About

What it does

• Predicts outcomes for every upcoming fixture across 8 leagues.
• Picks a side with a confidence grade (Strong / Lean / Coin-flip).
• Suggests a "spicy" pick — longer odds the model rates higher than the bookies.
• Tracks itself honestly — every pick is graded against the actual result.

What is "Elle's Pick"?

Elle's Pick is a single recommended outcome per fixture (Home Win / Draw / Away Win), graded by confidence. It's derived from the model's probabilities, but not always the same as "the most likely outcome." The current production model is v0.6 — see the model write-up for the version history.

The grading rule:

• Strong — top side ≥ 55% probability. High confidence pick.
• Lean — top side 45-55%. Lower confidence, still our pick.
• Coin flip — top side < 45%. The fixture is genuinely close.

Why Elle's Pick may not match the highest probability

There are three reasons this happens:

• Coin-flip → DRAW preference. When the top side beats DRAW by less than 5 percentage points (e.g. Home 38% / Draw 34% / Away 28%), we pick DRAW. Reason: a near-tie between a side and the draw is statistically a draw signal — the model is saying "I can't separate these two outcomes."
• Draw post-processor (v0.4). In La Liga, Serie A, Ligue 1 and English League One, on coin-flip fixtures, we apply a league-cultural rule: if both teams have been blowing leads recently and sit in the bottom half of the table, we override the model's pick to DRAW. This is based on a measurable historical pattern (those leagues genuinely convert chaos into draws more than the others) — it would be statistical noise to apply it everywhere, so it's targeted.
• Modal draw. If DRAW has the single highest probability across the three outcomes, we pick DRAW (with appropriate confidence grade). Standard but worth flagging.

We surface the raw probabilities on every fixture so you can see the underlying numbers. The Pick is our recommendation; the probabilities are the truth the model produced.

What is the "spicy" pick?

A separate, deliberately-contrarian alternative on each fixture: an outcome with longer odds where our model thinks the bookmakers have it wrong. It's not Elle's primary pick — it's a "if you fancy a flutter" suggestion based on price-vs-probability gap.

How accurate is it?

Bookmaker accuracy on top-five European leagues hovers around 55%. We're not there yet. We're transparent about every miss — see the history page.

Performance over time, per model version

Every time we ship a better model, the older one stops growing — frozen at the day it was replaced — and the new one takes over from there. Newer versions also rescore the older fixtures once, so you can compare like-for-like and see whether the upgrade actually helped.

Switch between % right (raw accuracy), Right per week (how many calls landed each matchweek), or Confidence (how sure we were when the answer turned out to be the one we picked). The line is jagged because some weeks are calm and some are bedlam — that's real football, not bad maths.

The model

Three-stage pipeline: a Dixon-Coles statistical base, a gradient-boosted stacker that corrects per-fixture using ~77 contextual features, and a league-targeted draw post-processor for coin-flip games in defensive leagues.

Read the full model write-up →

Where the data comes from

API-Football is the single primary source: fixtures, finished results, starting lineups, injuries, xG, shots, possession, bookmaker odds, top-scorer rankings — everything driving both the model features and the displayed stats.

Picks regenerate every matchday once fresh fixture data lands. Results score automatically once games finish. The site rebuilds on every data refresh — so the numbers you see are always the latest available.

Who built it

Elle Bush. Independent.