Method and measured results
Everything below comes from a walk-forward backtest on 71,396 matches from
18 European leagues, 2015–2026, in which the model only ever saw matches from days strictly before
each fixture.
The model does not beat the odds. On Over/Under 2.5, the model's Brier is 0.2447 against the market's 0.2415.
The market wins in all 18 leagues and in all 11 seasons. Every betting strategy based on the gap between the
model and the odds lost money, and a "value" filter made the result worse the stricter it was. The "most probable
event" on the match page is a calibrated estimate, not a betting recommendation; its cumulative result, including
at real odds, is public in the Record.
Since v11.4, where bookmakers price the market, "Both teams to score" is shown as 25% model and 75% the market without its margin, and the corners as the market's own probability. Tested week by week on the Record's matches (843, 19 October to 16 November 2026, with the weights chosen only from earlier weeks), the model alone stayed below the market on both. With no price, the model's figure stays. Estimates recorded before v11.4 keep their figures.
Since v7.3 the app calculates by default on the
current season, regular phase only. The figures below are for the tested configuration, "last two
seasons", which remains available as an option. The newer market "over 1.5 goals in the 1st half" was not part of
the backtest.
How well each market is predicted
skill in %, against the base rate
| Market | v6 | v7 | v7 anchored + calibrated |
| Over 2.5 goals |
1.13% | 2.20% |
3.33% |
| Over 1.5 goals |
0.51% | 1.58% |
2.30% |
| Over 10.5 corners |
0.72% | 1.60% |
1.66% |
| Over 9.5 corners |
1.13% | 1.54% |
1.62% |
| Over 8.5 corners |
1.46% | 1.70% |
1.64% |
| Goal in the 1st half |
0.30% | 0.74% |
1.13% |
| Goal in the 2nd half |
0.06% | 0.58% |
1.03% |
| Over 0.5 goals |
0.11% | 0.55% |
0.80% |
| Both teams to score |
-0.22% | 0.62% |
0.72% |
Reference: on Over 2.5 goals the market consensus has a skill of
3.39%. The anchored variant reaches 3.33%, but that is the consequence of using the price, not a merit of the
model. The model's own merit is the v7 column.
How it is built
Likelihood estimation
Attack and defence strengths come from a time-weighted Poisson
estimate with L2 regularisation and Newton steps, not from ratios with hand-picked constants.
Dixon-Coles correction
The dependence between low scores (0–0, 1–0, 1–1) is estimated
from the data. Without it, "both teams to score" came out worse than the historical average.
One distribution
All goal markets and 1X2 come out of the same score matrix, so
they cannot contradict each other.
Halves, separately
A model of its own for the 1st half and for the 2nd half, instead
of an artificially bounded heuristic weight.
Negative-binomial corners
With over-dispersion estimated from the data, instead of a normal
approximation with a fixed deviation.
Anchoring to the market
Where an Over 2.5 price exists, the goal total is scaled until it
matches the price, and the other markets are recalculated from it.
How we know the measurement is right
- Positive control. The same pipeline, but with the model allowed to see the
future: Brier drops from 0.2464 to 0.2344 and the model moves ahead of the market. So the pipeline really
blocks look-ahead, and, conversely, any backtest in which the model beats the market leaks information from the future.
- External anchor. The market's hit rate on 1X2 in the Premier League comes
out at 54.83%, against 55.13% in an independently published benchmark; the market's Brier comes out at 0.1896
against 0.1901.
- The model on the site is the one tested. The functions are extracted from this
page and compared with the offline implementation: maximum difference 3.3·10⁻¹⁵.
- Naive reference. Averages over the last 20 matches + Poisson give a Brier of
0.2573. The model gives 0.2447. It is clearly better than the naive approach; the market is simply very good.
What it cannot say
There is no odds archive for corners or for a goal in the 1st half, neither in public data sets
nor at the API used here, which keeps odds only a few days back. On these markets the model cannot be compared
with the market retroactively. The only way to find out is daily collection, from now on, which is exactly
what the
Record page does.
Terms
- Brier
- The mean of the squared differences between the announced probability and what happened
(0 or 1). Lower is better. It punishes both the mistake and the unjustified confidence.
- Skill
- By what percentage the Brier is reduced compared with simply announcing the historical rate
of that market. 0% means the model adds nothing over the historical average.
- Bookmaker's margin
- How far the sum of the implied probabilities in the odds exceeds 100%. At a 5.91% margin, an
average bettor loses that money whatever they choose.
- Calibration
- If you take every case where the model said 70% and it happened in 70% of them, the model is
calibrated. The figure on the screen then means exactly what it says.