Why Our AI Picks a Winner: Explainable Predictions in Plain Language

A probability without a reason is hard to trust. When our model says Djokovic 57%, the obvious next question is why? Since June 2026 every match analysis on Predixsport answers it with a card called "Why the Model Favors…", which lists the factors behind the pick in win-probability points. This article explains where those numbers come from and how to read them.

The problem with black boxes

Our winner models are neural networks that read hundreds of inputs per match (about 700 in football, more than 1,100 in tennis and the NBA): serve and return rates by surface, the projected starting five, schedule load, chance creation from shot data, and so on. Nobody, us included, can look at that many inputs and say which ones decided a given prediction. So instead of writing a plausible story, we measure it on the model.

How it works

1. Measure the pull of every input. For each match we run Expected Gradients, a standard attribution method, on the network. It measures how far each input moved the win probability compared with an average matchup. It is a calculation on the model, not a summary written afterwards.
2. Group the inputs into factors a fan would recognise. A single input such as serve_hold_rate_clay means little on its own and its value is noisy. Added up into factors like Surface fit, Recent form, Starting five, Attack or Chance creation (xG), the numbers become stable and readable.
3. Express everything in win-probability points. A factor worth +5.3 moved the prediction 5.3 percentage points toward that side. Together the factors add up, within about a point, to the distance between the final probability and a 50/50 match.
4. Show only what is stable. Each factor is computed several times on separate samples. If the runs disagree on its direction, it is marked unstable and never presented as a reason.

Reading a real card

The card from Dino Prizmic vs Novak Djokovic in Rome, where the model gave Djokovic 57%:

That reads like a tennis argument: Djokovic's class and clay record against a 20-year-old's youth and form, with class winning narrowly. Note that the card is not one-sided. It shows what favors the other player too. A card where every line points the same way only happens in a clear mismatch.

The "Home advantage" line

In the NBA and in football the host wins more often across the whole league, before anything is known about the two teams. We show that general edge as its own factor. So when a football card reads Home advantage +16 next to Attack +4.7 to the visitors, it means the home side leads because of the venue, while the matchup itself leans the other way. Tennis has no host, so the card starts from 50/50.

What this is and is not

• It is a measurement of what the production model did for this match. Before launch, every factor we display was checked on thousands of held-out matches against the raw statistic it describes, and factors that did not line up were not shipped. How well the probabilities themselves hold up is on our model transparency page.
• It is not editorial opinion. Nobody picks the factors match by match. The same method runs on every prediction.
• It can be wrong about the match. After an upset the card shows which beliefs failed, and a probability you can question is more useful than one you can only accept.

Where you will find it

The card appears on tennis, NBA, Euroleague, Champions League and football match analyses, right under the prediction header. The article text on those pages is written from the same factors. Developers and AI agents get the same data as a key_factors field in our API and through the Claude connector, so asking Claude "why does the model favor X?" returns the measured factors.

The fine print

• Older articles, published before June 2026, have no card. The method was not running yet.
• In football the factors compare the two sides. The draw is priced by the same model but shown in the header. The card answers "who has the edge if someone wins?"
• Football has a second tab that explains the over/under 2.5 goals prediction the same way: which attack or defence pushes the total up or down.
• Very small factors are hidden. Anything that moves the prediction by less than half a point stays off the card.

In short: the model's own arithmetic, grouped into factors a fan would use in an argument, checked before launch, and shown on every prediction, including the ones we get wrong.