Skip to content

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

FRM® estimator bias: why the sample variance divisor matters

Estimator bias is an average error across repeated samples, not a verdict on one estimate. Learn why the variance divisor changes that bias and how to read it.

ConceptOmni Curriculum Team7 min read

Four blank ceramic tiles surround a wooden disc on a navy desk beside a ruler and closed notebook.
On this page

Estimator bias is the difference between an estimator’s average result over repeated samples and the population quantity it is meant to estimate. A sample estimate can miss the truth even when its method is unbiased; the question is whether its errors cancel over repeated sampling. For FRM® Part I, the practical example is variance: dividing squared deviations from the sample mean by the sample size tends to understate population variance, while the familiar one-degree-of-freedom adjustment removes that average shortfall under standard sampling assumptions.

A property of a method, not of one observed answer

A parameter describes a population and is treated as fixed. An estimator is a rule applied to data that could have been different if a different sample had been drawn. The resulting estimate is the particular number from the sample in front of you. Confusing the rule with its realised value is the first source of errors. Bias describes the rule’s distribution over hypothetical repetitions, even though an analyst usually observes only one sample.

Bias⁡(θ^)=E(θ^)−θ\operatorname{Bias}(\hat\theta)=E(\hat\theta)-\theta
Bias equals the expected value of the estimator minus the population parameter.

A positive value means the estimator tends to run high; a negative value means it tends to run low. Zero bias means the expected estimate equals the parameter. It does not mean every estimate is correct, that its errors are small, or that it is the best estimator for a particular decision. The expectation is taken over samples generated under the assumed sampling model. If the sampling process changes, the distribution of the estimator can change with it.

Imagine repeatedly selecting equal-sized independent samples from the same stable population and using the same rule each time. Plotting those estimates would produce a cloud. The centre of that cloud reveals bias; its spread reveals sampling variability. A narrow cloud centred away from the parameter is precise but biased. A broad cloud centred on the parameter is unbiased but imprecise. For a risk estimate, either defect can matter. The candidate must identify which property a prompt is actually testing before choosing an answer.

Why the sample mean is centred correctly

The sample mean adds observations and divides by their count. If each observation is drawn from a population with the same finite mean, linearity of expectation makes the expected average equal to that mean. Independence is useful for statements about uncertainty and precision, but linearity itself does not require independence. The point is a property of the estimator, not an assertion that a small sample will look representative. One unusual observation can still make its realised value far from the population mean.

E(Xˉ)=1n∑i=1nE(Xi)=μE(\bar X)=\frac{1}{n}\sum_{i=1}^{n}E(X_i)=\mu
When every observation has population mean mu, the expected sample mean is mu.

This distinction is worth keeping when reading results from a small historical return series. A sample mean of returns may be an unbiased method for estimating a stable population mean under the stated model, yet the particular estimate may have large uncertainty. A forecast based on it may also fail if the return-generating process has changed; the mathematical property does not repair a poor population definition or a broken model.

Why variance needs a different divisor

Variance concerns squared distances from a centre. If the true population mean were known, distances from that fixed mean would have a different calculation. In practice the mean is usually estimated from the same observations whose dispersion we want to measure. The sample mean sits at the centre of those observations by construction, so squared distances around it are unusually small compared with squared distances around the unknown population mean. Dividing their sum by the full sample size therefore yields a downward-biased estimate of population variance under independent, identically distributed sampling with finite variance.

σ^n2=1n∑i=1n(Xi−Xˉ)2\hat\sigma_n^2=\frac{1}{n}\sum_{i=1}^{n}(X_i-\bar X)^2
The sample variance using the full sample count in the denominator.
E(σ^n2)=n−1nσ2E(\hat\sigma_n^2)=\frac{n-1}{n}\sigma^2
Under the standard sampling assumptions, the expected full-count variance is below population variance.

Once the sample mean is fixed, the deviations from it must sum to zero. Knowing all but one determines the last, leaving one fewer independent deviation than observations. Replacing the divisor by the sample size less one corrects the expected shortfall. This is a correction for estimating the centre, not a magical claim that the realised sample dispersion equals the population dispersion.

s2=1n−1∑i=1n(Xi−Xˉ)2s^2=\frac{1}{n-1}\sum_{i=1}^{n}(X_i-\bar X)^2
The corrected sample variance uses one fewer than the observation count.

The correction requires more than a single observation. It targets variance, not standard deviation: taking the square root of an unbiased variance estimate does not generally produce an unbiased estimator of population standard deviation. That distinction matters if an answer choice casually calls a volatility estimate unbiased merely because its underlying variance used the corrected divisor.

A small illustrative return sample

Take four invented returns expressed in percentage points: two, four, six and eight. Their sample mean is five. The deviations from five are negative three, negative one, one and three percentage points. Squaring and adding gives twenty squared percentage points. These are illustrative numbers, not reported market returns. The full-count variance is twenty divided by four, or five squared percentage points. The corrected sample variance is twenty divided by three, or approximately 6.67 squared percentage points.

Notice what this comparison proves and what it does not. Both calculations use the same realised sample; the corrected answer is larger because its denominator is smaller. We do not know the true population variance from these four observations, so we cannot say which realised number is closer to it. The statement that the corrected estimator is unbiased is about its expected value across all possible samples, not about this one set of data. This is the most common conceptual trap in a question that shows a tiny table of returns.

For an even simpler way to see the direction, imagine a sample whose observations happen to be close together. The sample mean adapts to exactly that sample. It therefore makes those deviations as small as possible for any chosen centre. Correcting the divisor counteracts the systematic shrinkage caused by estimating that centre. It does not compensate for every form of data problem: missing observations, a change in regime, measurement error and selection of only calm periods are separate issues that a divisor change cannot cure.

How to read a bias calculation

Suppose an illustrative population variance is twelve squared units and an analyst uses the full-count variance on samples of four observations. The expected value of that estimator is three quarters of twelve, or nine squared units. Its bias is nine minus twelve, or negative three squared units. The sign is negative because the method runs low on average. This is an expectation calculation, not a second realised data sample. If the sample count grows while assumptions remain in place, the proportional gap gets smaller, even though it is still present for a finite count.

A candidate can work backwards as well. If a prompt gives an expected estimate below the stated population parameter, subtract parameter from expected estimate before assigning a sign. Do not subtract in the reverse order just because a question calls the magnitude an “error”. If it asks for an absolute bias, then report a non-negative magnitude; if it asks for bias itself, preserve the direction. Check the units too: a variance bias has squared return units, whereas a mean bias has return units.

Bias versus precision and usefulness

An unbiased estimator is not automatically the most useful one in a particular finite sample. Estimators also vary in dispersion. A deliberately smoother estimate can be slightly biased yet have less variation from sample to sample. A GARP-hosted technical paper explicitly chooses the full-count variance in one application because its authors prefer a biased estimator with smaller mean squared error. That choice is context-specific. A candidate should not infer that a biased method is preferable unless the question supplies a criterion for comparing them.

Consistency asks a different question: does an estimator approach the parameter as the amount of data grows? The full-count sample variance has finite-sample downward bias under the standard model, but the fraction of variance lost through the divisor shrinks as the count increases. Bias and consistency are therefore not interchangeable labels. Nor does a long sample guarantee relevance if the distribution itself changed over the observation period. The mathematical sampling model and the analyst’s real-world data choice need separate scrutiny.

A reliable exam sequence

  • Name the target parameter before looking at a formula. Population mean, population variance and standard deviation are different targets.
  • Identify whether the question asks about one estimate, the expected estimate, bias, variance of the estimator or consistency. Each refers to a different quantity.
  • For bias, calculate expected estimator minus target and keep the sign. An estimate being above the target in one sample does not establish positive bias.
  • For sample variance, check whether the centre was estimated from the same data and whether the question explicitly asks for the full-count or corrected divisor.
  • State the sampling assumptions when explaining why a correction works. The correction is not a cure for selection bias or a changing population.

GARP places Quantitative Analysis in its FRM study materials and describes the scope of its learning objectives. For a study route, the Omni FRM programme page and methodology page give the relevant internal context. When revising this topic, explain the distinction aloud before calculating: a correct divisor chosen for the wrong question is still a wrong answer.

Sources

  1. FRM Study Materials | GARP · garp.org · Retrieved
  2. FRM Learning Objectives | GARP · garp.org · Retrieved
  3. GARP-hosted technical paper on risk estimation · garp.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.