Skip to content

New to Omni?·Get 20% off your first purchase with code WELCOME20 at checkout.See plans

FRM® paired data: why do two sample means move together?

Two sample averages can be precise yet share estimation error. Learn how pairing affects covariance, a difference of means, and FRM Part I interpretation.

ConceptOmni Curriculum Team6 min read

Two columns of blank paired data cards on a navy desk, with pencils held beside the rows.
On this page

The two averages in a paired dataset are calculated separately, but their estimation errors can move together because each pair was observed in the same period. For FRM® Part I, the useful distinction is between finding each mean and describing their joint sampling behaviour. The covariance between the two sample means is the covariance of a single pair divided by the number of independent pairs.

Start with the unit of observation

Imagine measuring the returns of two desks at each month end. A row contains the return for desk X and the return for desk Y from the same month. The first average uses only X values; the second uses only Y values. Nothing about the arithmetic of an individual mean changes because another column exists. The pairing matters when a question asks how the averages vary across hypothetical repeated samples, or how uncertain their difference is. GARP places quantitative analysis among the tools tested in FRM Part I, and its learning objectives specify the readings and objectives for that exam.

A common mistake is to interpret “joint” as a request to average all observations in both columns into one number. That would answer a different question, and might even combine quantities with different meanings. Keep the two estimates separate. Then ask whether the same unusually good or bad periods tend to affect both. If so, the two estimation errors are related, even though the means were calculated separately.

A paired sample also has a clear sample size. If twelve months yield one X observation and one Y observation each, there are twelve pairs. There are not twenty-four independent time observations. Counting both columns as independent additions to the sample size would make both standard errors artificially small. This distinction is particularly useful when the question describes two strategies observed over the same dates rather than two unrelated groups.

What the covariance matrix adds

The joint mean can be pictured as a two-entry list: the X average followed by the Y average. Its uncertainty has three pieces. The variance of the X mean occupies one diagonal position, the variance of the Y mean occupies the other, and the covariance between the means occupies the two off-diagonal positions. The off-diagonal entries are equal because covariance does not depend on which variable is named first. A positive entry says samples that overestimate one population mean tend also to overestimate the other. A negative entry says their estimation errors tend to move in opposite directions.

Under independent, identically distributed pairs with finite second moments, averaging more pairs reduces each variance and the covariance by the same sample-size factor. The multivariate central limit theorem then makes the pair of sample means approximately jointly normal for a sufficiently large sample. This is an approximation to the sampling distribution, not a claim that each original monthly return is normally distributed. It also relies on independence between pairs. If returns are serially dependent, the simple division by the number of months can understate uncertainty; a time-series treatment would be needed.

The correlation between the two sample means does not automatically shrink simply because the sample gets larger. Both of their standard errors shrink, and their covariance shrinks, but the scale factors cancel when covariance is converted into correlation. The estimates become tighter around their respective population means while the direction of their errors can remain related. This is why “more data” and “independent estimates” are not synonyms.

A worked example with invented figures

Consider an illustrative pair of monthly desk returns. Suppose the population standard deviation of X is six percentage points, that of Y is three percentage points, and the correlation within a month is one half. Suppose the analyst has thirty-six independent monthly pairs. These are invented round inputs used only to show the method; they are not market observations. The population covariance, using returns in decimal form, is one half times 0.06 times 0.03, or 0.0009. The covariance of the two sample means is therefore 0.0009 divided by thirty-six, or 0.000025.

The standard error of X's mean is six percentage points divided by six, which equals one percentage point. Y's standard error is three percentage points divided by six, or half a percentage point. Check the relationship: dividing the covariance of the means, 0.000025, by 0.01 times 0.005 gives 0.5, the original correlation. That consistency check can catch a missing sample-size factor or a confusion between percentages and decimal returns.

Now suppose the exam asks about the estimated difference between the two population means. The variance of the difference is the variance of X's mean plus the variance of Y's mean minus twice their covariance. In decimal-return units, that is 0.0001 plus 0.000025 minus 0.00005, giving 0.000075. Its standard error is about 0.00866, or 0.866 percentage points. If the analyst ignored the positive covariance, the variance would be 0.000125 and the reported standard error would be about 1.118 percentage points. Positive co-movement makes a paired difference more stable because common monthly surprises partly cancel when one mean is subtracted from the other.

Reverse the sign of the correlation while keeping the two standard deviations and sample size the same, and the difference becomes less stable: subtracting twice a negative covariance adds uncertainty. The direction is worth understanding before calculating. A candidate who expects the same result for positive and negative correlation has lost the central information in the paired data. The separate averages would still be computed exactly as before; it is the inference about their difference that changes.

How to read an exam prompt

First identify whether it asks for two point estimates, their individual standard errors, the covariance of their estimators, or uncertainty about a sum or difference. The first task needs the columns separately. The remaining tasks require the dependence structure. Second, check whether the given covariance refers to observations in a single period or to the already averaged estimates. Divide by sample size only in the former case. Third, inspect the units: returns entered as percentages and returns entered as decimals produce covariances on different numerical scales, even though they describe the same relationship.

For a sum of two estimated means, a positive covariance increases variance because the two errors reinforce one another. For a difference, it decreases variance because common errors cancel. A negative covariance reverses those effects. If a portfolio weighting is involved, the covariance term must also carry the product of the weights. This is the same principle behind the two-asset risk calculations summarised by CFA Institute's portfolio risk reading, though the question here concerns uncertainty in estimated means rather than the realised volatility of a portfolio.

Do not confuse the covariance of returns with the covariance of their averages. The former describes how individual paired observations move together; the latter describes how estimates from repeated samples move together. Nor should a candidate treat two separate confidence intervals as if they automatically formed one joint confidence statement. Joint statements require the full dependence structure and an appropriate joint probability calculation.

There is also a distinction between conditioning on one particular sample and imagining the sampling process repeated. Once a dataset has been collected, its two averages are fixed numbers. A sampling distribution instead asks what values those averages would take if the same data-generating process produced many new sets of paired months. Covariance between estimators belongs to that repeated-sampling picture. This helps explain why a single observed month with both desk returns high is not itself evidence that the two estimated population means are biased. Bias concerns the average estimator over repetitions; covariance concerns how the two estimation errors vary together over those repetitions. The distinction becomes essential when interpreting a confidence interval or a test of equality between means. A question that supplies true population standard deviations and correlation is usually asking for theoretical sampling uncertainty. A question that supplies only historical observations may require estimates of those inputs first. In either case, state the assumption about independent pairs before applying the simple sample-size scaling.

For a practical study check, explain the same dataset in three sentences: each column provides its own mean; common dates create possible covariance between those means; sample size reduces the size of estimation errors but does not erase their correlation. Then calculate a paired difference once with positive correlation and once with negative correlation. That small comparison reveals whether you understand the sign before you rely on a memorised formula. The FRM programme page gives the wider Part I context, while Omni's methodology page is the relevant place to assess how its explanations are organised.

Sources

  1. FRM Study Materials | GARP · garp.org · Retrieved
  2. FRM FAQs | GARP · garp.org · Retrieved
  3. Portfolio Risk and Return: Part I | CFA Institute · cfainstitute.org · Retrieved

Explore FRM study resources

Educational content, not investment advice. This article is written to help candidates prepare for professional exams. It is not a recommendation to buy, sell or hold any security, and it does not take your personal circumstances into account.

Omni Finance Academy is an independent exam-preparation provider and is not affiliated with or endorsed by CFA Institute, GARP, CAIA Association or CFP Board. CFA®, FRM®, CAIA® and CFP® are trademarks of their respective owners. Full disclaimer.