match previews<\/a> as a projected result and score for the game; in the AI Predictor<\/b> pages as every pick for a competition laid out in one table with a running record on top; on the outright pages as title, top-five and relegation odds set against our model; and on club pages and the homepage. Wherever you see a Signal number it comes from the same place and is measured the same way.<\/p>\n\n\n\nA note on the two names, because both are ours. \u201cSquawka Signal\u201d is the model \u2014 how it works. \u201cAI Predictor\u201d is the page where you can read all of its picks for a competition in one place.<\/b> Same engine, two names: one for the method, one for the surface. This page documents the model, so it covers every surface, not just the AI Predictor table.<\/p>\n\n\n\nHow Signal works<\/h2>\n\n\n\n
Honest framing first. Signal is not a deep neural network with a billion parameters. It is a team-strength model you can interrogate: it rates every club’s attack and defense straight from results \u2014 two full seasons of them \u2014 using a Poisson maximum-likelihood fit, with a home-field adjustment and more weight on recent games. Those two ratings give an expected-goals figure for each side, which the model turns into a full probability distribution across plausible scorelines. The most likely scoreline is the pick; the win, draw and loss probabilities come off the same distribution.<\/p>\n\n\n\n
Two refinements keep it honest. A thin sample is treated as a thin sample<\/b>: a newly promoted side with no top-flight history is pulled toward a league-average prior and the little data it has is discounted, so a good three-game run is not mistaken for a new baseline (early-season calls on new teams move more than settled ones, by design). And a low-score correction<\/b> nudges the tight scorelines \u2014 0-0, 1-0, 1-1 \u2014 toward what football actually produces, because independent scoring math slightly under-counts draws.<\/p>\n\n\n\n
From that one distribution Signal reads more than the result. The projected scoreline<\/b> is the most likely score within the outcome it favors. Both teams to score<\/b> and over\/under 2.5 goals<\/b> fall out of the same goal distribution. And for anytime goalscorer<\/b>, the team’s expected goals are shared across its scorers by their scoring rate, so the players the model likes are the players it expects to be involved \u2014 availability and lineups are then applied editorially before anything is published.<\/p>\n\n\n\nSignal against the market<\/h2>\n\n\n\n
Every Signal number is shown against the market price, because a prediction is only really interesting where it disagrees with the money. Early in a season a young model has little to work with, so Signal leans on the market<\/b> \u2014 which has already priced the summer’s managers and signings \u2014 and hands over to its own read as real games bank, reaching the pure model by around the tenth matchweek. So an early-season number is mostly the market with a little model in it; a mid-season number is mostly the model. This is deliberate: it stops a still-forming model from shouting a bold call off three games of data.<\/p>\n\n\n\n
That gives two distinct readings that we keep separate. The prediction<\/b> is the most likely outcome \u2014 who we think wins. The value<\/b> is where Signal most disagrees with the market \u2014 where the edge is, if there is one. They are not the same thing: a side can be the value (the market underrates them) without being the prediction (they are still not the likeliest winner). We label them separately so one never masquerades as the other.<\/p>\n\n\n\nHow we measure Signal<\/h2>\n\n\n\n
A predictor that does not track its own performance is not a predictor; it is marketing. The running record below is the model’s report card, rebuilt daily as matches settle, and it is measured out of sample<\/b> \u2014 every pick scored as if locked in before kickoff, with the ratings fit only on matches already played. Three numbers matter.<\/p>\n\n\n\n
Result hits.<\/b> Did Signal call the correct outcome \u2014 home win, draw or away win? The loosest measure, but the one that maps directly to a match-result bet.<\/p>\n\n\n\n
Exact-score hits.<\/b> Did it call the exact scoreline? Correct-score is notoriously hard \u2014 top tipsters hit under 10 percent, and any system claiming a big edge here should be treated with suspicion. We expect Signal to land in the 8 to 12 percent band over a season.<\/p>\n\n\n\n
Signal against the market.<\/b> For every fixture we freeze what Signal said and what the market was paying at that moment, and grade the disagreement, not just the call. That is the number that tells you whether an edge is real or noise. It is the newest of the three and its ledger is still filling, so we publish it only once enough calls have settled to mean something.<\/p>\n\n\n\n
Underneath all three sits calibration<\/b>, the one that matters most: do the things Signal calls 60 percent actually happen about 60 percent of the time? That comparison \u2014 predicted probability against realized frequency \u2014 is in the track record below, measured out of sample. Where the model’s confidence is honest we show it, and where it drifts we show that too.<\/p>\n\n\n\nWhat Signal does not do<\/h2>\n\n\n\n
It does not know about injuries or suspensions.<\/b> The model reads strength from results; it has no feed from the medical room. If a key forward is a late withdrawal, the ratings still include his past games. That is why every published pick is paired with the team news in the match preview, and why a named goalscorer is checked for availability before it runs.<\/p>\n\n\n\n
It does not model tactical surprises.<\/b> A side that changes shape for one big game will not be caught until that pattern repeats in the data. The numbers lag the soccer at exactly those moments.<\/p>\n\n\n\n
It is not a substitute for the preview.<\/b> A Signal number gives you a pick, a price and an edge. The full preview gives you team news, probable lineups, tactical context and set-piece angles. Use them together.<\/p>\n\n\n\n
Coverage varies by competition.<\/b> Where our data feed carries a league, the model rates its clubs from results; where a side’s league is not covered, a promoted or cross-border newcomer leans on a strength prior until it has played enough. The surfaces flag when a read is data-rich and when it is thin.<\/p>\n\n\n\n
It is not a guarantee.<\/b> A season across four competitions will throw up results no model saw coming. Signal’s job is to be honest about its inputs and accountable for its outputs \u2014 not to promise a winner.<\/p>\n\n\n\nTrack record<\/h2>\n\n\n
\n\n