Methodology

This is the full breakdown of how Pro Edge^™ actually prices a selection - the data behind it, the models that turn it into a probability, the discipline that decides what gets published, and, just as importantly, what the approach doesn't claim. For why we grade ourselves on closing-line value rather than a win-rate headline, see the homepage or our FAQ.

1. Data ingestion and normalisation

Every model run starts from two streams: team and player data across the covered leagues, and real-time price data pulled from multiple bookmakers. Both are normalised before a single probability is calculated - adjusted for confirmed lineup news, fixture congestion, and known structural effects (home advantage by league, referee tendencies, travel distance) that would otherwise bias a naive read of the raw numbers.

Team and player strength are not read off season-to-date totals - a plain average overweights early-season form and is slow to react to a genuine change in personnel or tactics mid-season. We instead use exponentially-weighted rolling windows, so a rating adapts to recent form while still anchoring on a sample long enough to avoid overreacting to a single freak result - regression to the mean, applied to the inputs rather than left to distort the output. The same pipeline continuously checks for distribution drift: a rule change, a shift in refereeing strictness, or any other structural break can silently move the baseline a league was calibrated against, and a drift severe enough to break that calibration pulls the affected league from publication until it is re-estimated.

2. An ensemble, not a single model

No individual model captures every relevant signal, so each selection is priced by three independent methods, then reconciled:

1. Poisson regression - the workhorse for scoring-based markets (totals, both-teams-to-score, correct score). Goals are modelled as a Poisson process: the probability of a side scoring exactly k goals given a rate λ is P(X = k) = e^−λλ^k/k!, where λ itself is estimated as a log-linear function of attacking and defensive strength: log λ_home = attack_home − defence_away + home-advantage (and symmetrically for λ_away). Pairing an independent rate estimate for each side yields a full scoreline matrix, from which any goal-based market can be priced by summation. In practice the two goal counts are not quite independent - low-scoring draws occur slightly more often than the naive product predicts - so we apply the standard Dixon-Coles correction to the handful of low-scoring cells (0-0, 1-0, 0-1, 1-1) rather than treat the two sides as perfectly separate processes.
2. Gradient-boosted trees - captures non-linear, situational interactions that a log-linear model like Poisson regression cannot represent without being told where to look. Rather than fitting one model, boosting builds a sequence of shallow decision trees, each one trained on the error the current ensemble still leaves unexplained - a form of gradient descent carried out over functions rather than parameters, scored on log loss since the output is a probability, not a point estimate. Shrinking each tree's contribution and capping its depth are standard regularisation choices that trade a slower fit for one far less prone to memorising noise in the training sample. The payoff is that interactions the model was never explicitly told to look for - fixture congestion mattering far more when compounded with a long-haul away trip than either factor does alone, say - fall out of the tree splits directly.
3. Bayesian updating - revises the prior estimate as new information arrives close to kick-off (confirmed lineups, late team news), rather than freezing the model's view at the start of the week. Formally, the posterior probability of an outcome given the news is proportional to the prior probability of that outcome multiplied by the likelihood of the news under it: P(outcome | news) ∝ P(outcome) × P(news | outcome). The prior here is the ensemble's early-week estimate, built from a full season of data; the news is whatever has changed since. Updating this way, rather than discarding the early estimate and refitting from nothing, means the published probability reflects both the weight of the season's evidence and the latest team news, without letting a single late-breaking data point swing the estimate further than it warrants.

Each method outputs a probability with a confidence interval. Where the methods disagree sharply, that disagreement is itself information: it widens the uncertainty band rather than being averaged away.

3. From probability to edge

The reconciled modelled probability is compared against the market's price, adjusted for overround so the two figures sit on the same scale. The gap between them is the edge. A positive edge alone isn't sufficient: it has to have survived backtested transaction-cost and variance testing across our historical sample before it clears the bar for publication. Most candidate selections fail this gate and are discarded silently - see our Glossary for the precise definitions of edge, overround, and the other terms used here.

Before any edge is treated as real, the ensemble's calibration is checked directly: among every historical selection assigned a given modelled probability, did that outcome actually occur at close to that rate? A model can look profitable by luck over a short sample while being poorly calibrated, which is why backtests are scored on log loss rather than accuracy alone - a rule that penalises confident, wrong predictions far more heavily than cautious ones. Where the historical sample is too thin to judge confidently, we run a Monte Carlo simulation over the same historical fixtures to estimate the range of results a genuinely fair-priced edge of that size could plausibly produce by chance alone, and require the observed edge to sit outside that range before it clears the bar for publication.

4. Staking methodology

A unit is defined by splitting the bankroll into equal parts: never more than 1/20th (5%) of the total, and often smaller. Once set, that unit size is held fixed for a minimum of 30 days rather than recalculated bet to bet - the period, not any individual result, is what triggers a resize.

Every published selection is staked at exactly 1 of those units, regardless of the size of its edge. A single, fixed stake keeps performance comparable selection to selection and removes any temptation to size up after a loss or down after a win - the discipline holds independently of how the previous selection settled. Stakes are unit-based throughout, so the published figures hold regardless of an individual's bankroll.

The Kelly criterion - staking the fraction of bankroll that scales with the edge relative to the odds - is a legitimate alternative to flat staking; see our Glossary for the formula. We do not apply it to our own published figures, and we do not manage or recommend it for members. A member who judges their own risk tolerance differently may choose to size their own stakes this way against our published edge, entirely at their own discretion and risk.

5. Validation and ongoing tracking

Every published selection is logged and timestamped before kick-off, then settled and graded against the closing line once the market has finished pricing the event. Closing-line value (CLV) - computed the same way as edge, but as the percentage-point gap between the closing price's implied probability and the price we actually took - is the primary ongoing check on whether the model's edge is real, not raw win/loss rate: a model can lose a string of bets while still beating the close on all of them, and that combination is a stronger signal of a working model than a winning streak with negative average CLV. We publish the average CLV across every settled selection on the homepage, and the full, unfiltered record - wins and losses alike - on our Statistics page.

Limitations

No quantitative model removes all uncertainty, and we don't represent ours as doing so. Markets are not static: rule changes, new public data sources, or a shift in how efficiently a market is priced can erode an edge that backtested well historically, sometimes without obvious warning. Past performance, including any verified CLV or yield figures we publish, is not a guarantee of future results. Where a model's confidence interval is wide, we treat that uncertainty as a reason to stay silent rather than to publish anyway - see our FAQ for how this plays out in practice.