Cascade clustering

Concentration of liquidation prints into a local cascade.

Observed: read at the minute itself
Unit: ratio
Valid range: 0 to 1

In plain English

Summarizes uneven spacing between observed liquidation events. It is a smoothed timing estimator using up to 500 events, not a liquidation-volume total.

How to read it

The estimator transforms the variation in positive inter-event gaps and blends it into a state initialized at 0.8. More variable gaps push the estimate higher; regular spacing pushes it lower. It becomes present after 10 events, so early readings can still be dominated by initialization. A value of 0.5 is an index level, not 50% of liquidations in a cascade. Values stay below 1 by construction and hold through quiet periods.

Why a researcher might use it

Study whether liquidation timing is more clustered alongside the current 15-second intensity. A held high index and zero current intensity describe different aspects of the same history.

What it does not prove

This moment-based estimator is not a fitted Hawkes model or evidence that one liquidation caused another. It cannot establish a self-sustaining cascade, current activity or a future cascade probability. Publication throttling distorts the timing it observes.

Use in Research

What this cannot see

Binance throttles liquidation publications, so these readings describe the observed feed and miss some liquidation records. Count estimation is still under validation. It would require uncertainty, support and freshness checks, and would not recover hidden event times or validate a price-level heatmap or cascade forecast.

Technical details

Field ID

feature.cascade_clustering_index

Valid query operators

at least, at most, between

Window operators: lowest, highest, average, latest or change in, over a trailing window.

Editable query preset

This preset is a starting point, not a recommendation. Edit it before running the search.

{
  "schema_version": "research_query.v2",
  "normalization_version": "archive_normalization.v1",
  "feature_version": "feature_defs.v1",
  "target": "record_occurrences",
  "where": {
    "all": [
      [
        "feature.cascade_clustering_index",
        "gte",
        0.5
      ],
      [
        "times.anchor_time",
        "gte",
        "2025-07-15T00:00:00.000Z"
      ]
    ]
  },
  "sort": [
    "times.anchor_time",
    "desc"
  ],
  "page": {
    "limit": 5,
    "cursor": null
  }
}

The search reports how often the record held this condition and what followed. It does not decide whether an idea works.