Public state

Methodology

Version public-state/v1+liqpressure+canon1 · effective 18 August 2026

This page describes exactly what the public market state on riskloom.ai/live measures, how it is computed, and what it does not claim. If you are evaluating RiskLoom, read this before the numbers.

What the public state measures

The public state is a count.

Every five minutes, RiskLoom measures how many of the monitored perpetual futures markets have liquidation volume over the preceding fifteen minutes that exceeds that market's own calibrated threshold. Each market's threshold is derived from its own history, so a market with a thin book and one with a deep book are treated on the same basis.

The public state is a band placed on that count. Nothing else feeds it.

Association, not prediction. Across the measurement window, a higher count of markets in concurrent liquidation is associated with an increasing conditional probability that a qualifying cascade begins somewhere in the monitored universe within the following three hours. That is the entire claim.

It is not a forecast. It does not identify which market. It does not state that one market's liquidations cause another's. The relationship is an observed association between concurrent liquidation activity and near-term cascade frequency.

The states

The probability column is the observed frequency with which at least one qualifying cascade began anywhere in the monitored universe within three hours of a reading in that state.

StateMarkets in concurrent liquidation ObservationsShare of window Cascade began within 3h
QUIET018,90272.9%13.3%
BUILDING15,08119.6%19.3%
ELEVATED2 to 31,3695.3%31.0%
STRESSED4 or more5682.2%43.1%
25,920100%

Measured over 2026-05-20 00:00 to 2026-08-17 23:55 UTC — ninety complete days at five-minute resolution. The window is stated explicitly because the archive continues to grow and a later ninety-day window will produce different occupancy.

Read these as conditional probabilities, not signals. A cascade began within three hours of 13% of QUIET readings. Cascades occur during quiet states. The states describe how the odds shift, not whether an event will happen.

STRESSED also includes any reading where at least one market's liquidation volume exceeds its cascade threshold, regardless of the count.

The proportion of time spent in each state is not fixed. Across three consecutive thirty-day blocks the share of time in QUIET ranged from 64% to 88%. The ordering of the states was stable across all three blocks measured independently; the occupancy was not. Any statement of the form "the market is QUIET x% of the time" would be a property of a particular period, not of the method.

Per-market bands

Each market is shown in one of three bands, on the same fifteen-minute basis.

BandMeaning
NORMALLiquidation volume below this market's calibrated threshold
ELEVATEDAt or above this market's calibrated liquidation threshold
CASCADE PRESSUREAt or above this market's cascade threshold

The threshold values themselves are not published. They are calibrated per market from that market's own trailing distribution.

What a qualifying cascade is

A documented liquidation cascade in the RiskLoom archive: a period of sustained forced liquidation on a single market, exceeding that market's calibrated cascade threshold, recorded with its start, peak and end minute.

Every cascade in the archive is reconcilable to per-minute liquidation data. The full archive is browsable at riskloom.ai/replay.

Two kinds of evidence, kept separate

These are not equivalent and the site does not present them as such.

Archive reconstruction

Cascades documented from historical data, identified after the fact by applying the methodology to the recorded liquidation series. This is the depth of the archive — hundreds of events across dozens of markets. It demonstrates that the method identifies real events. It cannot demonstrate that a warning existed before an event, because the analysis was performed afterwards.

Prospectively timestamped record

From 18 August 2026, every public state is written to an append-only ledger at the moment it is computed, recording both the time of computation and the time of publication.

The ledger is append-only by design. The process that writes it holds insert permission and nothing else — it cannot update or delete a row. A database trigger rejects any update or deletion attempt, including from a privileged account. And each entry is cryptographically chained to the one before it, so removal or alteration of any historical entry breaks the chain and is detectable. This is tamper-evidence, not a claim that alteration is physically impossible.

An entry demonstrates that a state was recorded before whatever followed it. That is a materially stronger claim than a reconstruction, and it is why the two are labelled separately everywhere on the site.

The prospective record begins from zero. It accumulates.

Data freshness

The state is computed every five minutes from liquidation data that is itself a few minutes behind real time. Each ledger entry records how stale its inputs were at the moment of computation.

"Last updated" on the live page is derived from the publication time of the most recent entry, not from input freshness. A stale reading and a stopped process are different failures, and the site distinguishes them.

If inputs are too stale to compute a state honestly, the state is recorded as UNKNOWN rather than omitted. A gap in the record would be indistinguishable from downtime; a recorded UNKNOWN is self-documenting.

Coverage

The public state covers the monitored perpetual futures universe as of the most recent reading. The count of markets is shown alongside the state and changes as markets are added.

Markets enter the monitored universe once they are non-provisional with a computable threshold. A market that has not met that bar is not counted.

What is not published here

The numeric cascade risk index. Per-market threshold values. Calibration method and parameters. Confirmed signals and their timing.

The public state itself is available programmatically at api.riskloom.ai/public/v1/state, unauthenticated and rate-limited, returning exactly the fields rendered on /live and nothing further. That is a deliberately narrow public surface and is not the institutional API, which carries confirmed signals, per-market detail and history under evaluation.

These are the institutional product. The public state is a view of market structure, not a substitute for it.

Versioning

This methodology is versioned. Every entry in the public ledger records the version under which it was computed, so a state recorded today remains interpretable after the method changes.

public-state/v1+liqpressure+canon1 — effective 18 August 2026

Bands on count of markets in concurrent liquidation over a trailing fifteen minutes, against per-market calibrated thresholds. Boundaries at 0 / 1 / 2–3 / 4+, with cascade-threshold breach promoting to STRESSED.

Changes are additive, never retrospective. If the boundaries or the underlying calibration change, a new version is issued and applies from that date forward. Historical entries retain the version under which they were written. No entry is ever recomputed — the ledger records what was published, not what a later method would have published.

Method notes

Measurement window

2026-05-20 00:00 to 2026-08-17 23:55 UTC — ninety complete days, 25,920 five-minute observations, across the monitored universe. The window is pinned rather than relative, so the figures are reproducible.

Leakage

The outcome counts only cascades that began after the reading. An event already in progress cannot contribute to its own probability.

Out of sample

The per-market thresholds were calibrated from June 2026, so the majority of the measurement window falls outside the calibration period. The ordering of the states holds in both the in-sample and out-of-sample portions, and holds independently in each of three consecutive thirty-day blocks.

Boundary sensitivity

The ordering of the states survives moving the boundaries — tested at 0 / 1 / 2 / 3+ and 0 / 1–2 / 3–4 / 5+ as well as the published 0 / 1 / 2–3 / 4+. The boundaries were not selected to produce the result.

Questions on methodology, or a request for the underlying detail under evaluation: trials@riskloom.ai