A correlation matrix that says two assets move together at 0.8 feels like it settled something. It rarely did. Most people building those matrices reach for Pearson because it is the default in every spreadsheet and library, and then they trust the number without asking what it actually measured. Pearson measures how close two return series sit to a straight line. That is a strong assumption to make about crypto and macro, where a single ugly day can drag the whole estimate around.
I ran into this the boring way. Two tokens looked tightly correlated in the matrix, so I sized a hedge around it, and the hedge behaved nothing like the number promised. When I pulled the scatter apart, most days the two barely tracked each other. What was holding the correlation up was a handful of shared crash days where everything sold off at once. Pearson saw those days and decided the relationship was strong. On any normal Tuesday, it was not.
What each one is actually measuring
Pearson correlation is about magnitude. It takes the actual return values, and a day where one asset drops a lot and the other drops a lot pulls the coefficient up hard. Because it works on raw magnitudes, the biggest days carry the most weight. In a fat-tailed series, the biggest days are the outliers, so your correlation is disproportionately a story about the tails, whether you wanted it to be or not.
Spearman throws the magnitudes away. It ranks every day from smallest return to largest, then runs the same correlation math on the ranks instead of the values. So the question changes from "when this asset moves a lot, does the other move a lot in proportion" to "when this asset has a relatively big day, does the other tend to have a relatively big day too." A crash that is three times worse than the second-worst day is still just the worst-ranked day. Its outlier size stops mattering. It counts once.
That single change fixes the specific failure I keep hitting. Because ranks compress the tails, no single session can dominate the estimate. You also stop assuming linearity. Two assets can move together consistently in a curved, non-proportional way, and Pearson will underrate that relationship while Spearman picks it up, because Spearman only cares about the ordering, not the shape of the line.
When the gap between them tells you something
The useful move is to compute both and look at the spread. If Pearson and Spearman land close together, the relationship is roughly linear and reasonably clean, and you can trust either. When they diverge, the difference is the signal.
- Pearson much higher than Spearman usually means a few extreme joint days are inflating your number. On a typical day the co-movement is weaker than the headline. This is the trap I fell into, and it is common in crypto pairs that mostly do their own thing but crash together.
- Spearman much higher than Pearson usually means the assets do track each other most days, but the relationship is non-linear or a couple of odd sessions are dragging Pearson down. The real co-movement is stronger than Pearson admits.
- Both low and close together means they genuinely are not moving together, and no amount of measure-swapping will change that.
A rough tell I use: if the two coefficients are within roughly 0.1 of each other, I stop worrying and move on. When they split by more than that, I pull up the scatter plot before I trust either number, because the gap is telling me the distribution has something Pearson cannot handle on its own.
Which one belongs in your matrix
For most cross-asset work in markets with fat tails, I default to Spearman for the co-movement view and keep Pearson around as a diagnostic. The reasoning is practical rather than ideological. I want a correlation number that describes how these things behave on ordinary days, not one that is secretly a summary of three crash sessions. Spearman gives me that, and it does not fall over when one asset has a return distribution that looks nothing like a bell curve.
The honest exception is when magnitude is the whole point. If you are building something where the size of the joint move matters directly, like sizing a hedge ratio or estimating portfolio variance, Pearson is the right tool because variance and covariance are defined on magnitudes, not ranks. Spearman will not give you a hedge ratio. So the split I land on is roughly this: use Spearman to decide whether two things move together, and use Pearson, carefully and with your eyes open to the tails, when you need the actual sensitivity for sizing.
A short workflow that has saved me from bad matrices:
- Compute both Pearson and Spearman for every pair, not just one.
- Flag any pair where the two disagree by more than roughly 0.1 and look at its scatter before trusting the relationship.
- Check whether a small number of days are carrying a high Pearson value. If dropping the two or three most extreme sessions collapses the correlation, that correlation is a tail artifact, not a stable relationship.
- Pick your measure by the job. Co-movement view goes to Spearman. Sizing goes to Pearson, with the tails checked.
None of this makes correlation trustworthy on its own. Correlation is unstable across regimes, it tends to spike toward one exactly when you need diversification most, and both measures are backward-looking. But swapping in the rank-based version removes one specific way the number lies to you, which is a single bad day pretending to be a relationship. That is a cheap fix, and it is the first thing I reach for when a matrix tells me two assets are joined at the hip and my gut says they are not.