How to read a CoinPredict price prediction

A coin page shows one range per horizon. The first sentence under the heading states it plainly — a low, a base case, a high and a confidence — and everything below the fold is the evidence for that sentence. If you read nothing else, read that line and the update time next to it.

Low, base and high are quantiles

The three numbers are not a best case, a normal case and a worst case. They are the 10th, 50th and 90th percentiles of the model's distribution for that horizon:

• Low is a level the price should finish above about 90% of the time.
• Base is the median. Half of the modelled outcomes sit above it, half below.
• High is a level it should finish below about 90% of the time.

That makes the band an 80% interval. It should contain the outcome about four times in five, and it should miss about once in five. A miss is not the model breaking; a miss is the interval doing what an 80% interval does.

The width is the honest part. A coin with high volatility gets a wide band because its outcomes really are spread out. If a site shows you a narrow band and a confident direction on a volatile asset, ask what the band is made of.

Confidence is a claim you can check

Next to the direction there is a number: the probability the model gives to that side. It is not a raw classifier score. It has been through a calibration step fitted on data the classifier never saw, so calls published at 70% should be right about seven times in ten.

You do not have to take that on faith. The calibration chart on the track record page plots stated confidence against observed hit rate, with the number of graded calls behind each point. If the points sit near the diagonal, the numbers mean what they say. If they drift, the chart says so before you have to.

A missing row is a decision

Some horizons are absent from the table. That is not a rendering bug. A daily horizon publishes its directional model only when the walk-forward backtest shows it beating a coin flip, and it publishes a range at all only when that range held the price within the expected coverage.

Failing either check removes the row rather than showing a stale one. The methodology page lists the gates and the model card records which horizons passed on the version that made the call.

"Neutral" is the most common state

Most calls are neutral. That is the design: publishing a weak lean as though it were a strong one is how prediction sites lose the one thing that matters, which is being believable over time.

The hourly board has the same idea in different clothes. When confidence does not clear the threshold, the hour is published as no setup and graded as no call. The board shows its no-setup rate next to its hit rate, because a board that calls something every hour is either lying or overtrading.

The update stamp is the honest timestamp

"Updated" is the time the newest number on the page changed — a forecast run, a price tick or a regenerated analysis. It is not the time the page was served from cache, and it does not reset because someone visited.

For the underlying numbers: prices arrive every few minutes, forecasts are published daily, hourly signals are republished each hour, and written analysis is regenerated weekly or when a coin moves far enough to justify it.

What it is still not

A forecast is a distribution, not a promise, and no range covers the risks that matter most in crypto: a hack, a delisting, a regulatory change or a liquidity hole. The site is information, not advice, and a single call — right or wrong — tells you nothing on its own. The track record is the claim; read that next.