Skip to content

FRM Exam Part II · Empirical Properties of Correlation: How Do Correlations Behave in the Real World?

Correlation Distributions, Skewness and Persistence

Updated 11 October 2026 · Fact-checked

Observed correlations are not constant. Over time they form a distribution that is bounded between -1 and +1, often skewed and clustered, and they persist: high-correlation periods tend to follow high-correlation periods. You test this with histograms, skewness, autocorrelation of correlation changes, and stability tests across rolling windows.

Understand Correlation Distributions, Skewness and Persistence

A correlation estimated from data is a random variable. If you compute 60-day rolling correlations between two equity indices for ten years, you get a series of numbers. That series has a mean, a spread, a shape and a memory. This topic is about those four features.

The distribution is limited to the range -1 to +1. Because of this boundary, a distribution with a high mean (say 0.7) cannot spread freely above it. It piles up near +1 and has a long tail on the lower side. That is negative skewness. Equity correlations are typically high on average and negatively skewed. A low-mean distribution would show the opposite pattern.

Persistence means correlations have memory. The autocorrelation of a correlation series is the correlation between today's value and the value k periods earlier. Rolling-window correlations show high autocorrelation, partly because overlapping windows share data. So persistence in a rolling series is partly mechanical, not only economic. Correlation *changes* are usually far less persistent, and often show negative autocorrelation because of mean reversion.

Clustering means that high-correlation episodes bunch together, as in crises when many assets fall together. This resembles volatility clustering. It matters for risk: a model that assumes a constant correlation underestimates joint losses when correlations are in a high regime.

Stability testing asks whether the correlation in one sub-period equals that in another. You can compare rolling estimates, split the sample, or test the difference using the Fisher z-transform. The standard error of a sample correlation is large for short windows, so some apparent instability is just estimation noise.

Key formulas to remember

Sample correlation
ρ̂ = Cov(X, Y) ÷ (σX × σY)
Always between -1 and +1. Estimated over a chosen window, so it changes with window length.
Skewness of a correlation series
Skew = E[(ρ − μ)³] ÷ σ³
Negative means a long tail toward lower correlations. Expect this when the mean correlation is high.
Autocorrelation at lag k
AC(k) = Cov(ρt, ρt−k) ÷ Var(ρ)
Applied to a correlation series. Close to +1 means strong persistence. Rolling windows inflate it through overlap.
Fisher z-transform
z = 0.5 × ln[(1 + ρ) ÷ (1 − ρ)]
Makes the sampling distribution of ρ approximately normal, for approximately bivariate normal data.
Standard error of z
SE(z) = 1 ÷ √(n − 3)
n is the number of observations. Shorter windows give wider confidence bands.
Test of equal correlations (two independent samples)
Z = (z1 − z2) ÷ √[1 ÷ (n1 − 3) + 1 ÷ (n2 − 3)]
Compare with the standard normal critical value, such as 1.96 at 5% two-sided.

How to solve Correlation Distributions, Skewness and Persistence questions

Use this sequence for any question on the shape, memory or stability of correlations.

  1. 1Identify what is given: a mean correlation, a series of rolling values, autocorrelations, or two sub-period estimates with sample sizes.
  2. 2Check the boundary. If the mean is near +1, expect negative skew. If near -1, expect positive skew. If near 0, expect little skew.
  3. 3For persistence questions, check whether the series is levels or changes. Levels of rolling correlations are highly autocorrelated. Changes are much less so.
  4. 4Note the window overlap. High autocorrelation in rolling data partly reflects shared observations.
  5. 5For stability tests, apply the Fisher z-transform to each estimate, then compute the standard error using n − 3.
  6. 6Compute the test statistic and compare to the critical value. State whether you reject equal correlation.
  7. 7Interpret for risk: persistent high-correlation regimes mean diversification benefit cannot be assumed constant, and stress correlations should be used.

Quickest way: Boundary, memory, noise

When to use it: For conceptual MCQs where you need an answer in under a minute.

  1. Boundary: high mean correlation means negative skew; low mean means positive skew.
  2. Memory: rolling-window correlation levels are persistent; changes are weakly persistent or mean reverting.
  3. Noise: a short window gives a wide confidence band, so not every change is real instability.
  4. Eliminate options that claim correlations are constant, normally distributed, or independent over time.

Common mistakes in Correlation Distributions, Skewness and Persistence

  • Saying correlation distributions are symmetric or normal.

    Students carry over the normal assumption from returns.

    Fix: Remember the bounds of -1 and +1. A high mean forces a long lower tail, so skew is negative.

  • Linking positive skew to high-mean correlations.

    Confusing the direction of the tail with the location of the mass.

    Fix: Skew follows the long tail. Mass near +1 with a tail to the left is negative skew.

  • Treating high autocorrelation of rolling correlations as pure economic persistence.

    Overlapping windows are ignored.

    Fix: Overlap creates mechanical autocorrelation. Check non-overlapping windows or correlation changes.

  • Using n instead of n − 3 in the Fisher standard error.

    Mixing it up with simple standard error formulas.

    Fix: Write SE(z) = 1 ÷ √(n − 3) every time.

  • Testing the difference of raw correlations instead of z values.

    Skipping the transform to save time.

    Fix: Transform each ρ to z first, then take the difference.

  • Concluding correlations are unstable from one noisy short-window change.

    Ignoring estimation error.

    Fix: Compare the change with its standard error before calling it a regime shift.

Worked examples

Example 1

Over 103 daily observations in a calm period, the correlation between two equity indices is 0.50. Over 103 daily observations in a stress period it is 0.80. Assuming independent samples, test at 5% (critical value 1.96) whether the correlations are equal.

Show the solution
  1. z1 = 0.5 × ln(1.5 ÷ 0.5) = 0.5 × ln 3 = 0.5 × 1.0986 = 0.5493.
  2. z2 = 0.5 × ln(1.8 ÷ 0.2) = 0.5 × ln 9 = 0.5 × 2.1972 = 1.0986.
  3. Difference = 1.0986 − 0.5493 = 0.5493.
  4. Standard error = √[1 ÷ 100 + 1 ÷ 100] = √0.02 = 0.14142.
  5. Z = 0.5493 ÷ 0.14142 = 3.88.
  6. 3.88 > 1.96, so reject equal correlations.

Answer: Z ≈ 3.88, so reject the hypothesis of equal correlations at 5%. The stress-period correlation is significantly higher.

Example 2

A risk team computes 60-day rolling correlations between two banks' stock returns over many years. The mean is 0.75 and the series has a lag-1 autocorrelation of 0.98 using daily steps. Describe the likely skewness and explain how to read the autocorrelation.

Show the solution
  1. The mean of 0.75 is close to the upper bound of +1, so values cannot spread much above it.
  2. Occasional drops to low correlation create a long left tail, so skewness is negative.
  3. A lag-1 autocorrelation of 0.98 with daily steps and a 60-day window means consecutive windows share 59 of 60 observations.
  4. So most of that persistence is mechanical overlap. It does not prove strong economic memory on its own.
  5. To test genuine persistence, use non-overlapping windows or examine correlation changes.

Answer: Skewness is likely negative. The very high autocorrelation is largely caused by window overlap, so confirm persistence with non-overlapping windows.

Exam tips

  • When a question gives a mean correlation, use the boundary argument to name the skew direction first.
  • Watch for overlapping windows in any autocorrelation question. It is a favourite trap.
  • For the Fisher test, write z, then SE using n − 3, then the statistic. Do not skip steps.
  • Link results to risk: persistent high correlation in stress means diversification fails when it is needed most.
  • Reject options that say correlations are constant or independent through time.

Practice questions from Empirical Properties of Correlation: How Do Correlations Behave in the Real World?

Correlation Distributions, Skewness and Persistence: frequently asked questions

Why are equity correlations negatively skewed?

Their average level is high, so they sit close to the upper bound of +1. Values can only fall away on the lower side, creating a long left tail.

What is autocorrelation of correlation?

It is the correlation between a correlation estimate and its own earlier value, at some lag. High values mean correlations are persistent, and clustering of high-correlation periods is likely.

How do I test whether correlation is stable over time?

Estimate correlation in different sub-periods or rolling windows, apply the Fisher z-transform, and test the difference using standard error based on n − 3. Also compare with the estimation noise for short windows.

Does correlation clustering matter for VaR?

Yes. If high-correlation regimes cluster, a VaR model using a long-run average correlation understates joint losses in those regimes. Stress correlations or regime-aware models help.