On this page
For FRM® Part I, choose the covariance calculation from the kind of data you have: observed paired returns call for a sample estimate, while a complete scenario distribution uses stated probabilities. Correlation then scales covariance by the two standard deviations so that the strength of linear co-movement can be compared across pairs. The main error is to mix the divisor or units from one setting with those of another.
What is being estimated?
Covariance asks whether two variables tend to sit on the same side of their respective means in the same observation. If both deviations are positive, or both are negative, their product is positive. If one is above its mean and the other below, the product is negative. The average of these paired products describes the direction of linear co-movement. The word paired matters: the return for asset A in one month must be combined with asset B’s return from that same month, not with an independently sorted list.
GARP identifies quantitative analysis as part of FRM Part I, and its study materials organise learning objectives by reading. The assigned Omni curriculum note covers sample covariance and correlation, but an exam answer requires more than a remembered equation. You need to identify the observation unit, use consistent units and decide whether the question describes a sample or a probability model. A GARP-hosted technical paper on covariance estimation also emphasises that a sample covariance matrix is an estimate with limitations, rather than a permanent property of a portfolio.
Population covariance is an expectation over an underlying probability distribution. When the distribution is supplied as scenarios with probabilities that sum to one, multiply each pair of deviations by its scenario probability and add. A historical sample is different: its population means are unknown, so you replace them with sample means. Using estimated means consumes information, which is why the conventional unbiased sample covariance uses one fewer than the number of paired observations as its divisor. Always follow a question’s explicitly stated estimator if it specifies one.
Build a sample estimate from paired observations
Here is an invented, deliberately small set of four paired percentage returns for two hypothetical series. At successive observation dates, A records 1%, 3%, 5% and 7%; B records 2%, 4%, 4% and 6%. The figures exist solely to show the arithmetic. A’s sample mean is 4%, and B’s is also 4%. Subtract the correct mean from each member of each pair before multiplying deviations.
The paired deviations, measured in percentage points, are A: minus 3, minus 1, plus 1 and plus 3; B: minus 2, zero, zero and plus 2. Their paired products are 6, zero, zero and 6, giving a total of 12 percentage-point squared. The conventional sample covariance is 12 divided by three, or 4 percentage-point squared. Dividing by four instead would give 3; that is a different divisor convention and does not answer a request for the conventional unbiased sample estimate.
Now calculate the sample dispersions using the same observations. A’s squared deviations sum to 20, so its sample variance is 20 divided by three and its sample standard deviation is about 2.582 percentage points. B’s squared deviations sum to 8, so its sample variance is 8 divided by three and its standard deviation is about 1.633 points. The correlation is the covariance, 4, divided by the product of those standard deviations, about 4.216. The resulting correlation is approximately 0.949. This describes strong positive linear association in these invented observations; it is not a forecast or proof of a stable future relationship.
The numbers provide two quick checks. Each deviation column adds to zero, as it should when deviations are taken from that column’s own sample mean. Correlation falls between minus one and plus one. If a computed correlation lies outside that range, review the pair matching, the squared-deviation sums, the units and the divisor before interpreting it. Do not change a result merely to force the bound: identify the arithmetic error.
Why the correlation divisor can appear to disappear
If the same sample convention is used for covariance and both variances, the common divisor cancels when covariance is divided by the product of the standard deviations. In this example, using the divisor four everywhere would produce a covariance of 3, a standard deviation of about 2.236 for A and about 1.414 for B; 3 divided by their product is again about 0.949. The covariance estimate changes with the convention, but this consistently calculated correlation does not.
That cancellation does not excuse inconsistent working. Dividing covariance by three while using standard deviations formed with divisor four produces an incorrect correlation. Nor does it mean that sample covariance and population covariance have identical meanings. State what has been estimated. Covariance also has scale: if returns are entered as percentages, its units are percentage-point squared. If they are entered as decimals, the numeric covariance is smaller by a factor reflecting the unit conversion, while correlation remains the same. A small-looking decimal covariance is not evidence of weak association.
The sign is easier to interpret than covariance’s raw size. Positive covariance means same-direction deviations dominate, negative means opposite-direction deviations dominate, and near zero means little net linear co-movement in the observations. The magnitude alone cannot be compared across differently scaled series. Correlation removes the units and is bounded, making a comparison of linear association more meaningful. Yet a correlation near zero does not establish independence: nonlinear dependence can remain, and a short sample can conceal behaviour outside its observed range.
A scenario table calls for different weighting
Consider a separate invented scenario model, not an additional historical sample. Suppose a quiet state has probability 0.25 and returns of minus 2% for A and 6% for B; a middle state has probability 0.50 and returns of 4% and 3%; and a strong state has probability 0.25 and returns of 10% and zero. The probabilities add to one. The expected A return is 4%, and the expected B return is 3%, so the deviations in the outer states are symmetric.
In the quiet state, the deviation product is minus 18 percentage-point squared; in the middle state it is zero; and in the strong state it is also minus 18. Weighting by their stated probabilities gives minus 4.5, zero and minus 4.5. The resulting model covariance is minus 9 percentage-point squared. A’s probability-weighted variance is 18 and B’s is 4.5, giving standard deviations of about 4.243 and 2.121 percentage points. Correlation is minus 9 divided by their product, or minus one. Within this deliberately simple model, the relationship is perfectly negative and linear.
There is no sample degrees-of-freedom adjustment in that scenario calculation. The probabilities describe the model distribution provided in the question. Applying an extra division by the number of states, or by one fewer than that number, would double weight the observations and break the calculation. However, if a question says that the state frequencies are an empirical sample rather than model probabilities, its wording may require a sample estimator. The data description controls the method.
What an examiner can change to test understanding
An exam item may give variances rather than standard deviations. Take square roots before standardising covariance; dividing by variances is a different operation and can yield an impossible correlation. Another item may provide correlation and two standard deviations and ask you to recover covariance. Multiply the three quantities, keeping their units consistent. A third may change the scale of one series, which changes covariance but not correlation if the scale factor is positive. These variants probe the relationship among measures rather than separate memorised formula lists.
Before calculating, write a short margin note: paired observations or probability scenarios; percent or decimal units; and required divisor. During calculation, check that both means use the same dates as the covariance pairs. After calculation, inspect the sign and correlation bound. These checks cost less time than rebuilding a full table after a wrong answer. They also make the conclusion appropriately modest: a sample estimate summarises those observations and may shift when the sample or market regime changes.
For structured revision, the FRM programme page situates Part I quantitative analysis within the broader route, while Omni’s methodology explains its learning approach. Practise one small paired-return table and one probability table side by side. The purpose is to make the choice of estimator automatic only after you have read what the data actually represent.
Sources
- FRM Exam FAQs | GARP · garp.org · Retrieved
- FRM Study Materials | GARP · garp.org · Retrieved
- A Gentle Introduction to Value at Risk · garp.org · Retrieved



