Research · Altcoins
How BTC–alt correlation is measured
Last reviewed 2026-09-21Source: Methodology description; CoinGecko and Coin Metrics daily price seriesNo coefficient is reported on this page. Every figure would need its window, asset set and vintage stated.
The convention: rolling Pearson on daily log returns
Correlation is computed on returns rather than on prices. Two price series that both trend upward over a long sample will show a high correlation almost regardless of what they have in common, because the trend dominates the variation. Differencing the series removes the trend and leaves the day-to-day variation that the coefficient is meant to describe.
This site uses log returns: the natural logarithm of the ratio of consecutive daily closes. Log returns are additive across time, so a weekly return is the sum of five daily log returns, and they are better behaved for assets whose price can move by a large percentage in a day. The alternative — simple percentage returns — gives slightly different coefficients, particularly on large-move days, and a published figure is only reproducible if the convention is stated.
The coefficient itself is the Pearson correlation of the two return series over the window. It is bounded between minus one and one, it is symmetric, and it is unit-free. The window is what makes it "rolling": the coefficient is recomputed at each date using the trailing N days, so the series of coefficients shows how the relationship changed rather than collapsing it to a single number.
Why the window decides the answer
| Window | Observations | What it shows | What it costs |
|---|---|---|---|
| 30 days | 30 | Short-lived shifts, including the days around a single large market event | Very noisy; a handful of days can move the coefficient a long way, and the estimate is unstable |
| 90 days | 90 | A quarter's behaviour, long enough to smooth a single week but short enough to register a regime change | Still sensitive to a few outsized days; two adjacent windows can disagree sharply |
| 365 days | 365 | The prevailing relationship over a year, which is the figure most often quoted | Averages across regimes, so a genuine decoupling inside the year is hidden by the surrounding months |
Last reviewed 2026-09-21Source: Methodology descriptionObservation counts are the window length in trading days; crypto trades every day, so a 365-day window is a calendar year.
The window is the single largest source of disagreement between published correlation figures. Two analysts can compute the same statistic on the same assets over the same period and report different numbers, because one used a 30-day window and the other a year. Neither is wrong; they are answering different questions. A short window answers "what is happening now", and a long window answers "what has typically been true".
There is also a mechanical effect at the boundary. A rolling window drops one observation and adds one each day, so a single extreme day enters the calculation, sits in it for the length of the window, and then leaves. On a 30-day window a liquidation day can dominate the coefficient for a month; on a 365-day window its influence is diluted across the year. A reader comparing two charts should check that the windows match before concluding that the relationship itself changed.
What the coefficient does and does not say
A high coefficient says the two return series tended to move together over the window. It does not say which moved first, it does not say that one caused the other, and it does not say that the relationship will persist. It is also not a statement about levels: two assets can have a high return correlation while one has lost most of its value against the other, because correlation is computed on the shape of the daily changes rather than on the price path.
Correlation is also not the same as beta. Beta measures how much one asset tends to move for a given move in another, and it is scale-dependent; correlation measures how tightly the two move together, and it is not. An asset can have a high correlation with bitcoin and a beta well below one, meaning it moves in the same direction but by less. The companion page on beta to BTC covers that distinction.
Finally, a coefficient is an estimate from a sample, not a measured constant. With 30 observations the sampling error is wide enough that a coefficient of 0.6 and one of 0.8 are not reliably distinguishable. Reporting a correlation to two decimal places without the window and the sample size implies a precision the statistic does not have.
Dataset, period, method and limitations
Dataset. Daily closing prices for bitcoin and for the non-bitcoin assets under comparison, from CoinGecko's historical price endpoint, with Coin Metrics community data used as a cross-check where a series exists. Both aggregate across venues, so a venue-specific dislocation is smoothed before the return is computed.
Period. The window actually used must be stated with any figure. Because the large alts have different listing dates, a comparison across a fixed calendar window necessarily uses a different number of observations for each asset, and that difference should be disclosed rather than hidden.
Method. Pearson correlation of daily log returns over a trailing window, recomputed at each date. Log returns are the natural log of consecutive closing prices. No adjustment is made for non-synchronous closes, which is a real limitation for assets whose deepest market closes at a different hour from bitcoin's.
Limitations. Daily closes discard intraday sequence, so a same-day lead-lag relationship is invisible. Correlation is sensitive to outliers and to the window length, and it is not robust to regime change: a single coefficient computed over a multi-year sample describes an average that may never have held at any point in the sample. And because correlation is computed on returns, it says nothing about the relative performance of the two assets.
Sources and references
The measurement convention follows standard practice in quantitative finance; the data sources are the same aggregations used elsewhere on this site.
- Historical price data. CoinGecko, CoinGecko API documentation: the historical daily price endpoint used for both series.
- Cross-check series. Coin Metrics, Community Network Data: free daily reference rates used to sanity-check the aggregated closes.
- Correlation and its interpretation. NIST/SEMATECH, e-Handbook of Statistical Methods — Measures of Association: the definition of the Pearson coefficient and its sensitivity to sample size and outliers.
- Volatility and its windows, for context. Bitcoin Data Guide, Volatility Explained: why a volatility figure without its window says very little, which applies equally to a correlation figure.
Related reading
- Research HubEvery dataset on the site, with methodology and provenance.
- Altcoin ResearchAltcoins measured against Bitcoin: design intent, consensus, execution, scaling and market structure.
- The ETH-BTC Correlation RecordHow the correlation is measured, how it behaves across windows, and where it breaks down.
- The ETH/BTC RatioWhat the ratio measures, how to read its trend, and why it is not a forecast.
- ETH During Bitcoin Bull PhasesAssociation within a common market factor, and what co-movement cannot establish.
- ETH During Bitcoin Bear PhasesDrawdown depth and duration compared over identical windows, and the limits of the comparison.