Prediction market signals: what a hit rate is worth without its denominator

Every product that sells prediction market signals eventually publishes a percentage. Ours is 82.25%, measured on 17 August 2026. That number is worthless on its own, and the purpose of this article is to explain exactly what it takes to make it worth something — because the answer generalises to any Polymarket signals you are shown, including the ones that are not ours.

Three questions decide whether a hit rate is a measurement or an advertisement. What is the denominator? What happened to the cases that never resolved? And what would the same rate have been if you had simply followed the market price instead? A signal provider who cannot answer all three has published a number, not a result. Below, we answer them for our own, with the figures that make us look worse alongside the ones that do not.

Every figure in this article comes from a single snapshot of the public register, generated at 2026-08-17T10:25:08Z. The register is open without an account, it moves as new slots resolve, and the numbers below will therefore have moved by the time you read them — which is the point of dating them.

What counts as a prediction market signal

A signal is a directional statement, attached to a specific market, emitted at a specific instant, that is checkable against an outcome. Everything in that sentence is load-bearing.

Directional rules out commentary. Specific market rules out the sentiment take that could be scored against half a dozen instruments. Specific instant is the one people skip, and skipping it is how track records get inflated — a claim made "sometime during the week" can be scored against whichever moment flatters it. Checkable against an outcome rules out any signal whose resolution is a matter of opinion.

Our own signals sit on five-minute Polymarket markets for BTC, ETH, SOL and XRP. Four assets, 300-second slots, 288 slots each per day, on the order of 1,150 slots written every 24 hours — a rolling count the register measures, not a constant it derives. Each slot produces a directional read on which side the market resolves. That is a narrow terrain, deliberately: it is narrow enough that every emission is recorded, every outcome is checked, and the whole thing can be published as a register rather than as a claim.

One structural point that changes how the number should be read. The signal is emitted in two stages: an initial read shortly after the slot opens, then a confirmation at T+180 seconds against how the market has actually moved in the interval. The confirmed signal is the product, and the hit rate below is the confirmed signal's rate. A reader who assumes the forecast was made at the opening bell is reading a different, more impressive claim than the one we make.

The denominator problem, worked through on our own numbers

Here is the full accounting, measured in a single transaction so the totals reconcile (2026-08-17T10:25:08Z):

Read the second row again. Nearly half the slots we emitted have no known outcome. Not a wrong outcome — an unknown one: the resolution service failed to retrieve the closing price for that window. Those slots are excluded from the denominator, and that exclusion is the single largest lever anyone has on a published accuracy figure.

It is also a lever that can be pulled in the wrong direction by accident. Those unverifiable slots were once stored as incorrect rather than unknown; they are stored as null today, which is why they are dropped instead of counted. Counting them as wrong again would divide the same correct slots by all 183,017 emitted, and print 43.05% — a number that is not a measurement either, in the opposite direction. Same rows, same database, same day: the entire difference is in how the missing cases are handled.

Because coverage is not constant over time, the monthly series is published next to the rate rather than in a footnote (2026-08-17T10:25:08Z, with August still in progress):

The best-looking month in that table is May, at 87.42% — and it is the month where 73.76% of the slots have no outcome, leaving 9,369 verified out of 35,710 emitted. March, with complete coverage on 23,548 slots, prints the worst rate in the table at 76.83%. That relationship is the whole argument. A rate computed on a quarter of the data is not comparable to a rate computed on all of it, and if coverage is not published month by month, nobody can tell which one they are looking at.

The general form of the rule: a hit rate is a fraction, and a fraction is two numbers. Any provider quoting the numerator alone has, at minimum, not thought about it, and at worst has chosen the denominator that produced the headline.

Confidence is not probability

The second thing sold with prediction market signals is a confidence score. It is routinely read as a probability — a signal at 0.83 is assumed to be right 83% of the time — and on our own data that reading is wrong in both directions.

A confidence score is a ranking device. It orders signals by how strongly the market has already declared itself; it is not an estimate of the probability that the signal is right. The only way to know what a given level is worth is to look up what that level actually did. The register publishes one cut of that, at a confidence of 0.80, measured at 2026-08-17T10:25:08Z:

Two readings, both of which we would rather you take from the table than from us.

The top band beats its own label. A signal carrying a confidence of 0.83 sits in a band that realised 92.14% over 64,775 verified slots. Reading the label as a probability understates what that band did.

The lower band undershoots its label. A signal carrying 0.72 sits in a band that realised 61.59% over 31,011 verified slots. That is the direction that costs a reader money, and it is why the observed rate is now printed next to the score rather than left to be inferred from it.

The two rows are 30.55 points apart, and that gap is what makes the top row an argument rather than a label. If the score separated nothing, both lines would print the same number and the score would have no content at all. Publishing the weak row is not modesty; it is the only thing that gives the strong row meaning.

What the rate would have been anyway: calibration against the price

The hardest question to ask about any signal is the one that most often goes unasked. A prediction market already publishes a probability — the price. A signal is only worth something if it improves on that. So the relevant comparison is not "how often was the signal right" but "how often was the price right, and did the signal beat it".

We measure the market's own calibration by reading the price at T+180s of a 300-second slot, on the engaged side, and comparing it with the observed outcome. All eight bands, including the two where the market resolves less often than its price implies (2026-08-17T10:25:08Z):

95,439 slots, and the eight rows sum to that total — a small point that is worth checking on any table you are shown. The 347 verified slots with no recorded price at T+180s sit outside the table, which is why 95,439 and 95,786 differ, and the difference is published rather than absorbed.

Three things fall out of it. First, the market price is a strong forecast: across most of the range the realised frequency sits within a few points of the price. Second, the two low bands are negative: cheap sides resolve less often than their price implies, the long-shot effect that has been documented in betting markets for decades. Third, and this is the part that governs whether any of it is actionable, the largest positive gap is +3.44 points, in the 0.80–0.90 band — and taking liquidity on Polymarket costs roughly 3.5% of notional, the figure we publish next to the calibration table itself. The measured discrepancy and the cost of acting on it are the same order of magnitude. We publish that fact rather than the gap alone, because the gap alone would read as a return, and it is not one.

An earlier version of this table showed six bands and concluded that the gap was positive everywhere. It got there by omitting the two bands where the gap is negative — 4,692 slots on 17 August 2026. A calibration table with its unflattering rows removed is not a calibration table, and the same test applies to anyone else's.

Where the hit rate comes from, decomposed

If a signal's accuracy comes overwhelmingly from one component, that component is the product and the rest is packaging. Ours splits by whether the market's own move at T+180s agreed with the direction that had been read (2026-08-17T10:25:08Z):

The three rows sum to the 95,786 verified slots in the first table, which is the point of publishing them together.

This is the least flattering table on the site and the most informative. The headline 82.25% is produced almost entirely by the confirmation step. Where the market's move at T+180s agreed with the direction, the read was right 84.69% of the time. Where it did not agree, the read was right 41.94% of the time — worse than a coin flip, on 5,227 verified slots. The third row rests on 347 slots and should not be read at all; it is printed because leaving it out would break the reconciliation above.

That is a real result about what the signal is: it is a fast, disciplined read of an already-moving market, not a forecast of an unmoved one. Anyone selling you prediction market signals should be able to produce this decomposition. If the accuracy survives removing the component that reads recent price action, the claim is about forecasting. If it does not, the claim is about reaction speed — which is a legitimate product, just a different one, and it should be sold as such.

Per-asset, the rates are close enough to be uninteresting, which is itself reassuring: BTC 81.72% on 22,316 verified slots, ETH 82.49% on 24,449, SOL 82.37% on 23,805, XRP 82.37% on 25,216, all measured at 2026-08-17T10:25:08Z. A signal that worked spectacularly on one asset and not the others would be a warning sign, not a feature.

How to evaluate any prediction market signals you are sold

A checklist that works on our numbers and on anyone else's.

• —Ask for the denominator, in the same breath as the rate. Not "82%" but "82.25% of 95,786 verified slots, measured at 2026-08-17T10:25:08Z". If the second half is missing, the first half is not information.
• —Ask what happened to the unresolved cases. They were dropped, counted as wrong, or counted as right. All three are defensible; only one of them was disclosed. Ours drops them, and publishes the share dropped: 47.66%.
• —Ask for the coverage series over time. A single all-time figure hides the months where coverage collapsed. Our best month is our thinnest month, and you can only see that because the series is printed.
• —Ask what the price alone would have delivered. If the signal is not compared against the market's own probability, the comparison being invited is against a coin flip, and almost anything beats a coin flip on a market where prices are informative.
• —Ask what it costs to act. A 3.44-point discrepancy against a 3.5% cost of taking liquidity is not a strategy. The odds and probability converter is useful here: it turns a price into an implied probability so you can see how large your disagreement with the market actually is before you price the fee on top.
• —Ask for the decomposition. Which component carries the accuracy, and what is the rate without it?
• —Ask when it was measured. Every figure in this article carries the same snapshot, 2026-08-17T10:25:08Z. A rate without a date is a rate that can quietly stop being true.

The formal treatment of what calibration means, and why a well-calibrated forecast can still be useless, is in forecast calibration. The counterpart question — how to judge a wallet rather than a signal — runs on exactly the same logic, and is worked through in Polymarket whale tracking: of the 145,141 wallets we track, 36,354 show a win rate of exactly 100% (17 August 2026), which is the same failure as an 87.42% month computed on 9,369 of 35,710 slots. And the reason a small measured discrepancy is so rarely enough to trade on is the cost structure set out in why most prediction-market traders lose money.

Frequently asked questions about prediction market signals

What is a good hit rate for a prediction market signal?

The question has no answer without the denominator and the baseline. On five-minute crypto markets where prices are already strong forecasts, a rate in the eighties is unremarkable — the price alone realised 88.64% in the 0.80–0.90 band on 17 August 2026. The number that would be remarkable is a persistent gap against the price, net of the roughly 3.5% cost of taking liquidity, on a sample large enough to rule out luck.

Why do you publish the months where your coverage was bad?

Because the alternative is publishing a figure that cannot be checked. May 2026 shows our highest rate, 87.42%, on 9,369 verified slots out of 35,710 emitted. Quoting the 87.42% alone would be technically true and materially misleading, and anyone who later found the coverage number would be right to distrust everything else on the site.

Is a confidence score the same as a probability?

No, and on our data the gap is large in both directions. Measured at 2026-08-17T10:25:08Z, the 0.80-and-above band realised 92.14% over 64,775 verified slots; everything below 0.80 realised 61.59% over 31,011. Reading either label as a probability gets you a materially wrong number. This is why the observed rate and the verified count are published alongside the score.

Can I trade on these signals?

We publish measurements; we do not execute trades, manage funds, or make recommendations. What a signal gives you is a dated, checkable directional read with its historical accuracy and its sample size attached. Whether that is worth acting on, at what size, against a cost of roughly 3.5% of notional, is a decision that stays entirely with the reader.

Which assets and which tiers get live signals?

Live signals are part of Operator and Principal. Operator covers BTC and ETH; Principal covers BTC, ETH, SOL and XRP. That restriction is the behaviour of the API, not a marketing distinction, and we state it before purchase rather than after — a customer discovering post-payment that they have two assets out of four is a complaint we would deserve. The pricing page carries the full grid.

What is the difference between your signal and the market price?

The price is a probability set by everyone trading the market. The signal is a directional read produced by observing that price move over a fixed window and confirming at T+180s. The decomposition above shows how much of the accuracy comes from the confirmation step: 84.69% on the 90,212 verified slots where the move confirmed the direction, against 41.94% on the 5,227 where it did not, both measured at 2026-08-17T10:25:08Z.

Everything above is on the public register, which is open without an account: the totals, the monthly coverage, the eight calibration bands including the negative ones, and the decomposition that shows where the accuracy comes from. The Observer tier costs nothing and requires no card. Read the register first, check the arithmetic, and decide whether a measurement published with its own weaknesses is worth more to you than a percentage published alone.