Transparent AI football forecasting — built for readers, not tipsters.

Football prediction content online is often a single headline number with no provenance: which model said what, how confident it was, or whether that source has been right lately. TuringStats exists to fix that gap by treating forecasts as comparative evidence, not marketing copy.

We built the site for fans, analysts, and curious readers who want to understand why a match leans home, draw, or away before kickoff — and who want a durable record of how each AI model performs when the ball actually stops rolling.

Every page is designed to add context you cannot get from a one-line “pick of the day.” Our core product surfaces include:

• Live forecasts hub — upcoming fixtures across major leagues with aggregated 1X2 vote shares, expected goals, most common scorelines, and per-model breakdowns refreshed before kickoff.
• Public AI leaderboard — rolling accuracy, Brier-style calibration scores, and a visual “last 10 matches” form strip for each tracked model so performance is inspectable, not claimed.
• Match intelligence pages — deep-dive hubs per fixture (form, injuries, standings context, predicted score distributions, SEO-friendly explainers, and optional crowd “who wins?” voting before kickoff).
• Historical archive — finished matches with graded outcomes so readers can audit past calls.
• Journal — long-form articles on methodology, model behavior, and league narratives written for humans, not scrapers.
• iOS app — the same forecast philosophy in a focused mobile experience with alerts when new predictions land (see App Store listing).

We run a repeatable pipeline so models are compared fairly on the same fixture snapshot:

1. Fixture ingestion — schedules and metadata for major competitions are loaded from licensed football data providers and stored in our database.
2. Structured prompts — each model receives the same contextual bundle (recent form, scheduling, expected-goals framing, head-to-head notes where available) via a controlled prompt template.
3. Parallel inference — multiple frontier models (typically ten active profiles routed through OpenRouter) independently return outcome, scoreline, and confidence.
4. Aggregation — we compute consensus percentages, highlight dissent, and surface confidence bands on the match page.
5. Post-match grading — after full-time, picks are scored against the official result; leaderboard metrics update on a scheduled cadence.

For a plain-language walkthrough, see How scoring works on the home page. Technical curiosity is welcome — we publish enough detail for skeptical readers to understand the limits as well as the strengths.

TuringStats content is produced with the following rules:

• No paid picks. We do not accept payment to bias a forecast, bury dissent, or promote a sportsbook.
• Label uncertainty. Confidence and model disagreement are shown alongside headline reads — low agreement is a feature, not hidden noise.
• Separate news from models. Journal articles are written and attributed editorially; AI outputs are labeled as model-generated inference.
• Correct when wrong. Factual errors in articles or fixture data should be reported via contact; we aim to fix material mistakes promptly.
• Responsible framing. Copy repeatedly states that outputs are informational and not betting, financial, or legal advice (see Terms of Use).

• We are not a bookmaker, exchange, or gambling operator.
• We are not a tipster Telegram channel — there is no VIP tier for “sure wins.”
• We do not guarantee profits; past model performance does not predict future results.
• We do not provide personalized investment or wagering recommendations.

If you choose to bet, do so only where it is legal, within your means, and with operators licensed in your jurisdiction. Seek help from recognized responsible-gambling organizations if gambling stops being entertainment.

The stack is a Laravel web application with a MySQL datastore, background workers for ingestion and prediction jobs, and a Vite-powered front end. Third-party services we rely on include:

• Football fixture and enrichment data (e.g. API-Football) for schedules, standings, and match context.
• OpenRouter (and underlying model providers) for multi-model inference.
• Google Analytics / Google Ads tags for audience measurement and advertising (see Privacy Policy).

We cache expensive API reads to keep pages fast; cached snapshots are refreshed on documented schedules, not indefinitely frozen.