Football predictions, and how they are actually produced

The robot starts from market activity rather than from a fixture list. It looks for football matches where the 1X2 market is genuinely traded, filters out anything without a clean competition, kick-off time and team pairing, and only then considers whether the price is doing something interesting.

Each price observation is stored as a new timestamped snapshot, never overwriting the last. A drop is therefore a measured change between two real observations. Because thin markets move on a single large ticket, a minimum traded-volume requirement is applied before a drop is treated as meaningful at all.

A candidate becomes a selection only when it clears every gate: sufficient market volume, a well-supported price movement, a fresh price snapshot, acceptable data quality, and a positive edge measured against the margin-free fair probability. Failing any one gate ends the candidate — the gates are not traded off against each other.

A Poisson goal model, an Elo rating and a recent-form component are blended into a single probability. Elo is computed only from matches that had already finished at prediction time, so the model cannot benefit from information that did not exist yet. Implausibly large edges are treated as a sign of a data problem rather than as an opportunity.

Publishing a prediction for every fixture would be easy and worthless. If the market price is already fair, there is nothing to say. Empty days are a feature of a system with real thresholds, and this site does not create match pages simply to have a URL for every fixture.

A probability built on stale team data or a stale price is a confident-looking guess. Every candidate carries a data-quality assessment and a snapshot age, and both can veto a selection outright — the robot skips the match rather than bet on a number it no longer trusts.