Risk Modelling and Survival Analysis · Introduction to copulas

Dependence and Limitations of Correlation in Risk Modelling

Updated 11 October 2026 · Fact-checked

Linear (Pearson) correlation measures only the strength of a straight-line relationship. It can miss non-linear dependence, depends on the marginal distributions, and may not exist for heavy tails. Rank correlations, Spearman's rho and Kendall's tau, use only ranks, so they are unchanged by increasing transformations and capture monotonic dependence.

Understand Dependence and Limitations of Correlation

Start with the idea of dependence. Two risks are dependent if knowing one changes what you expect of the other. In risk modelling, claims from different lines, or losses in different markets, are often dependent. Ignoring this understates the risk of large combined losses.

The usual measure is the linear correlation coefficient ρ(X, Y) = Cov(X, Y) ÷ (σ_X σ_Y). It lies between -1 and 1. It is a good measure when the variables are jointly normal, or more generally elliptical. Outside that setting it has real limits.

The main limitations are these:
- Independence gives zero correlation, but zero correlation does not give independence. If X is standard normal and Y = X², then Cov(X, Y) = E[X³] - E[X]E[X²] = 0, yet Y is fully determined by X.
- It only captures linear dependence. A perfect non-linear but monotonic relationship gives a correlation below 1.
- It is not invariant under non-linear increasing transformations. Corr(X, Y) is usually not equal to Corr(ln X, ln Y).
- It depends on the marginals. Given two marginal distributions, not every value in [-1, 1] can be attained.
- It requires finite variances. For very heavy-tailed losses, such as a Pareto with a small shape parameter, it may be undefined.
- It says little about tail behaviour. Two models with the same correlation can have very different chances of joint extreme losses.

Rank correlation fixes several of these. It uses only the ordering of the data, not the actual values. Spearman's rho is the Pearson correlation of the ranks. Kendall's tau compares pairs of observations and counts how many are in the same order (concordant) and how many in opposite order (discordant). Both are unchanged when you apply a strictly increasing transformation to X or Y. Both equal 1 for perfect increasing dependence and -1 for perfect decreasing dependence. They depend only on the copula, not the marginals, which is why they appear in this chapter.

Rank correlations still have a limit. A single number cannot describe the whole dependence structure, and it still says little about the tails. That is the motivation for copulas.

Key rules to remember

Linear correlation: ρ(X, Y) = Cov(X, Y) ÷ (σ_X σ_Y); Needs finite, non-zero variances. Measures linear dependence only. Takes values in [-1, 1].
Sample covariance form: r = Sxy ÷ √(Sxx Syy), where Sxy = Σxy - (Σx)(Σy)/n; Use the same form for Sxx and Syy. This is the sample Pearson coefficient.
Spearman's rho (no ties): ρ_S = 1 - 6 Σd² ÷ (n(n² - 1)); d is the difference between the ranks of each pair. This shortcut holds only when there are no ties. With ties, compute Pearson correlation on the ranks.
Spearman's rho (definition): ρ_S(X, Y) = ρ(F_X(X), F_Y(Y)); The linear correlation of the probability-integral transforms (the ranks scaled to the unit interval).
Kendall's tau (sample, no ties): τ = (c - d) ÷ (n(n - 1)/2); c is the number of concordant pairs and d the number of discordant pairs, out of n(n - 1)/2 pairs.
Kendall's tau (population): τ = P(concordant) - P(discordant); For two independent copies (X₁, Y₁) and (X₂, Y₂) of the pair. Pair is concordant if (X₁ - X₂)(Y₁ - Y₂) > 0.
Invariance property: If T₁, T₂ are strictly increasing, then ρ_S(T₁(X), T₂(Y)) = ρ_S(X, Y) and τ(T₁(X), T₂(Y)) = τ(X, Y); Pearson correlation does not have this property in general.

How to solve Dependence and Limitations of Correlation questions

Use this method for any question on the limitations of correlation or on calculating rank correlation.

1Read what is asked: a comment on a measure, a calculation of Pearson, Spearman or Kendall, or a comparison of the three.
2If it is a comment question, name the property involved: linearity, dependence on marginals, invariance, existence of variance or tail behaviour. Support it with a short example.
3If it is a calculation, rank each variable from smallest to largest. Give tied values the average of the ranks they would occupy.
4For Spearman, find d = rank difference for each pair, then Σd². Apply 1 - 6Σd² ÷ (n(n² - 1)) if there are no ties.
5For Kendall, compare every pair of observations. Count concordant pairs c and discordant pairs d. Check that c + d = n(n - 1)/2 when there are no ties.
6Compute the coefficient and check it lies between -1 and 1 and has a sensible sign.
7Interpret the result in context. Say what it shows about monotonic dependence and what it does not show, such as tail dependence.
8State any assumptions, such as no ties, so the examiner sees the formula used was valid.

Quickest way: Rank, difference, sum

When to use it: Use this for small data sets (about 5 to 10 pairs) in a calculation question with no ties.

Write the data in a table with columns for x, y, rank of x and rank of y.
Rank x and y from smallest (rank 1) to largest.
Add a column d = rank x - rank y and a column d².
Sum d² and plug into 1 - 6Σd² ÷ (n(n² - 1)).
For Kendall, sort the pairs by x. Then count, for each y, how many later y values are larger (concordant) and smaller (discordant).
Check c + d = n(n - 1)/2 before you compute τ.

Common mistakes in Dependence and Limitations of Correlation

Saying zero correlation means the variables are independent.
It is true for the multivariate normal, so students assume it always holds.
Fix: State that independence implies zero correlation, not the reverse. Quote the X and X² example when asked for a counter-example.
Using the 6Σd² shortcut when there are tied values.
The formula is memorised without its condition.
Fix: Check for ties first. If there are ties, use average ranks and compute the Pearson correlation of the ranks.
Ranking in opposite directions for x and y.
Rushing, or ranking one column from largest and the other from smallest.
Fix: Rank both variables from smallest to largest. Reversing one only changes the sign.
Claiming Pearson correlation is unchanged by any transformation.
Confusing it with its invariance to linear changes of scale.
Fix: Pearson is invariant only under strictly increasing linear transformations (positive slope). Spearman and Kendall are invariant under any strictly increasing transformation.
Treating a high rank correlation as proof of strong tail dependence.
Students assume one number describes the whole dependence structure.
Fix: Say that rank correlation measures overall monotonic association. Tail dependence is a separate property and needs its own measure.
Miscounting pairs in Kendall's tau, so that c + d does not equal n(n - 1)/2.
Comparing pairs in a disorganised way.
Fix: Sort by x first and compare each observation with all those after it. Always check the total number of pairs.

Worked examples

Example 1

Five pairs of claim sizes (in ₹ lakh) from two portfolios are: (2, 5), (4, 9), (7, 6), (9, 12), (12, 20). Calculate Spearman's rank correlation coefficient.

Show the solution

There are no ties, so use ρ_S = 1 - 6Σd² ÷ (n(n² - 1)) with n = 5.
Rank the x values 2, 4, 7, 9, 12 as 1, 2, 3, 4, 5.
Rank the y values: 5 is 1, 6 is 2, 9 is 3, 12 is 4, 20 is 5. So the y values in the original order 5, 9, 6, 12, 20 have ranks 1, 3, 2, 4, 5.
Rank differences d = 0, -1, 1, 0, 0, so d² = 0, 1, 1, 0, 0 and Σd² = 2.
n(n² - 1) = 5 × 24 = 120.
ρ_S = 1 - 6 × 2 ÷ 120 = 1 - 0.1 = 0.9.

Answer: Spearman's rho = 0.9, showing strong positive monotonic dependence.

Example 2

Using the data (1, 3), (2, 1), (3, 2), (4, 4) for (x, y), calculate Kendall's tau. Then explain why Kendall's tau for (ln x, ln y) would be the same, but the Pearson correlation would generally differ.

Show the solution

There are n = 4 observations, so n(n - 1)/2 = 6 pairs. The data are already sorted by x.
Compare y values pair by pair. Pair 1-2: 3 then 1, discordant. Pair 1-3: 3 then 2, discordant. Pair 1-4: 3 then 4, concordant.
Pair 2-3: 1 then 2, concordant. Pair 2-4: 1 then 4, concordant. Pair 3-4: 2 then 4, concordant.
So c = 4 and d = 2. Check: 4 + 2 = 6.
τ = (c - d) ÷ 6 = (4 - 2) ÷ 6 = 1/3.
The logarithm is strictly increasing on positive values, so it preserves the order of x values and of y values. Every pair keeps its concordant or discordant status, so τ is unchanged at 1/3.
Pearson correlation depends on the actual values, and ln is a non-linear transformation, so the covariance and variances change and the coefficient generally differs.

Answer: Kendall's tau = 1/3. It is unchanged under the log transform, whereas the Pearson coefficient generally changes.

Exam tips

Learn the standard list of limitations of linear correlation. Written questions often ask you to 'discuss' them, so give four or five points with a brief reason each.
In calculations, show the rank table. Marks are given for correct ranks even if the final arithmetic slips.
Always state whether you used the no-ties shortcut and that it is valid for the data given.
MCQs often test invariance. Remember: Spearman and Kendall are unchanged by strictly increasing transformations of either variable; Pearson is not.
Link the topic to copulas. Rank correlations depend only on the copula, which is why they are used to fit and compare copulas.

Practice questions from Introduction to copulas

Dependence and Limitations of Correlation in other exams

The same ground in other exams, if you are preparing for more than one or want another angle on it.

Dependence and Limitations of Correlation: frequently asked questions

Why is Pearson correlation not enough for risk modelling?

It measures only linear dependence and needs finite variances. It also depends on the marginal distributions and says little about joint extreme losses. Risk models need to capture how large losses occur together, which a single linear coefficient cannot show.

What is the difference between Pearson and Spearman correlation?

Pearson uses the actual values and measures linear association. Spearman uses only ranks and measures monotonic association. Spearman is unchanged by strictly increasing transformations of the data, while Pearson is not.

Kendall's tau vs Spearman's rho: which should I use?

Both are rank measures with the same invariance property, and both depend only on the copula. Kendall's tau has a direct interpretation as P(concordant) minus P(discordant). Spearman's rho is easier to compute by hand for small samples without ties. Their values generally differ for the same data.

Can two variables have zero correlation but still be dependent?

Yes. If X is standard normal and Y = X², the covariance is zero because E[X³] = 0, yet Y is a function of X. Zero correlation rules out only linear dependence.