⚡

Pitch Engine Data Center

Ten models · match analytics

中文
Ten models online
0 matches

About Pitch Engine Data Center

Ten models, ten conclusions, one match. Pitch Engine Data Center puts them side by side so you see not just a number, but how much that number is worth trusting.

Every fixture goes to six statistical models and four large language models at the same time, each deriving win/draw/loss probabilities, the most likely scoreline and a goal distribution on its own. Statistical models converge by construction; the LLMs frequently argue — same two inputs, and DeepSeek says 2-1 while GLM says 1-1. That disagreement is itself the signal.

Statistical models: Poisson · Bayesian blending · Elo rating · Dixon-Coles low-score correction · regression fit · weighted consensus

Large language models: DeepSeek · Zhipu GLM · MiniMax · Xiaomi MiMo

How the maths runs

Expected total goals are inverted into a scoring rate, then split between the sides using the strength gap. A Poisson model expands that into an 8x8 scoreline matrix — sum the matrix and you have win/draw/loss probabilities, the chance of each exact scoreline and which goal band is most likely. The six models each correct the strength gap differently: Dixon-Coles suppresses low scores, Elo uses rating difference, the Bayesian variant pulls in a prior, and a weighted average becomes the consensus.

The disagreement score

Take the standard deviation of the home-win probability across the six statistical models and map it onto 0-100. It does not answer "who wins" but "how much do the models agree". On high-scoring fixtures no single model is worth much — which is exactly what this site wants you to notice first.

Common questions

Where does the data come from?

Fixtures, pairings and kickoff times come from a public sports information service. The models use only two numbers from it: the home side's expected goal difference and the expected total goals. If the network fails, a local snapshot takes over and the page never goes blank.

Do the LLMs know the market?

No. The four models receive the same two lines of data, and the prompt explicitly forbids any mention of wagering, stakes or settlement. They make independent calls from limited information rather than restating a price.

Are these probabilities accurate?

They are deductions under stated assumptions, not confidence in a result. Football is highly random — upsets and heavy scorelines are normal. This site publishes the process, not a guarantee.

Does it cost anything?

No, and there is no sign-up. Add it to your phone's home screen and it opens full-screen, offline included, using the cached fixtures.

Disclaimer: This site provides statistical analysis of football fixtures for research and reference only; it constitutes no advice of any kind. It does not provide, participate in or point towards any wagering service, and contains no odds, pricing or settlement information. Football is inherently random — read the numbers accordingly.
Page views and click events are recorded by an analytics service; no personally identifying information is collected. Some content is generated by AI — see the privacy policy.
This is a technical demonstration of statistical modelling and is not intended for users under 18.