Combined ensemble
62.5%
Accuracy on 285 holdout games
AUC-ROC: Not reported
Transparent training, accuracy, market, and methodology reporting
Model construction and transparency
See how game data becomes a probability, why three models are combined, and which inputs the tree-based component used most. Construction details explain the system; they do not establish future performance.
The shortest path from source data to a published home-win probability.
Prediction assembly line
How one final probability is constructed and then evaluated.
01 · Source
Historical team, matchup, availability, and game-context inputs.
02 · Transform
Validated columns form the candidate pool before component-specific selection.
03 · Three lenses
Tree-based and regularized linear models estimate home-win probability.
04 · Run config
Configured contribution shares combine the three component probabilities.
05 · Published output
One final home-win probability becomes the model prediction shown to visitors.
Holdout · not used for training
285 games test the completed pipeline
Evaluation measures this historical sample; it does not guarantee future performance.
Component roles, same-sample evaluation, configured weights, and agreement coverage.
Each component sees the same game inputs through a different mathematical lens. Their home-win probabilities are combined with configured weights. The weights are implementation inputs, not proof that one model is better.
62.5%
Accuracy on 285 holdout games
AUC-ROC: Not reported
Pattern finder
Finds nonlinear patterns and interactions in the shared feature matrix.
Signal reducer
Fits a regularized linear view that can reduce the influence of redundant inputs.
Probability lens
Provides a simpler probability estimate from linear relationships.
All reported accuracy values above are directly comparable only because they use the same holdout sample. A component scoring higher on this sample does not establish that it will remain better on future games.
Source: game outcome default config.
QuantCup ensemble
62.5%
holdout accuracy
Stream length is a visual aid for configured contribution only; it does not represent model quality.
Agreement describes whether component picks pointed to the same team; it does not measure correctness.
of 285 comparable games
of 285 comparable games
excluded from agreement rates
Registered inputs are checked for configuration, availability, and selection before component-specific training matrices are built.
Candidate inputs
70
Linear models
53
Tree model
55
Selected union
55
Ranked evidence
55
Unavailable
0
70 candidate inputs produced a 55-feature selected union: 53 for the linear components and 55 for XGBoost.
Feature category constellation
Node size represents the number of reported features, not importance.
Selected union
55
model features
The five highest source-ranked features in this run.
rolling_8g_points_for_diff
A home-versus-away difference for 8-game rolling points scored.
#1
source rank
rolling_4g_win_rate_diff
A home-versus-away difference for 4-game rolling win rate.
#2
source rank
rolling_16g_point_diff_diff
A home-versus-away difference for 16-game rolling point difference.
#3
source rank
away_qb_available
A model input describing away qb available.
#4
source rank
rolling_16g_epa_offense_diff
A home-versus-away difference for 16-game rolling epa offensive.
#5
source rank