About

English

What the site publishes

Every hour the site publishes, for BTC/USDT, a 24-hour price range: the predicted high and low at two confidence levels (q50 — "half the time the high stays below this"; q90 — "nine times out of ten the high stays below this"), plus a range for the close 24 hours later. Alongside it comes a volatility level, which is the predicted 24-hour amplitude placed against its own history: Low (below the 30th percentile), Normal (30–70), High (70–90), Very high (90th percentile and above). The site does not publish up / down direction predictions: in our research, 24-hour direction probabilities were no better than a constant probability. The main output is the range and the volatility alert.

How to read the coverage numbers

Coverage is the share of past predictions where the actual high stayed below the predicted q90 high, and the actual low stayed above the q90 low; the target is 90% for each side, and 80% for the close range. Every coverage figure is printed with its window and sample size (for example "30 days · 720 predictions"). A number without those two labels is not published. Small samples move a lot; read the window before the percentage. We do not publish a direction hit rate.

Data sources

DataSourceUsed for
BTC/USDT hourly candlesBinance spotprice features and targets
Taker buy volumeBinance spotbuy-sell pressure
Funding rateBinance USDT-M futureslong-short crowding
Open interest, long-short ratiosBinance USDT-M futuresleverage and positioning
Premium indexBinance USDT-M futuresspot-futures basis
Fear & Greed indexalternative.mesentiment
DVOL implied volatilityDeribitoptions-market expectation and benchmark

All sources are public. Each feature only uses values that were already known at the time of the prediction; when a source is delayed the page marks it grey.

The Technical indicators panel on the dashboard shows these model inputs, calculated exactly as the model uses them from closed hourly candles: RSI (14), MACD (12, 26, 9) and its histogram, ADX (14), ATR (14) as % of price, Bollinger position (20, 2), deviation from the 24-hour VWAP, volume ratio (24h), 24-hour realized volatility, and implied minus realized volatility (from Deribit DVOL). The panel only explains what each number measures.

Method

The range model is a LightGBM regressor for the 24-hour amplitude, whose output is turned into high / low / close quantiles by a piecewise calibration fitted on the last part of the training window. Models are evaluated by walk-forward testing: each day is predicted using only data available before it. Direction models (gradient boosting, random forest, logistic regression) have been evaluated the same way against a plain baseline; in that research, 24-hour direction probabilities were no better than a constant probability, so the site does not publish direction predictions. The range and the volatility alert are the product.

Range calibration

The site can widen its ranges based on recent misses. When this is switched on, once a day at 03:10 UTC, right after the hourly reconciliation, it measures how far the actual high and low went beyond the raw model bounds, as a share of the price at prediction time, over the reconciled predictions of the last few days; for the q90 bounds it takes the 90th percentile of that overshoot and moves the bounds of the following predictions out by that share of the current price. By default this is switched off; when on, by default only the q90 bounds are calibrated and only widened (a negative value counts as zero, so a bound is never pulled in), and nothing is moved until enough predictions have been reconciled. The window length, the minimum number of reconciled predictions and the maximum widening are settings; the line below shows the ones in use. Only predictions whose 24-hour window has closed and been reconciled are used. The ranges on the site and all coverage figures use the bounds as published; the raw model output is stored next to them.

Current figures are read from the API when the page opens.

Update cadence

Data is pulled every hour; a prediction is made five minutes past each hour; 24 hours later it is reconciled against the actual high, low and close, and the result goes into the history page and the coverage figures. Models are retrained weekly, and earlier if coverage degrades.

Limits

Ranges are statistical estimates and will be exceeded — by design about one time in ten at q90. Sudden news, exchange outages or data-feed delays are not something the model can see in advance. Nothing here is investment advice; see the terms.

中文

说明

本站输出什么

本站每小时发布 BTC/USDT 的未来 24 小时价格区间:两档置信度下的预测最高价与最低价(q50 ——「一半概率最高价不超过它」; q90 ——「九成情况最高价不超过它」),以及 24 小时后收盘价的区间。同时给出波动预警级别,即预测振幅在自身历史中的分位:偏低(低于 30%)、正常(30–70%)、偏高(70–90%)、极高(90% 及以上)。本站不发布涨跌方向预测:研究里 24 小时方向概率不优于常数概率。主输出是区间与波动预警。

覆盖率怎么读

覆盖率 = 过去的预测中,实际最高价没有超过预测 q90 上沿、实际最低价没有跌破 q90 下沿的比例;两侧目标各为 90%,收盘价区间目标 80%。 每个覆盖率数字都带统计窗口和样本数(例如「30 天 · 720 次」),没有这两个标注的数字本站不发布。 样本少时数字波动很大,先看窗口再看百分比。本站不发布方向命中率。

数据来源

数据来源用途
BTC/USDT 小时 K 线币安现货价格特征与预测目标
主动买入量币安现货买卖力量
资金费率币安合约多空拥挤度
持仓量、多空比、大户比、主动买卖比币安合约杠杆与分歧
合约溢价指数币安合约期现价差
恐惧贪婪指数alternative.me情绪
DVOL 隐含波动率Deribit期权市场预期与标杆

见上表(全部为公开接口)。每个特征只用预测时刻已经可以知道的数据;某个源延迟时页面会标灰。

看板的「技术指标」面板显示以下模型输入,用已收盘的小时 K 线、按与模型完全相同的算法计算:RSI(14)、MACD(12, 26, 9)与柱、ADX(14)、 ATR(14)占价格、布林位置(20, 2)、相对 24 小时 VWAP 的偏离、成交量比(24 小时)、24 小时已实现波动、隐含 − 已实现波动(用 Deribit DVOL)。 面板只说明每个数值衡量什么。

模型方法

区间模型是对 24 小时振幅做回归的 LightGBM,再用训练窗口末段拟合的分段校准把输出转换成最高 / 最低 / 收盘价的分位数。 评估采用 walk-forward:每一天只用它之前的数据训练与预测。方向模型(梯度提升、随机森林、逻辑回归)也按同样的方式与朴素基线比较过: 研究里 24 小时方向概率不优于常数概率,所以本站不发布涨跌方向预测,区间与波动预警才是产品。

区间校准

本站可以按最近的越界情况把区间放宽。打开时,每天 03:10 UTC(那一轮对账之后)统计最近几天已对账的预测里,实际最高 / 最低价越过模型原始边界的幅度(按预测时现价折成比例),q90 两侧各取第 90 百分位,之后的预测把 q90 边界按 「这个比例 × 当前价」向外移动。默认关闭;打开时默认只校准 q90、只向外扩(算出负数按 0,不会把边界往里收), 已对账的预测不够多时不移动。窗口长度、最少条数、最多外扩多少都是设置项,下面一行写的是当前在用的值。 只用 24 小时窗口已经结束、已经对账的预测。 页面上的区间与全部覆盖率用的都是实际发布的边界,模型原始输出另行保存。

当前数字在页面打开时从接口读取。

更新节奏

每小时拉取数据;每小时第 5 分钟预测一次;24 小时后与实际最高 / 最低 / 收盘对账,结果进入历史页与覆盖率。模型每周重训,覆盖率下滑时提前重训。

局限

区间是统计估计,会被突破——按设计 q90 大约十次里有一次。突发消息、交易所故障、数据源延迟都不是模型能提前看到的。 本站内容不构成投资意见,详见用户协议。