A price, an opinion, and the gap between them.

Projections are collected from more than a dozen independent forecasters and blended into a single number for every game, player and race. The blend is not an average. No source is trusted on reputation. Each one is scored against real outcomes, market by market, and a forecaster that cannot beat the category average is dropped to zero weight rather than quietly averaged in.

Agreement between sources is not treated as evidence either, and that is a measurement rather than a caution. We tested whether the sources agreeing predicts a better result, market by market and projection band by projection band, and it does not. In only ten of twenty-four cells did disagreement widen the error at all, and the middle case moved it by about six percent — noise, next to what a reader would reasonably infer from a number that says nine of nine.

There is a structural reason to expect that. Not every source is independent: several are scraped from overlapping upstreams, so some of the agreement on any given line is two copies of one opinion rather than two opinions. That is also why the site shows you the count and the split and stops there. It is context for reading a line, not a score, and it is deliberately drawn in the same flat grey whether the field is unanimous or evenly divided.

That number is then priced against the live market — the same market a reader can actually bet into — and what this site publishes is the disagreement, not the forecast on its own.

A projection is a point; a bet needs a distribution. We track how far each market’s projections actually land from the result and size the uncertainty from that spread, instead of a constant chosen once and left alone. A market the model reads well earns a tighter distribution than one it reads badly, and the difference shows up in the rating rather than being smoothed away.

Where a market’s outcomes run wider than a simple count model allows, we use a distribution that fits — but only where the evidence earns the change. Coverage runs across NFL, NCAA FB, NBA, WNBA, NCAA MBB, NCAA WBB, MLB, NHL and Soccer; game lines, totals, team totals and player props.

Hit rate and calibration are tracked market by market and published, including the places the model has been wrong. That is the whole argument for it: a model that is never scored is just an opinion with a decimal point.

Every race is modelled on the history of that track and that surface — a dirt sprint is not scored on turf routes run somewhere else. The model learns finish order directly, ranking a field rather than rating horses in isolation, across speed and class figures, pace shape and running style, post position, and the recent form of the jockey, the trainer, and the two of them together.

Form is read as of race day. When the model is tested against last season it sees only what was knowable at the time, so a good backtest cannot be an artefact of hindsight — which is the most common way a racing model flatters itself.

Exacta, trifecta and superfecta prices are built from the same field, using the long-established correction that a horse’s chance of running second is not simply its chance of winning. Skipping that step is what makes naive exotic prices look far more attractive than they are.

The same bet is priced differently everywhere, and the difference is the whole game. A spread at −108 at one book and −118 at another is a positive bet at the first and a negative one at the second. Every board here shops the books you have selected and shows the best price among them, with the operator named on the row.

A new account starts on 5 national books and can change that in a click. The full list is here — sixteen operators, including two prediction exchanges, which price the same events differently.

Expected value alone is not confidence. A big edge on a market the model knows little about is worth less than a small edge on one it knows well, so every bet carries a rating that combines the size of the edge with how much the model trusts itself on that market.

• 4–5 — a large edge the model is confident in. A handful a week.
• 3 — real but modest. The bulk of what gets published.
• 1–2 — a low edge, low confidence, or both. Shown, not recommended.

The bar never moves to fill a quiet day. Some Tuesdays nothing rates above a 2, and the board says so rather than promoting the best of a bad slate.

Every prediction the model has settled, grouped by the rating it gave itself. Where the model broadly agrees with the price, it is calibrated — it says 17% and 18% happens. Where it departs furthest from the market, the claim and the outcome come apart.

• Everything — claimed 33.6%, happened 33.5%, a gap of +0.1 points across 221,687 settled predictions.
• Rated 0 — claimed 16.8%, happened 18.8%, a gap of −2.0 points across 132,174 settled predictions.
• Rated 1 — claimed 70.5%, happened 72.1%, a gap of −1.6 points across 46,638 settled predictions.
• Rated 3 — claimed 30.2%, happened 22.0%, a gap of +8.1 points across 26,916 settled predictions.
• Rated 4 — claimed 69.2%, happened 63.5%, a gap of +5.7 points across 11,608 settled predictions.
• Rated 5 — claimed 72.6%, happened 56.0%, a gap of +16.6 points across 4,351 settled predictions.

This is measured, not argued, and it is why the rating is capped rather than trusted at the top of its range: a model that is most confident exactly where it is least reliable has to be stopped from compounding the two. It is also why this site publishes no return by rating — the ordering of these buckets is stable and their levels are not. The whole table, market by market and updated nightly, is on the record page.

• No promised return, no edge over the market, no win rate. Not because they would be hard to write, but because they would have to be re-proved every month — and the measurement is the thing that stays true on a losing week.
• There is no bet slip and no button that places anything. We do not hold money, and there are no affiliate links on a price.
• No lock of the day, no free pick, no capper theatre. Every bet the model rates is published and graded, whatever it does.
• This is research rather than advice. The model is wrong often enough that treating it as advice would be a mistake, which is why the record is free and the losses are in it.