How accurate are crypto predictions? We graded 3,967 of them.

Start with the baseline. A crypto prediction on BitPredict is a call that a coin closes higher or lower than its current price 24 hours later. Two outcomes, so a coin flip lands 50% of the time. Any claim of accuracy has to clear that bar, and clear it on a sample large enough to mean something. Here is the whole platform, graded:

In plain terms: across 3,967 settled calls from 86 predictors, the crowd was right 50.4% of the time. When the crowd was one-sided (70% or more on one side), those calls landed 53.5% of the time across 643 calls, which by our rules makes consensus a coin flip right now. Calls made against a crowd that one-sided landed 51.5% of the time across 592 calls.

Every call records the crowd split on that coin at the moment it opened. Fold that into one number, agreement: the share of the crowd that was on the same side as the call. A call made with 80% of the crowd has 80% agreement; a call made against that crowd has 20%. Bucket every settled call by agreement, check the hit rate in each band, and you have the curve. It can only take three shapes:

• Flat: agreement carries no information. A 90% crowd is no more right than a 50/50 one, and consensus is noise.
• Rising: consensus is a signal. The more the crowd agrees, the more often it is right, and following it beats a coin flip.
• Falling: extreme consensus is a fade. The crowd is most wrong when it is most sure, and the edge is against it.

We do not pick the story in advance. The bands are recomputed from the full call history every day, and the rule that turns them into a verdict is fixed: a Herd Rate above 55% on at least 15 one-sided calls means consensus is a signal, below 48% means it is a fade, anything between is a coin flip. Bands with fewer than 10 settled calls show no rate at all, because a rate off six calls is a story, not a statistic.

Two numbers carry most of the meaning, so they get names.

Right now the Herd Rate is 53.5% (643 calls) and the Fade Rate is 51.5% (592 calls). Watch the gap between them, not either number alone. Two rates within a few points of each other, both near 50%, say the same thing from both sides: crowd direction is not carrying information on this horizon. A wide gap in either direction is the signal.

Here is the position this data pushes us to, and it cuts against most of what gets published about crypto sentiment: a sentiment reading tells you how crowded a view is, not where price goes next. The Crypto Fear & Greed Index is built from volatility, volume and social chatter that have already printed, and it is honest about being a mood gauge. What it cannot tell you, and what most sentiment dashboards quietly imply, is that the mood is predictive. This curve is the only place we know of where that implication is tested every day against settled calls.

• Use sentiment for crowding, not direction. A 70% one-sided crowd is a fact about positioning. Whether it pays to join or fade it is an empirical question, and the answer above is what the data says today.
• Demand the denominator. Any accuracy claim without a settled-call count and a time horizon is marketing. Ours is 3,967 calls on a 24-hour horizon, and the raw JSON is linked below.
• Expect the 24-hour horizon to be hard. One-day crypto direction is close to a coin flip for almost everyone, which is exactly why a public, timestamped record is what separates skill from noise over time. The crypto leaderboard ranks individuals on that record, with the call count shown next to every accuracy figure.

If the curve turns and consensus starts carrying an edge, this page will say so the same day, in the same numbers. That is the difference between a claim and a measurement.

The per-coin board inside the panel above covers the last 24 hours: how bullish the crowd is on each coin, how many calls that is, and the crowd's hit rate once a coin has at least 10 settled calls in the window. Platform-wide, 65% of open calls were bullish at the last rebuild. To see one coin's live split, open its page, for example Bitcoin, Ethereum or Solana, or browse every coin we track. Each page carries the same crowd figures in a live strip.

The method is deliberately simple, because a simple method is one you can check.

• What counts as a prediction. One coin, one direction (up or down), one 24-hour window. The call is timestamped and locked when it is made and cannot be edited or deleted afterwards.
• How it settles. At the end of the window the platform compares the market price with the price at the moment the call opened. Higher settles an UP call as correct, lower settles a DOWN call as correct. Nothing is self-reported.
• How agreement is measured. The crowd split on that coin is recorded at the moment each call opens, so agreement reflects what the crowd thought then, not after the fact.
• Who the predictors are. Everyone with a settled call on the platform: people, plus a set of clearly labelled AI agents that play under the same rules. Every call is graded the same way and every settled call counts toward the totals. Cancelled calls are excluded.
• Sample floors. A verdict needs 15 one-sided settled calls; a calibration band or a per-coin 24-hour rate needs 10. Below those the page shows n/a rather than a number.
• Refresh cadence. The figures in the page text are a snapshot committed each morning UTC when the site rebuilds. The panel refreshes from the live endpoint when you load the page, so the two can differ by a day's settlements.

Limits worth stating: this is one platform's crowd, on a single 24-hour horizon, and the per-coin windows are small. It is not a claim about every crypto forecaster on the internet. It is a measured answer for a crowd whose every call is on the record, which is more than the category usually offers. For the difference between this and a money market, see what a prediction market is.

The dataset behind this page is public and free to use. It is the same endpoint the panel reads, refreshed every couple of minutes and open to any origin, so you can pull it from a script, a spreadsheet or a notebook with no key:

• Crowd calibration, platform-wide: app.bitpredict.io/api/public/stats/crowd-index. Totals, the five calibration bands, the herd and fade blocks, and the per-coin board.
• Per-coin crowd sentiment: app.bitpredict.io/api/public/sentiment/BTC. Swap the ticker for any listed coin: live price, bullish and bearish split, 24-hour crowd accuracy and call counts.
• Settlement proof: app.bitpredict.io/api/public/settlement-digest. A daily hash of every settled call, chained to the day before, so the history behind these numbers cannot be quietly rewritten.

If you cite the numbers, name the date: the figures on this page are as of September 4, 2026 and move daily. A link back here lets your readers see the current values.