EV Methodology

What EV means here

Expected value is the answer to one question: if our probability is right and you took this price, what would you earn per dollar on average? In its plain form that is our model's probability multiplied by the payout at an actually-payable price, minus one. A +15% EV means a dollar staked at that price returns fifteen cents of expected profit — if our probability is right, which is the entire bet.

Three rules apply everywhere on the site, in every sport:

• EV is computed against real prices, not fair ones. The price in the formula is a live quote from a named venue. Where we also show a “fair” probability, it comes from de-vigging that venue's own board — and where the venue's margin makes the de-vig genuinely ambiguous (futures boards can carry 60%+ of overround), we say so rather than pretend the ambiguity away.
• Negative EV is published, not hidden. Most of any honest board is red, favourites included. A page that only shows green numbers is a tout sheet; the reds are what make the greens mean something.
• Every probability is walk-forward. The models behind these numbers train strictly on data from before the events they predict, and each sport publishes its backtest the same way. Nothing is fit to the games it is graded on.

Two ways we compute it

Analytic EV — used by most sports on the site (NFL, MLB, tennis, WNBA, and others). The model produces a calibrated probability for the exact market being priced, and EV is that probability against the venue's posted odds. When two venues quote different prices for the same outcome, they get different EVs — tennis futures show this explicitly with separate exchange and sportsbook columns, so a disagreement between venues stays visible instead of being blended into one confident-looking number.

Calibrated EV — used by NBA, our longest-running model. Rather than trusting the model's probability directly, the displayed EV maps the size of the model-vs-market edge to the ROI that edge size actually produced across five seasons of walk-forward backtesting. It answers a more conservative question: not “what does the model believe?” but “what did believing the model historically pay?” Moneyline results and spread results before 2023-24 are withdrawn while the backtest file is corrected. The calibration curves were fitted on that file.

When we refuse to publish a number

The most important part of an EV methodology is knowing when the formula's output isn't information. Every sport on Predictium ships explicit abstain guards: when one fires, the row stays visible and shows “no claim” with a plain-language reason — the row is never silently hidden, and the EV is never quietly clamped. Examples of guards live on the site today:

• Longshot floors. On tennis futures, any runner we give under a 1% title chance shows no EV — at 200-to-1, a tiny error in our probability becomes amplified noise, however appetising the price looks. The threshold is on our probability, never the market's, so venue coverage can't decide who gets a claim.
• Stale inputs. If a player hasn't appeared in a match we observe for an extended stretch, the market may be pricing news — an injury, a layoff — that our ratings cannot see. We publish our probability and the market's price side by side, and decline to claim an edge against people who can see more than we can.
• Event state. Once a match starts, its pre-game EV is dead and the chip disappears — a model number priced against a pre-game line must never be read against a live one.
• Ambiguous rosters and thin markets. WNBA boards abstain — with the reason shown — when a returning player invalidates the model's team state, when a market has no line, or when the sample behind a projection is too thin to price.
• Unearned trust. Our newest sections launch with EV dormant on purpose: soccer claims no edges at all until a forward early line value pass has earned the right to. Publishing probabilities came first; claiming edges has to be paid for with a track record.

One thing we deliberately do not guard on: disagreeing with the market. Wide model-vs-market gaps at long prices are ordinary on outright markets, and suppressing our biggest disagreements would flatter the track record by deleting exactly the claims most worth grading.

How the claims get graded

An EV number is a claim, and claims need receipts. Each sport grades its own in public, two ways:

• Walk-forward backtests, published per sport, unflattering seasons included. Where a page shows the market's line next to ours, it is a calibration reference, not a contest we are trying to win on game day.
• Live early line value tracking. For flagged picks, we record the price we quoted and measure whether the line moved toward it by close, then settle results in units. That is the test we hold ourselves to: being right before the market gets there. The scoreboards accumulate in public and stay “unavailable” until they reach a minimum settled sample rather than debuting on a lucky streak.

By sport

Appendix: the NBA calibration curves

Moneyline results and spread results before 2023-24 are withdrawn while the backtest file is corrected. The NBA calibration curves were fitted on that file, and no live NBA surface prices EV from them in the meantime.