Model and selection: how a probability becomes an opportunity

An opportunity on DumbleScore comes about in two steps. First, the model calculates how likely each outcome of a match is. Then the selection checks whether the price is worth it and chooses at most one opportunity per match. The model answers the question “How likely is what?”, the selection “Are the odds better than they should be by our calculation?”. We deliberately keep the model and the selection separate: that way each one can be checked and improved on its own.

Step 1: the model calculates

The model calculates the expected goals (xG) for both teams. In league and European competitions, since October 13, 2026, the final analysis takes them from the ratings of the eleven players who actually start. Previews have used the same player model since October 15, on our forecast of the starting XI. Cup matches, internationals and incomplete lineups are calculated with team values. All probabilities follow from the expected goals: win, draw, double chance, over/under, “both teams score” and the most likely scores. The AI does not calculate a single number; it only summarises the most important factors in your language.

Whether the calculation is good does not show in a single match. Over many matches, we measure how close the probabilities come to the actual goals after 90 minutes and to the outcome (win, draw, loss). A single surprising result says nothing about it.

Step 2: the selection decides

The selection compares our probabilities with the market. For the fair market price, we take the odds from Pinnacle, one of the sharpest markets (or from Bet365 if they are missing), and remove the margin. Measured against this fair price, our edge must be large enough, and the higher the odds, the larger it has to be. On top of that come strict filters: no very low odds, no extreme values, and at most one opportunity per match. That is why the answer for many matches is “No edge”.

The selection has its own yardstick, too. You can tell a good selection by the odds at the time of the analysis being better than the fair odds shortly before kick-off. The market then moves in our direction. That shows up faster and more reliably than the profit or loss of a single week.

New versions run alongside for comparison

How do we test a change? Not on the performance record you see. With every final analysis, several variants run in the background, both calculation variants and selection variants: on the same matches, at the same moment, with the same odds. Their results are stored and later measured against the actual results. They are not visible, and they do not flow into the performance record. We call this shadow mode. New models and new selection rules first run in the shadow like this; only then do we decide whether they go live.

What this means for you

• What you see is always the version that is live. Experiments stay invisible.
• Changes come with a date. For the performance record, this means: until October 13, 2026 with the previous team model, since then with the player model in league and European competitions, all in the same performance record.
• Nothing is recalculated retroactively. Every opportunity stays as it is, with the model that produced it: all counted opportunities from final analyses before the switch come from the previous team calculation. Nothing is changed after kick-off.

You will find every opportunity and its outcome in the performance record. Past results are no guarantee of future results.