How it works

From raw match data to a calculated probability

Every forecast on the platform goes through the same pipeline. Here is what happens at each stage, what the output means and what it can't tell you.

The pipeline

Five stages, run for every fixture across every competition we cover.

  1. 1

    Collect and validate

    For each fixture we pull recent form, expected goals, home and away splits, rest days, match context, head-to-head history and how the wider market has priced the match. Every field is checked before anything else uses it, because one bad score corrupts every calculation built on it.

  2. 2

    Build the features

    Raw data becomes the measurements the models actually read: form weighted so recent matches count more, attacking and defensive strength adjusted for the opponents faced, and league baselines so a goal in one competition isn't treated the same as a goal in another.

  3. 3

    Run the models

    The fixture goes to the models trained for its competition group: one per market, with separate home and away models for goals, corners and cards. Each reads its own set of checked inputs and weights every signal by how much it mattered in comparable matches during training.

  4. 4

    Calibrate

    Raw model output is adjusted so the published numbers mean what they say. Most platforms skip this step. If we publish 70%, roughly 70% of those situations should go that way over a large sample. Otherwise the figure is decoration.

  5. 5

    Rank and publish

    Calibrated probabilities are compared with a normalised consensus of publicly available market pricing to find where the model disagrees most. Each forecast is published with a confidence signal and its context, then logged so the outcome can be recorded.

What the numbers actually mean

This is the part worth reading slowly, because it is the most commonly misunderstood thing on any prediction site.

When the model publishes 70% for an outcome, it has calculated a 70% chance. That doesn't mean the outcome will happen. Across many similar situations, about seven in ten go that way and about three in ten don't.

A well-calibrated model at 70% should be wrong about 30% of the time. If it were never wrong at that level, the number would be incorrect. So when a high-probability forecast fails, the model isn't broken. The probability is doing exactly what a probability should.

No single result tells you anything meaningful about model quality, in either direction. Only large samples do.

The confidence signal

Every forecast carries a signal from Neutral to Strong. Strong ratings are rare on purpose: the scale is there to separate the clear-cut from the merely likely.

  • Neutral

    The model sees no meaningful separation. Read it as a coin-flip with extra steps.

  • Solid

    A measurable lean, but well within the range where the alternative is unremarkable.

  • Good

    A clear directional read supported by more than one input.

  • High

    A wide probability gap, with the supporting data pointing the same way.

  • Strong

    The model's highest rating, issued sparingly. Still not a certainty, and never presented as one.

The six markets we model

Each is modelled separately rather than derived from the match result, because the factors that drive them genuinely differ.

Match result (1X2)

Home win, draw or away win. The foundational market, and the one where team strength and form carry the most weight.

Double chance

Two of the three outcomes combined. Useful where the model sees a clear favourite but rates the draw as live.

Both teams to score

Driven far more by defensive record and attacking consistency than by which side is stronger overall.

Total goals

Over and under lines, built from expected goals for both sides and the pace at which the competition typically plays.

Total corners

Corners accumulate gradually across ninety minutes, which makes them less vulnerable to a single moment than a goals line.

Total cards

Referee tendency matters as much as the fixture itself, so this market reads officiating history alongside team discipline.

Value analysis

Next to each probability, the model compares its own figure with a normalised consensus of publicly available market pricing for the same outcome. When the two disagree by a meaningful amount, the fixture is flagged for a closer look.

Normalising matters. Published market pricing includes a built-in margin, so comparing against it raw would make almost everything look like a disagreement. We remove that margin first, so we're comparing two estimates of the same thing.

A flagged disagreement is not a recommendation and not a promise of any outcome. It means our model and the wider market have reached different conclusions about the same match. Sometimes the model has read something the market hasn't. Sometimes the market knows something the model doesn't, such as a late fitness doubt. Use it as a starting point for your own judgement, not a substitute for it.

Where the limits are

Stated plainly, because a model presented without its limitations is a marketing document, not a tool.

Data quality sets the ceiling

Coverage is thinner in smaller competitions. Where the underlying data is sparse, the model is less confident and you should be too.

Late news moves faster than any model

A fitness call an hour before kick-off can change a match materially. We update to kick-off, but no system reacts to information it has not yet been given.

Football is genuinely random

A deflection, a marginal offside or one refereeing decision can overturn the most reasonable analysis. That irreducible randomness is not a flaw to be engineered away.

Past accuracy is not future accuracy

Our published record describes what has already happened. It is not a projection, a target, or a rate you should expect to personally experience.

Read a real forecast

Every fixture page shows the full breakdown: probabilities, the data behind them and the confidence signal. Up to four best picks a day are free.