Liquidation Forecast

An indicator in its own pane: the probability that the instrument will have at least one forced liquidation in the next 15, 60 or 240 minutes. It answers "when", not "which way".

What it is

Liquidation Forecast is the second half of the liquidation model. Liquidation Map shows where in price the forced exits sit. Liquidation Forecast shows how likely it is that the instrument's liquidation flow wakes up at all in the next few minutes.

The indicator is called [SFM] Liquidation forecast; on the chart it is labelled [SF] Liq Forecast. To add it: the Indicators button on the chart → search for "Liquidation forecast".

• Plan: Pro and Max. On Free the indicator draws only bars older than 6 hours; the latest bars stay empty.
• Exchanges: Binance and Hyperliquid. Each exchange has its own model: the liquidation flows of Binance and Hyperliquid arrive in different forms, and their features are on different scales.
• Where it is drawn: in a separate pane below the chart, on a percent scale. Not over the candles: probability and price on one axis would only confuse.

The same 15-minute probability is shown in the 15m column of the Liquidation Spike screener. The screener shows it for the whole market at once; the indicator shows the history for one instrument.

What it draws

Three lines, one per horizon. The value is the probability of at least one liquidation within the horizon, in percent.

Line	Color	Width	Default
Probability over 15 min	light blue	2	on
Probability over 60 min	purple	1	off
Probability over 240 min	teal	1	off

The colors are cool on purpose. On the platform red and green mean the side of the sweep, and the probability says nothing about the side or the price direction.

A bar's value is the forecast at its close: the last minute inside the bar is taken, not an average. The forecast is recalculated once a minute.

A gap in the line means "not calculated", not zero. If there was no model, seasonal base or volatility for a minute, the point is skipped. A zero on the chart would be a confident "nothing will happen" where the model simply knows nothing.

Settings

Parameter	Type	Default	What it does
Horizon 15 min	checkbox	on	Shows the 15-minute line
Horizon 60 min	checkbox	off	Shows the 1-hour line
Horizon 240 min	checkbox	off	Shows the 4-hour line

Which horizon to pick

It depends on the instrument, so the choice is left to you.

• Liquid instruments (BTC, ETH, SOL and the like) — watch 15 minutes. Over an hour, let alone four hours, they almost always have at least one liquidation: the line sits at the ceiling and tells nothing apart.
• Thin instruments — the opposite: turn on 60 or 240 minutes. Over 15 minutes their probability is almost always low, and the differences show on the longer horizon.

How it is calculated

There is one model for all instruments of an exchange. Only the input is specific to the instrument: its seasonal base, volatility and recent flow.

For each minute and each horizon $h$ (15, 60, 240 minutes), five features are taken:

Feature	What it is
$x_1 = \ln(1 + B_h)$	Expected liquidation volume over the horizon from the instrument's seasonal base, $
$x_2 = \ln \sigma_{60}$	Volatility of one-minute returns over the last hour
$x_3 = \ln(1 + L_{60})$	Liquidation volume over the last hour, $
$x_4 = \ln(1 + A)$	Minutes since the last liquidation, capped at 1440
$x_5 = \ln(1 + L_{15})$	Liquidation volume over the last 15 minutes, $

The seasonal base is the instrument's expected flow at this hour of the day, separately for weekdays and weekends. How it works is described in the Liquidation Spike article.

From the features, a logistic regression gives a raw probability:

$$z = a_0 + \sum_{i=1}^{5} a_i x_i, \qquad p_{\text{raw}} = \frac{1}{1 + e^{-z}}$$

Then the raw probability goes through calibration — a monotonic piecewise-linear correction $p = C(p_{\text{raw}})$. Logistic regression orders minutes well, but its level can drift from the actual frequency of events. The correction moves the level and keeps the order. The coefficients $a_i$ and the correction are fitted on history separately for each horizon.

How far to trust it

• Calibration. In a calm market the promised percentages match the observed frequency with an error of about one point: of the minutes with a 30 % forecast, about 30 % had a liquidation. In a market-wide cascade the error grows roughly threefold: the model looks at one instrument and cannot see a regime shift of the whole market.
• Ranking. In testing the model tells minutes before a liquidation from quiet ones better than random guessing: AUC 0.68–0.78, for both large and the thinnest instruments. That is a moderate result, not a precise forecast.
• Expected volume is not shown. The model computes it too, but it is reliable only on dense instruments: $R^2$ 0.44 for the most liquid versus 0.09 for thin ones. A line that is right for a fifth of the market would do more harm than good.

What the indicator does not show

• Price direction. A high probability means a liquidation flow is likely. Whose positions get swept and where price goes after that, the model does not predict.
• The side of the sweep. Longs or shorts — see Liquidation Map and the liquidations feed.
• Size. One small liquidation and a million-dollar cascade are the same event for this probability.

tip

A working combination: Liquidation Spike finds an instrument whose flow is already above normal, Liquidation Forecast shows whether it stays hot, and Liquidation Map shows how far to place the stop.