foxintheboxd

Predictions

How scoring works

Most prediction leagues pay a flat 3 points for an exact score and 1 for the result. That rewards backing favourites. This one prices every call by how surprising it was.

The idea in one line

Your payout is proportional to the information your call carried. In plain terms: the less likely your scoreline was, the more calling it correctly is worth — far more than someone who called the obvious one.

points = k × −log₂(p)

where p is the probability the model assigned to what you predicted. That quantity, −log₂(p), is the surprisal of the event, measured in bits. The model runs server-side and its probabilities are never published: what you see is the payout, not the odds.

Where the probabilities come from

  1. 1

    Attack and defence strengths per club

    Every finished result in the database is read to work out how many goals each club scores and concedes relative to the league average. Clubs with few games played are shrunk toward the average, so three fixtures cannot produce a wild rating.

  2. 2

    Expected goals for the fixture

    Home strength times away weakness, adjusted for the home-advantage effect measured across the whole dataset. This gives two numbers — the expected goals for each side.

  3. 3

    A full scoreline grid

    Those two expectations produce a probability for every scoreline from 0-0 to 8-8. The grid gets a Dixon–Coles low-score correction, because plain independent Poisson reliably under-counts 0-0, 1-0, 0-1 and 1-1 — draws happen more often than naive maths expects.

  4. 4

    Your price, frozen

    When you submit, the probability of your exact scoreline and of your result are both stored on the prediction. Later results change the model but never your payout — the price you were quoted is the price you get.

A worked example

Priced live, right now, for Man City v Brighton. Same fixture, different calls — note how the payout climbs as the scoreline gets bolder.

Example payouts for different scorelines in the same fixture
Your callExact paysResult only
1071
2171
1176
3171
23138
33146
43151

Read the top and bottom rows together: the chalk scoreline pays 7, the wild one pays 15. Same fixture, same effort, very different reward for being right.

The rules, plainly

Exact scoreline

3–40 pts

Scaled by surprisal

Right result only

1–12 pts

Also scaled

Wrong result

0 pts

Never negative

  • An exact call always pays strictly more than the result alone, even on the rare fixture where the maths would say otherwise.
  • Payouts are capped at 40 points so a 7–6 lottery ticket cannot outweigh a season of careful calls.
  • A fixture opens for predictions a fortnight before kickoff and closes an hour before it. You can change or withdraw a call at any point in between; changing it re-prices against the model as it stands at that moment.
How prediction scoring works · foxintheboxd