nudge model to handle toss-up games without distorting confident predictions.
Added rest/travel adjustments, HCA recalibration (2.0 to 3.5 baseline), and star player absence detection.
Venue-scaled boost in close games. Backtested across 4,122 games: 75.91% to 76.40%, MAE 9.34 to 9.27.
Tested 7-day NET rank trajectory. Found directional signals < 60% accurate in toss-ups always hurt overall accuracy. Shelved.
Conference HCA dampener (0.70), margin compression (0.90), HCA recency weighting (0.92 decay). Validated on 2,275 games: +7.14pp accuracy gain.
Frequently Asked Questions
How accurate is the CHD Scout prediction model?
CHD Scout has achieved 76%+ winner accuracy across 4,000+ games in the 2025-26 season, with a mean absolute error of approximately 8.2 points. For high-confidence games (predicted margin 12+), accuracy reaches 85-93%. True toss-up games (predicted margin under 3 points) sit around 55%, which is expected since those games are essentially coin flips where small in-game momentum swings decide the outcome.
Is College Hoops Data free to use?
Yes. Every prediction, accuracy metric, NET ranking, quad record, and player stat on College Hoops Data is completely free with no paywall. Unlike KenPom ($24.95/year), every feature on CHD is accessible without a subscription.
How does CHD Scout compare to KenPom and BartTorvik?
Like KenPom and BartTorvik, CHD Scout uses adjusted efficiency metrics as its foundation. CHD adds several layers on top: player form tracking (hot/cold analysis), venue-specific home court advantage with conference dampening, competitive game boosts, rest/travel adjustments, and a nudge system that reweights signals in toss-up games. CHD also provides transparent real-time accuracy tracking so you can verify every prediction.
When are predictions published for each game?
Predictions are generated and locked before tipoff for every Division I game. The system captures predictions when the game status changes to live, ensuring no look-ahead bias. Every prediction includes the model version, predicted scores, win probability, and confidence rating — all permanently stored and graded after the game ends.
Why are some games harder to predict than others?
Games between evenly matched teams (predicted margin under 3 points) are inherently unpredictable because tiny factors — a referee call, a momentum run, a cold shooting stretch — can swing the outcome. The model is strongest when there are clear efficiency and quality gaps between teams. Conference games are also harder because teams know each other well, home court advantage shrinks, and the talent gap narrows.
What is the nudge system?
The nudge system activates for toss-up and moderate games where the initial efficiency-based margin is small. Instead of relying primarily on efficiency (which has less predictive power in close matchups), the nudge reweights the signal mix to 20% efficiency, 50% NET rankings, and 30% recent form. This reflects the finding that NET rank is the strongest predictor in games between evenly matched teams.
Can I see historical accuracy data?
Yes. The Accuracy page on College Hoops Data tracks every prediction the model has ever made, broken down by confidence tier, time period, and game type. Every prediction is permanently recorded before tipoff and graded after the game, creating a fully transparent and auditable track record.