Models

A configurable, weighted model registry

SportsAI.ML does not hard-code model versions. Administrators register providers and models with capabilities, pricing, specializations, calibration and reliability — then the consensus engine weights each one.

Providers

Supported AI providers

Routed through the Vercel AI Gateway, with direct connections where commercial terms require.

OpenAIAnthropicGooglexAIMetaMistralCohereDeepSeekAmazon BedrockAzure AIGoogle Vertex AISelf-hosted OSS

Registry

Model registry

Masked configuration view. Secrets are never displayed after entry — only masked values and last-rotated dates.

Model	Status	Max context	Latency	Retention	Error rate
Grok 4.5 xAI	Active	256K	1.3s	Zero	0.4%
Grok 4.3 xAI	Active	256K	1.1s	Zero	0.5%
Grok 4.2 Reasoning xAI	Active	256K	2.2s	Zero	0.6%
Poisson + Elo ensemble SportsAI.ML	Active	n/a	40ms	n/a	0.1%

Performance

Model-performance leaderboard

Weights are continuously recalculated per sport, league and market.

Model	Specialization	Accuracy	Calibration	Reliability	Weight	Sample
Poisson + Elo ensemble SportsAI.ML	Soccer, Hockey	56.7%	0.94	0.99	1.31×	2,210
Gradient-boosted trees SportsAI.ML	Basketball, Football	57.8%	0.93	0.98	1.27×	2,050
Grok 4.5 xAI	Basketball, Football	58.4%	0.92	0.98	1.24×	1,840
Grok 4.3 xAI	Soccer, Tennis	57.1%	0.90	0.97	1.18×	1,620
Grok 4.2 Reasoning xAI	Baseball, Hockey	55.9%	0.88	0.96	1.05×	1,490

Weights derive from historical accuracy, calibration, specialization, data completeness, freshness and reliability. Demonstration data.

Proprietary layer

Statistical prediction models

General-purpose LLMs are never used alone. A modular statistical layer runs alongside them.

Logistic regressionPoisson & bivariate PoissonElo & adjusted EloBayesian hierarchicalGradient-boosted treesRandom forestsNeural networksTime-series forecastingMonte Carlo simulationExpected-goals / expected-pointsPace & efficiencyInjury-adjusted projectionsEnsemble stackingIsotonic / Platt calibration

Final model weight

Weight = Historical Accuracy × Calibration Quality × Sport Specialization × Market Specialization × Data Completeness × Data Freshness × Reliability

Backtests use only information available before each event started to prevent data leakage. Simulation counts shown reflect simulations actually run.