Live prices are currently unavailable — the exchange feed could not be reached and no recent cached reading is held.

Bitcoin Data GuideBitcoin Reference

Research · Altcoins

Do correlations rise in market stress?

The claim that correlations converge toward one in a crisis is one of the most repeated statements in market commentary. It is testable, it has a mechanism behind it, and it comes with a caveat that is usually dropped: a rise in correlation is not evidence that one asset caused another's move.

Last reviewed 2026-09-21Source: CoinGecko historical daily prices; Coin Metrics community network dataMethod and test design only. No coefficient is asserted without its window, asset set and stress-window definition.

The claim and its mechanism

The claim is that assets which are only loosely related in calm markets move together when markets are stressed. The mechanism is straightforward. In a stress event, the dominant flow is not asset-specific: it is deleveraging, redemption or a risk reduction that is applied across a portfolio. When the same seller is selling everything, the asset-specific factors that normally differentiate the assets are overwhelmed by the shared one, and the measured correlation rises.

A second mechanism reinforces the first. Liquidity withdrawal is correlated across assets, because the same market makers reduce size everywhere at once. When depth falls, price impact rises, and the common selling pressure produces larger moves in every asset simultaneously. The rise in correlation is partly a rise in the common component and partly a fall in the noise that normally differentiates the assets.

The claim is testable, which is what makes it worth stating carefully. It predicts a specific pattern: correlation measured over a stress window should be higher than correlation measured over a calm window for the same assets. If that pattern does not appear, the claim is wrong for this market, and the page should say so.

Named stress windows

The stress windows used to test whether alt-Bitcoin correlation rises, what defined each window, and the caveat that applies to each.
WindowHow it is definedCaveat
March 2020The broad risk-asset selloff in the second week of March, defined by a stated decline in bitcoin over a fixed number of daysThe window is short, so the estimate has wide sampling error and a few days dominate the result
May 2021The decline following the mid-May peak, defined by a stated peak-to-trough move over a fixed windowThe event had asset-specific components, so a rise in correlation is partly a shared reaction to shared news
June 2022The credit-driven decline in the second quarter, defined by a stated peak-to-trough moveThe failure of a specific lender affected assets with direct exposure differently, which cuts against a uniform rise
November 2022The exchange failure and its aftermath, defined by a stated peak-to-trough moveThe event was concentrated in one venue's assets, so the window tests contagion as much as correlation

Last reviewed 2026-09-21Source: Market-history description; CoinGecko daily series for the underlying pricesWindows are defined by a stated rule so the test is reproducible; the boundaries depend on the rule chosen.

Correlation is not causation

A rise in correlation during stress is exactly what the common-factor mechanism predicts, and it is also what a causal story predicts. The two are not distinguished by the coefficient. If bitcoin's decline caused the alts' decline, correlation would rise; if a shared factor drove both, correlation would also rise. The measurement cannot tell them apart.

There is a further trap specific to stress windows. The windows are selected after the fact, because a stress event is only identifiable once it has happened. Selecting the window on the outcome biases the test toward finding the effect: any period chosen because prices fell together will show elevated correlation, whether or not stress systematically raises it. A defensible test defines the window by an observable rule — a stated decline over a stated number of days — rather than by the narrative of the event.

The honest conclusion is therefore narrow. The data can show whether correlation was higher in the named windows than in the surrounding calm periods. It cannot show that stress caused the rise, and it cannot show that bitcoin caused the alts' moves within the window. Where this site reports the pattern, it reports it as a description of the window and stops there.

Dataset, period, method and limitations

Dataset. Daily closing prices for bitcoin and for the largest non-bitcoin assets by market capitalisation, from CoinGecko's historical price endpoint, with Coin Metrics community data as a cross-check. Aggregated closes across venues.

Period. The stress windows listed above, each compared against a calm window of the same length immediately preceding it. Matching the length is what makes the comparison meaningful, because the sampling error of a correlation estimate depends on the number of observations.

Method. Pearson correlation of daily log returns, computed separately over the stress window and over the matched calm window, for the same asset set. The window definition rule is stated before the result is reported, so the boundary is not chosen to produce a particular answer.

Limitations. Stress windows are short, so every estimate carries wide sampling error and a few days dominate. Window selection is inherently retrospective and biases the test toward finding the effect. Daily closes discard intraday sequence, which is the resolution at which a contagion mechanism would be visible. And the asset set changes over time, so a comparison across windows years apart is not computed on an identical basket.

Sources and references

The test design follows standard practice for correlation studies; the data sources are the aggregations used elsewhere on this site.

Related reading