Methodology

A market's fair probability is estimated from the prices of sharp reference sources — a small set of venues whose prices historically track true outcome frequencies most closely — with the bookmaker margin (the "vig") removed. When multiple sharp sources price a market, their de-vigged probabilities are combined into a consensus.

Sources are used in a strict order of measured reliability. When the primary sharp reference is unavailable, the estimate falls to a consensus of other sharp books, then to liquid betting exchanges, then to prediction markets, and finally to a no-vig consensus of the whole market. Each fair price is labeled with the source tier that produced it, and confidence is capped by that tier's measured historical reliability — a fair price from a weaker source is never presented with the certainty of a stronger one.

De-vigging uses standard published techniques (proportional, power and Shin-style adjustments, which account for the favorite–longshot bias documented in our settled data). Try it yourself with the free no-vig calculator.

EV is the standard definition: for decimal odds o and fair probability p, EV = o × p − 1. A +3% EV price pays 3% more than the estimated fair value of the outcome. EV at the moment of observation is recorded separately from live EV, which is recomputed as prices move.

EV is an estimate whose quality depends entirely on the fair price behind it. That is why every published edge carries its source tier, the number of independent sharp sources pricing the market, and a confidence score — and why we measure ourselves by closing-line value rather than by picked winners.

Confidence starts from the depth and agreement of the sources behind a fair price and is penalized for thin markets, single-source estimates, unusual disagreement and data-quality flags. Separately, segment calibration compares modeled EV against realized returns across thousands of settled signals per segment (sport × market type × odds range). Segments that historically under-deliver their modeled edge are discounted; segments that deliver are not. The calibration table is recomputed from settled results, not set by hand.

Signals are ranked by an expected-return model trained on settled outcomes. Its inputs are observable market facts (odds range, market type, sport, source depth); its target is realized return. We publish what it is — a model of what has actually paid — rather than its internals.

CLV measures whether a price beat the market's final, most-informed price: positive CLV means the market moved toward the signal after we flagged it. Because single results are dominated by variance, CLV is the professional yardstick for whether an edge was real — a losing bet with positive CLV was probably a good bet; a winning one with negative CLV probably wasn't. We record closing prices for tracked markets and publish realized CLV on settled records. Compare your own bets with the CLV calculator.

Identity is never guessed: teams, leagues and bookmakers are matched by strict, curated mappings — never by string similarity. Records that cannot be matched are excluded rather than approximated. Automated detectors flag impossible EVs, mismatched lines and contaminated fixtures, and flagged records are excluded from public numbers. Where correct output cannot be produced, no output is produced.

Fair prices are estimates, not truths. Quoted prices may be limited, unavailable or gone by the time you look. Positive EV is a long-run statistical property, not a per-bet promise: realized returns swing hard over any sample a human would call large. Nothing on this site is betting advice.