The market quotes a price.
The panel quotes one back.

Executable edge, in probability points

The big number is the model's probability minus what it costs to buy that side right now: the ask, spread included, not the midpoint. It is the only version of the number you could have acted on. Here the model says 81.7% and the offer is 61¢, so the edge is +20.7 points — while the gap to the book's midpoint, shown on the card below, is 21.2. The difference is the spread, and pretending it doesn't exist is how backtests lie.

The chip picks the side with the larger edge and the panel follows it. Under about two points, the number is inside the noise of estimating volatility, and the panel says so rather than dressing it up.

SERVER ⚡ = Pro With a Pro session, the headline stops being computed in your browser and starts streaming from our engine — the same numbers we paper-trade and publish every night on /performance. Free keeps the model probability; the executable edge is the paid half.

Both estimates, for the whole life of the window

Two probabilities of the same event over time: ours in amber, sampled once a second, and the market's — the order book midpoint — in blue. The shaded area between them is the disagreement, and its colour carries the direction.

The curve above is the real one from the archived window. For eleven minutes the two lines are on top of each other, never more than six points apart, because a liquid book is fast and mostly right and we say so. Then spot turns, our estimate turns with it, and the book stays priced for a move that already finished. The header keeps the current gap — Δ +21.2 pts — visible even when the card is folded.

Five books, one price, and how far ahead of the feed it is

The market resolves against a single published feed, and that feed is a relay: it summarises venues it is slower than. So the panel reads five exchange books over direct websockets — Binance, Coinbase, Kraken, OKX, Bybit — at roughly a hundred milliseconds each.

The five quote in different units: dollar-pegged tokens on some venues, actual dollars on others, tens of dollars apart. Each source therefore carries a per-source offset, an exponential moving average against the raw feed, applied before anything else; the composite is the median of the five corrected books, so one stalled venue moves nothing. The second tile is the lead — how far in front of the feed the composite currently sits — and it lights up when a move is happening on exchanges and has not reached the feed yet. The sparkline shows the last ninety seconds of both, with the price to beat as the dashed red line.

The window shown here pre-dates the 7 August 2026 switch to TWAP settlement, so its sparkline carries two series. On a market that resolves on the TWAP, a third line joins them in violet, dashed: the sixty-second average that actually settles the window, about thirty seconds behind the raw feed by construction.

the lead number is Pro

The shape the probability came out of

The distribution of the terminal move over the time that is left, with the distance to the price to beat drawn on it as a dashed line in σ. The green area is P(Up), the red one P(Down); they are the probability, not an illustration of it.

Switching the model to Student-t with ν=4 — renormalised so its variance still matches σ — visibly fattens the ends of this curve. Both curves agree to about 2σ and then part company by one to two orders of magnitude, which is exactly the region where a window nobody thought was live turns over.

Every input the probability was computed from

Spot as the model sees it, the price to beat, the distance in dollars and in σ, σ over the remaining horizon, annualised volatility with the candle window it used, the market's own implied probability, and the probability of the spot path touching the beat at least once before expiry.

Two conventions worth knowing. Distance σ is signed from the beat towards the spot: it reads −0.90σ here because the spot sits above the beat, which is the favourable side for Up — the dollar distance right next to it carries the plain-language sign. And P(touch) is indicative: it is twice the terminal probability by the reflection principle, computed on the spot path, while resolution rides a sixty-second average of a different series. The panel says so in its own tooltip rather than quietly presenting it as exact.

What the market assumes, against what actually happened

Invert the market price and you get the volatility the market is assuming — a binary-option implied vol. Put it next to realised volatility from one-minute candles and you can read the market's state of mind in numbers rather than vibes. On this window the book was pricing 189% annualised against 53% realised: it expected far more movement than the recent past, which is another way of saying it thought the move could still come back.

figures are Pro Free keeps the badge in the header, so you can see when implied and realised are far apart; the numbers themselves are paid.

An hourly ladder, priced strike by strike

Kalshi's crypto hourlies quote a ladder of strikes rather than a single Up/Down question. The panel prices every strike near spot with the same model and puts each one next to both asks — yes and no — so the side with an edge, if any, is the coloured one. Rows with nothing to say stay dimmed.

The probabilities in this example are computed from the model at the spot and horizon shown above; the asks beside them are illustrative — no Kalshi quotes are archived for this window. Kalshi's settlement source is not the one we calibrate the composite against, so the panel warns that sub-two-point edges there are noise.

The assumptions are yours to change

Normal or Student-t and its ν. The volatility window — automatic by market duration, or pinned to 60, 90, 120, 240 or 480 one-minute candles. Fast multi-exchange spot on or off, so you can see exactly what the composite is buying you. And an experimental short-term drift term, off by default, because very-short-term momentum is mostly noise and we would rather say that than sell it.