How the QUVIX model works

Last updated: 11 August 2026

QUVIX is a free prediction game: a statistical model publishes football probabilities before kick-off and stakes a virtual daily pool on one pick a day, in public, against everyone. Every result — including losses — is published. The model is designed to be beatable, and we grade it on closing line value, not cherry-picked wins.

1. The probability model

Match probabilities come from a Dixon-Coles model (a bivariate Poisson variant that corrects the low-score dependence classic Poisson gets wrong), fitted per league with exponential time decay (ξ = 0.25 per year, so a match from a year ago carries ~78% of the weight of yesterday's). Each league is trained in its own cluster — a Swedish Allsvenskan match is never priced with Premier League priors.

The model outputs a full scoreline distribution per fixture, which collapses into the three markets shown on the site: match result (1X2), over/under, and both-teams-to-score.

2. The market baseline

Odds shown on QUVIX are a consensus median across multiple bookmakers, captured hourly. For analysis we remove the bookmaker margin ("de-vig"): implied probabilities are normalised to sum to 100%, giving the market's true consensus price. The edge figures on the site are simply model probability minus de-vigged market probability — where the model and the market disagree.

3. The AI player's betting rules

The model doesn't just publish probabilities — it plays the same game as every user, with the same daily pool of 1000 virtual credits. Its rules are fixed, deterministic and public:

Rule	Value
Bets per day	Exactly one, full pool (all-in)
Candidate set	Every future fixture × every published market that day
Selection score	edge = p_max − league structural prior
Threshold	edge ≥ 5%; if nothing qualifies, fall back to the highest-probability 1X2 pick
Look-ahead	None — bets lock at pre-match odds, never re-priced, never after kick-off

The "league structural prior" matters: home teams win more everywhere, but not equally — the baseline 1X2 favourite probability is ~43.2% in the Premier League and ~47.8% in North American leagues. Subtracting it means the model bets genuine conviction, not "home teams exist".

4. How we grade ourselves: closing line value

Single-match results are mostly noise. The industry-standard test of whether a prediction had real information is closing line value (CLV): did the price you took beat the final pre-kick-off consensus? We snapshot odds hourly and mark the last capture before kick-off as the closing line — but only count it if that snapshot is within 90 minutes of kick-off, otherwise the bet is honestly reported as "no closing data" rather than silently graded against a stale line.

Every settled bet on QUVIX — the model's and every player's — gets a CLV figure by the same formula. The leaderboard has a CLV column. Nobody is graded on a different curve.

5. Radical transparency

The model's full bet history is published at /data/model-track-record.json — every pick, every price, every loss.
Its current record is honestly mediocre (negative ROI at the time of writing). We publish it anyway. The point of QUVIX is that the model is a fair, transparent opponent — not an infallible tipster.
The model's open pick is hidden until after kick-off (so you can't copy it), but everything settled is public forever.
Users and the model see identical data: the same probabilities, the same consensus odds, the same daily pool rules.

6. Known limitations

The model prices from historical results and market data only — it does not read injury news, rotation rumours, or weather.
Newly covered leagues start with structural priors and limited within-season data; early-season estimates are wider than mid-season ones.
Consensus odds are snapshots, not live prices; new leagues may have sparse early markets.

7. The plumbing (for the curious)

Fixtures, results and odds come from a commercial data API, captured by a scheduled pipeline; probabilities are recomputed and republished daily. The whole stack runs on Cloudflare Workers + D1 with a static front end — no servers to speak of. Bets are settled by an independent job that never sees the model's reasoning, and the model's confidence values are deliberately excluded from every public API response.

Think you can beat it? Play QUVIX — free, 1000 virtual credits a day. Credits are points, not money: QUVIX is not a gambling operator and nothing here is betting advice. Terms · Privacy