Crypto correlation is a measurement, not a property. It describes how two assets moved together over a chosen window, and it says nothing about whether they will move together tomorrow. Portfolios built on the assumption that it is stable tend to discover the error at the worst possible moment: during a drawdown, when the diversification was supposed to pay.
The pattern is old and well documented across asset classes. In calm markets, returns diverge and crypto correlation between large assets drifts lower, which flatters any diversification test run on that data. In a sharp sell-off, the same assets converge and the coefficient moves toward one. The risk that was supposed to be spread out arrives all at once.
- Correlation is regime-dependent. A coefficient estimated in a quiet month is a statement about that month, not a risk parameter.
- Convergence is asymmetric. Assets tend to fall together more reliably than they rise together, so diversification helps least when it is needed most.
- Liquidity is the transmission channel. In stress, sellers raise cash in whatever is liquid, which is why unrelated assets get sold in the same hour.
- Test the bad window. Measure on drawdown periods, not on the full sample that averages calm and crisis into a comfortable number.
What crypto correlation actually measures
It measures how closely two return series move together over a sample, between −1 and +1. A coefficient near zero means no linear relationship in that sample. It does not mean independence, and it does not mean safety: two assets can be uncorrelated on average and still crash on the same day.
Daily returns are the usual input, though weekly and hourly windows are common. Hourly data captures the intraday stampede; daily smooths it; weekly can hide it entirely. A figure quoted without a stated window is close to meaningless.
The coefficient, not the story
Two tokens can share a narrative and still show a modest coefficient in a calm market, because narrative is not a cash flow link. The reverse holds too: assets with no shared story can correlate tightly because the same marginal buyer funds both. Correlation measures flows and shared exposure, not resemblance.
Why the sample window matters
A 90-day window containing one violent week reports a high coefficient; a 365-day window containing the same week reports a moderate one. Neither is wrong, and both mislead alone. Compute the coefficient on several windows, and again only on days when the market fell hard. If the drawdown figure sits far above the full-sample figure, the portfolio carries a break the headline number conceals.
Why crypto correlation rises in a drawdown
It rises because the marginal seller is the same in every market. A fund facing redemptions or a margin call does not sell its best idea; it sells what it can, quickly, at a price the book will absorb. That transmits a shock from one asset to another without any fundamental link between them.
Three mechanisms do the work. Leverage: when collateral falls, positions across the book are liquidated together, and liquidations are price-insensitive. Collateral plumbing: funding stress in one place forces deleveraging everywhere that asset is posted. Market making: quotes widen and size shrinks at the same moment, so liquidity that looked independent turns out to be shared.
Convergence is not symmetric
Assets commonly fall together and rise apart, and the reason is mechanical rather than psychological. Losses force selling; gains do not force buying. A holder who is up can do nothing, while a holder facing a margin call cannot. That asymmetry is why realised diversification in a downturn is worse than the average suggests.
The correlation is not the risk
A high coefficient tells you positions moved together in the past. It does not tell you the size of the next move, the depth of the book, or whether you can exit. Two assets with a coefficient of 0.9 and thin liquidity are more dangerous than two with the same coefficient and deep books. Correlation describes co-movement; liquidity describes the price of being wrong.
How behaviour changes by regime
The table below is a framework, not a set of measurements. Actual coefficients vary by asset pair, venue, window and period, and should be computed from data rather than assumed.
| Regime | Typical pairwise behaviour | What diversification does | Practical implication |
|---|---|---|---|
| Calm, range-bound | Coefficients drift lower and disperse | Appears to work well | Backtests overstate the benefit |
| Trending up | Moderate and stable, with rotation | Works partially, lags the leader | Diversification costs return, not capital |
| Sharp drawdown | Converges toward one across majors | Largely stops working | Size positions on the assumption of one bet |
| Post-liquidation recovery | Disperses again, unevenly | Resumes with a lag | Correlation falls before confidence returns |
| Venue or plumbing stress | Rises across unrelated pairs | Fails entirely | Shared custody and collateral dominate |
Diversification is a claim about the future made from data about the past. The claim is weakest exactly when the market is loudest.
Testing a portfolio for correlation breaks
Testing is a repeatable procedure, and running it once is not testing. Rebuild the numbers on a fixed schedule and after every significant move.
- Collect daily returns for each holding over at least two years, or the longest common history available, and note the gaps.
- Compute the pairwise coefficient on the full sample, then on a rolling 90-day window, and plot both.
- Split the sample into days when the market index fell more than 5% and days when it rose more than 5%, then compute coefficients separately for each subset.
- Compare the drawdown coefficient with the full-sample figure. A large gap is the break you are trying to find.
- Recompute using a stress proxy: replace each asset’s return with its worst 10% of days and re-run the same calculation.
- Aggregate the result into a single figure, such as the share of portfolio variance contributed by the largest common factor, and track how it moves.
- Translate the stressed figure into a position size, and check whether the current book would survive a move of that size with the leverage actually in use.
Step seven is where most analyses stop being useful. A matrix is only an input; the output has to be a position size, a leverage ratio or a stop distance the book can honour under stress.
What to do about it
Do not try to forecast the break. Assume it, and size for it. The adjustments are unglamorous and they work: fewer simultaneous positions, smaller notional per position, collateral held outside the correlated cluster, and a written rule for what gets cut first when everything falls together.
Diversify the failure modes, not the tickers
Holding eight tokens on one venue with one custodian and one collateral asset is one position with eight names. Real diversification separates custody, venue, collateral type and strategy horizon. Those separations survive a break because they are not return series at all — they are structural differences in what can go wrong.
Keep a liquidity reserve outside the book
The cheapest hedge against convergence is cash you can deploy without selling a correlated asset. A reserve in a stable instrument makes a drawdown a decision rather than a forced sale, and forced sales are the mechanism that produces the convergence you were trying to avoid.
0 comments
No comments yet — ask a question and we will answer it here.