Daro

How our estimates work

Daro shows its own numbers before every match, says how they were made, and checks them against what happened. It never reads betting odds to make them.

How does Daro calculate its predictions?

Each statistic (goals, shots on target, corners, cards) has its own Poisson team-rating model, of the Maher and Dixon–Coles family. Every team has an attack and a defence rating, and the home side gets a home advantage. Goal ratings learn from both goals scored and expected goals (xG), so a lucky or unlucky result counts for less.

The ratings are fitted on every played match Daro holds before the day of kick-off, across all 59 competitions at once. Cup ties between leagues are what tell one league's strength from another's. Older matches count less: a match a year old weighs half as much.

A team seen in few matches is pulled towards the level of the league it plays in, so a handful of results never makes it look like a giant. A side needs at least three matches with a statistic, or we show no estimate rather than a guess.

The result (home, draw, away) and both teams scoring come from the goal ratings; fair odds are simply one divided by our probability. Total shots use each side's last ten matches, which tested better for that statistic.

Was it tested before release?

Before the model went live it was backtested week by week: fitted only on matches before each week, then scored on that week's matches. Over 9,190 matches from August 2025 to October 2026, its result forecasts scored a Brier score of 0.608, against 0.642 for simply guessing each league's average. Lower is better.

It is calibrated: when it gave the home side 45%, home sides won 46% of the time; at 64%, they won 65%.

Any change to the method has to beat the current version on the same test before it ships.

How does it compare with the bookmakers?

After we make our numbers, we show the bookmakers' prices next to them, from 10 bookmakers, with their margin removed so they read as probabilities. Where we and the market disagree is where questions are worth asking.

The market sees things our model doesn't: line-ups, injuries, motivation. Its closing prices are a hard benchmark, and our track record shows how we compare, honestly.

Does Daro learn from its mistakes?

Every finished match is scored against the result and against the market's closing prices. Each night an AI reads the biggest matches (events, statistics, line-ups) and writes down what our estimate missed. Each week the patterns become hypotheses to test, and only a tested improvement changes the model.

Every Monday the model also re-tunes its own settings (how fast old matches fade, how far a team is pulled to its league, how much goals lean on expected goals) and keeps a change only if it predicts better on recent weeks it was not tuned on.

What doesn't Daro know?

Injuries, suspensions and line-ups are shown on each match page and given to the chat, but the model itself doesn't use them yet.

Corners are close to a coin flip for every method at the usual 9.5 line. The Poisson model also understates how much corners and cards vary, so extreme lines look more certain than they are.

Teams from leagues we don't follow are rated from fewer matches, mostly cup ties. Their pages say so.

What can I ask Daro?

You can ask Daro about any match, a competition or a whole day, and ask it to build a parlay with the chance it all lands. On a match page it answers from that match's evidence (our estimates, the prices, form, statistics, absences), cites it, and says when something is unknown rather than guessing. It talks about football only, and it never tells anyone to bet.