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 natural next question is the one every analyst gets asked: why? As of June 2026, every match analysis on Predixsport answers it — with a card called "Why the Model Favors…" that lists the factors behind the pick, measured 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 weigh more than a thousand signals per match — serve and return rates by surface, starting-five production, schedule congestion, chance creation from shot-level data, and hundreds more. That depth is why the predictions are calibrated. It's also why nobody — including us — can eyeball the inputs and say what drove a specific prediction. Most prediction sites solve this with prose that sounds like an explanation but is really a writer's guess. We wanted the real thing: the model's own arithmetic, made readable.

How It Works, Honestly Stated

1. Measure every signal's pull. For each match, we use a technique called gradient-based attribution to measure how much every single input moved the win probability, compared against an average matchup. This is a calculation on the network itself — not a summary written after the fact.
2. Group a thousand signals into human factors. Individual signals are noisy and unreadable (serve_hold_rate_clay means little on its own). Summed into concepts — Surface fit, Recent form, Starting five, Attack, Chance creation (xG) — the attribution becomes both stable and meaningful.
3. Express everything in win-probability points. A factor worth +5.3 moved the prediction 5.3 percentage points toward that side. The factors approximately add up to the distance between the final probability and an even matchup — nothing is hidden in the gaps.
4. Publish only what survives a stability test. Each factor is estimated several times independently; if the runs disagree on direction, the factor is flagged unstable and never narrated as decisive.

Reading a Real Card

Here is the card from a real match — Dino Prizmic vs Novak Djokovic in Rome, where the model priced Djokovic at 57%:

That's a real tennis argument, not a vibe: Djokovic's class and clay record against a 20-year-old's youth and momentum — and class wins, but only just. Notice the card is not a chorus. A credible explanation concedes what favors the other side; a card where every line points the same way should (and does) only happen in genuine blowouts.

The "Home Advantage" Line

In NBA and football the host wins more often league-wide — before anything is known about the two teams. We show that baseline edge as its own factor, so when you read Home advantage +16 next to Attack +4.7 to the visitors, the card is telling you something precise: the home side leads because of the venue, while the matchup-specific numbers actually lean the other way. In tennis there is no host, so the card starts from a true 50/50.

What This Is — and Is Not

• It is a measurement of what the production model did for this match, validated on thousands of held-out games before launch (factor directions agree with the obvious stat gaps 70–99% of the time, depending on the factor — and we publish the method, not just the wins).
• It is not editorial opinion. No one chooses the factors per match; the same engine runs on every prediction.
• It can be honestly wrong. When an upset happens, the card shows which beliefs failed. We consider that a feature: a probability you can interrogate beats one you can only accept.

Where You'll Find It

The card appears on every new tennis, NBA and football match analysis, right under the prediction header. The article text on those pages is grounded in the same factors, so the chart and the prose can never tell different stories. Developers and AI agents get the same data as a structured key_factors field in our API and through the Claude connector — ask Claude "why does the model favor X?" and it answers from the attribution, not from guesswork.

The Fine Print (Kept Short)

• Older articles (published before June 2026) don't have the card — the engine wasn't running when they were written.
• Football factors compare the two sides. The draw is priced by the same model but shown separately in the header — the card answers "who has the edge if there's a winner?"
• Football gets a second tab. The same engine also explains the under/over 2.5 goals call: which team's attack or defense pushes the total up or down, in the same probability points.
• Tiny factors are hidden. Anything moving the prediction less than half a point stays off the card; it's noise to a reader even when it's signal to the model.

That's the system: the model's own arithmetic, grouped into factors a tennis or basketball fan would actually use in an argument, validated before we shipped it, and published on every prediction — including the ones we get wrong. That last part is the point.