The pricing model

A board we generated, right now

Live from the feed, no key needed. Each row carries the true probability, the price before margin, and the price we would offer.

Navarro/Eugenio vs Caldera/Eugenio

Anchored on: ranking:fip · Margin: 12% · Generated: 2026-10-21 14:20 UTC

What the API returns

The same rows, as your integration would receive them.

{
  "data": [
    {
      "matchId": "mch_6d345ebdcfef",
      "market": "total_sets",
      "selection": "2",
      "probability": 0.677268,
      "fairPrice": 1.477,
      "price": 1.362,
      "anchor": "ranking:fip",
      "margin": 0.12
    },
    {
      "matchId": "mch_6d345ebdcfef",
      "market": "total_sets",
      "selection": "3",
      "probability": 0.322732,
      "fairPrice": 3.099,
      "price": 2.593,
      "anchor": "ranking:fip",
      "margin": 0.12
    }
  ],
  "meta": {
    "count": 36,
    "matches": 1,
    "anchors": [
      "ranking:fip"
    ]
  }
}

Why padel can be priced from first principles

Padel inherits tennis scoring: points to 15/30/40, games to six, sets to two. That makes a match a Markov chain over points, and given how often each pair wins a point on its own serve, every market follows analytically — match winner, set betting, total games, handicaps. No bookmaker’s opinion is copied at any stage.

Golden point is not a detail

Premier Padel and WPT play a single deciding point at deuce instead of the advantage series. That measurably helps the returner — an advantage deuce lets the stronger pair grind out its edge, one point does not — so it compresses every price on the board. A model that priced padel with tennis maths would sell the favourite too cheaply.

What happened when we tested it

The model is tested against every padel match we hold with a winner and two complete sets — 4,928 of them, current and historical, back to 2022. For each it was given the first set’s score and nothing else, then asked to predict the rest of the match, so everything it is scored on is out of sample. Each baseline was fitted on those same matches, which makes it the strongest version of "the model adds nothing". The test is re-run as the archive grows and as the model changes; these are the current figures, not a one-off.

lower is better

The middle column is what the model does with a match it has never seen. Totals are scored the way the market works — for each of five lines, did it go over, and how sure were we — against a baseline that already knows how often each line goes over. Fifteen per cent better on the market padel books leave empty is the figure that matters, because that is the one a category launch has to fill.

A result nobody fitted

Padel resolves a deuce differently from tennis, and the model is sensitive to it: run with padel’s scoring it predicts better than with the tennis series on all three measures — 0.3948 against 0.3961, 0.5336 against 0.5355, 0.5740 against 0.5751. Nothing was tuned to produce that. It falls out of the chain being built the way padel is actually played, which is the strongest evidence here that the structure is right.

What we would claim, and what we would not

This is a pricing input. It produces markets nobody else prices, with the probability exposed so the operator applies its own margin and owns the price that reaches a customer. On that job the evidence is direct: out of sample across 4,928 matches, better than a fitted baseline on all three measures, and calibrated — of 2,200 matches called at 87.6%, 87.1% came in.

What it has not been tested for is whether these prices would have made or lost money as a book. That is a different question, and a well-calibrated model can still be picked off by a market that is faster or knows something it does not. We have not run that test and would not claim it.

Two blind spots worth naming. In production the match-winner number is anchored on the market, so what we add sits in the derived markets — and those we cannot compare against anyone, because nobody prices them. And the baseline here is a long-run average, not a sharp book’s closing line; beating the first is not beating the second. Both are consequences of pricing a category that is empty, which is the same reason the product exists.

What this does not show

• Nearly five thousand matches is a real sample, but padel’s recorded history is short — this reaches back to 2022 because there is not much before it. These numbers will keep moving.
• One number in the chain is fitted rather than derived: how much the gap between two pairs closes with each completed set. Everything else follows from the scoring. Worth saying, because a fitted parameter can flatter its own sample — though this one was chosen where the model predicts best rather than where the figures look best, and the two turned out to be the same place.
• Live boards are repriced from the score as it moves and withdrawn 45 seconds after they are computed. Point-by-point pricing is not built: the score reaches us game by game, so a board is only as current as the last completed game.
• Where no bookmaker has priced a match, the model can anchor on FIP rankings instead. That is the weaker signal, is quoted wider, and every row says which anchor produced it.

Want it against your own coverage?

The full feed is a Pro key. Every selection ships with the probability and the pre-margin price alongside ours, so you can load your own margin.

Request API access →API docs →What you are buying →