Erudil: an AI engine that prices every match outcome

The problem was to price dozens of outcomes of one match — winner, draw, totals, handicaps, both teams to score — from one coherent probability model, not a list of separate guesses; otherwise the model can claim a 60% home win and a 50% under-1.5-goals at the same time. The classical answer goes back to Dixon and Coles (1997), who modelled football scores with adjusted Poisson distributions and showed where the betting market was inefficient; the engineering problem is keeping such a model consistent, live and honest.

The model must also update when a line-up changes an hour before kick-off, cope with six data providers that disagree with each other, and — hardest of all — know when it has no edge and say so instead of forcing a pick.

Everything a user sees is derived from one correct-score grid: the 1X2 probabilities, totals, handicaps and BTTS never contradict each other because they are read off the same matrix.

It was built around one correct-score probability matrix per match, produced by a Poisson-based Monte Carlo simulation of ten thousand virtual matches and cross-checked by an ensemble of specialised modules. Every market is read off that matrix, so prices never contradict each other. Bookmaker odds are converted to implied probabilities with the margin removed and compared with the model; the pick is published only where the edge exceeds a calibrated threshold, and its success is judged against the closing line — the measure of skill that Pinnacle’s own betting resources (2024) describe as the most reliable long-run indicator.

• Outcome probability matrix. Goal expectancy per side from expected goals, expected points, form, head-to-head, venue and weather, simulated into a correct-score grid.
• Multi-module ensemble. Attack vs defence, historical trends, an NLP module that reads injury and line-up news, and a market module that tracks price and money flow.
• Market comparison. Where the model’s probability exceeds the market’s by more than the threshold, the pick is flagged as positive expected value; otherwise the engine returns a SKIP verdict.
• Public, tamper-evident record. Every prediction is hashed (SHA-256, the NIST FIPS 180-4 standard) before publication; hit rate and ROI are published by confidence tier, sport and competition, and nothing is ever deleted.
• Product surface. Web app with subscriptions and a token model, free daily picks, statistics pages, a B2B API and a white-label option, in four languages.

The table contrasts the usual tipster service with the engine behind Erudil. A tipster publishes a pick and a story; the engine publishes a probability, the market price it was compared with, and a verdict that can be SKIP. The difference is not accuracy on a good week — it is whether the record can be audited on a bad one.

Six data providers, a news-reading NLP module and an odds feed converge on the ensemble; every event that can move a probability — a confirmed line-up, an injury, a 3% odds move — triggers a recalculation.

The model's calibration is tracked publicly. Hit rate is reported per confidence tier so that a high-confidence pick means something; the Brier score and closing-line value are the two numbers we watch in production — a model that beats the closing line is a model with a real edge.

The page that makes the product honest: every pick is hashed before it is shown, graded automatically after the match, and counted in the tier statistics. A user can verify that a pick was not edited after the fact.

Modelling and simulation in Python (NumPy, PyTorch for the ensemble weighting); data pipelines pulling fixtures, statistics and odds from six providers with re-computation on news events; PostgreSQL and Redis; a Laravel back end with a React front end; containerised deployment on AWS. Model calibration (Brier score, closing-line value) is monitored in production and a drift alert opens a retraining ticket.

This approach is for any team whose decisions depend on “how sure are we?” as much as on the answer: demand forecasting in retail and logistics, dynamic pricing, credit and fraud scoring, capacity planning. Erudil is the kind of forecasting system we build for clients — a calibrated probability model that is honest about uncertainty, connected to live data, and shipped with monitoring rather than a slide deck. Deloitte’s technology and sports outlooks (2024) expect data and AI to keep reshaping how sports organisations and media price and present outcomes; the same pressure applies in every industry that sells against a moving market.