On this page
An estimator is consistent when, as the sample grows, the chance that its estimate is meaningfully far from the true population value tends to zero. For FRM® Part I, the key question is whether more observations make an estimator converge to the target, not whether the estimator is perfectly unbiased in every small sample. A useful check is whether its bias and its sampling variance both shrink towards zero; that combination is sufficient for consistency. The property describes a limiting pattern, not the accuracy of one estimate you happen to observe.
Separate the estimate from the rule that produces it
A population parameter is a fixed but usually unknown feature, such as a mean return under a specified data-generating process. An estimator is a rule applied to a random sample, such as taking the sample average. Before the data arrive, the estimator can produce different values across different samples, so it has a sampling distribution. After the sample arrives, its particular result is an estimate. This distinction explains why consistency is a property of the rule across growing sample sizes rather than a label attached to one realised number.
Imagine repeatedly drawing samples from the same stable population, once with a small number of observations and once with many. The sample mean from the larger samples should cluster more tightly around the population mean under suitable assumptions. Consistency says the probability of landing outside any fixed tolerance around the true parameter approaches zero as sample size increases without bound. The tolerance could be economically large or very small; the definition must work for either fixed choice. The underlying population and sampling assumptions matter because additional observations from a changed or biased process do not automatically improve the estimate of the original target.
GARP identifies Quantitative Analysis within the FRM learning objectives. CFA Institute's estimation and inference refresher describes the sampling distribution and standard error of the sample mean. The worked examples below are newly invented to unpack the same statistical reasoning at the level of the assigned FRM reading; their figures are illustrative rather than empirical returns.
Bias and variance give a practical test
Bias is the difference between an estimator's expected value and the parameter it aims to estimate. Variance of the estimator measures how much the rule's output varies across repeated samples. If bias approaches zero and estimator variance approaches zero as the sample grows, the mean squared estimation error approaches zero too. In that case the probability of an error larger than a fixed tolerance also approaches zero. This is a sufficient route to consistency that makes the idea operational in an exam calculation.
The two parts play different roles. An estimator can be centred correctly on average but remain widely scattered. A single estimate from that rule can still miss badly, and more available data do not help if the rule stubbornly ignores them. Conversely, a rule can become extremely precise about the wrong value. Its distribution narrows, but its centre stays away from the intended parameter. Neither low bias alone nor low variance alone is enough for this simple sufficient test. Think of convergence as needing both the centre to move towards the target and the spread around that centre to contract.
Do not silently replace estimator variance with population variance. The population can remain just as variable while the sampling variance of an average becomes smaller. Those are different objects. In a stable independent sample with finite population variance, averaging more observations reduces the variance of the mean estimator because idiosyncratic deviations partly offset one another. If observations are strongly dependent, or the target itself changes, the familiar reduction need not apply in the same way. State the assumptions before treating sample size as a cure.
Work through an invented sample-mean example
Assume a stable population of monthly changes in a risk measure, with a mean of five units and population standard deviation of twelve units. Treat observations as independent and identically distributed for this illustration. The sample mean is unbiased: across repeated samples, its expected value is five. Its standard error is the population standard deviation divided by the square root of sample size. With nine observations, the standard error is four units. With thirty-six observations, it is two units. With one hundred and forty-four observations, it is one unit. The population standard deviation stays at twelve throughout; only the sampling spread of the mean contracts.
These invented figures do not say that a particular nine-observation sample must lie within four units of five, or that a particular large sample must be within one unit. Standard error describes the spread across hypothetical repeated samples, not a hard bound on a realised estimate. The example does show the trend: multiplying sample size by four halves the standard error under the stated assumptions. As sample size continues to increase, the standard error tends to zero. Since the mean estimator is already centred on the true mean, both parts of the sufficient test are met.
A common exam shortcut is to see an unbiased sample mean and immediately conclude that any realised estimate is reliable. That skips the sampling variance. Suppose a rule takes only the first observation of every available sample and ignores the rest. It is unbiased for the population mean if that observation has the population mean. Yet its variance remains equal to the population variance, even when the file contains thousands of observations. More data have not changed the rule's output. This estimator is unbiased but inconsistent, which proves that unbiasedness alone does not establish convergence.
A biased rule may still converge
Now consider the familiar sample-variance calculation using the sample size itself as the divisor after subtracting the sample mean. Under independent, identically distributed observations with finite variance, this rule is biased downward in a finite sample: its expected value is the population variance multiplied by the factor one minus one divided by sample size. The bias is therefore the negative of the population variance divided by sample size. As sample size grows, that bias tends to zero. Under suitable regularity conditions the estimator's spread also contracts, so it is consistent despite its finite-sample bias.
Use invented round numbers to see the scale. If the population variance is twenty-five square units and the sample size is five, the expected value of this divisor-five estimate is twenty square units, a downward bias of five square units. If the sample size is one hundred, the expected value is twenty-four point seventy-five square units and the bias is one quarter of a square unit. The arithmetic uses the expectation of the estimator, not a claim that any one computed sample variance must equal those values. The bias never disappears completely at a finite sample size under this formula, but it vanishes in the limit.
The usual sample-variance rule with sample size minus one as divisor removes this particular finite-sample bias under the stated sampling assumptions. Both rules can be consistent. Therefore the question 'which rule is unbiased?' and the question 'which rule is consistent?' can have different answers. In the related sample-variance discussion, focus on the divisor's effect on expected value. Here the focus is the limiting behaviour as more data become available. The distinction is the whole point of this reader question.
What the limit does and does not buy you
Consistency is an asymptotic statement. It does not specify how large a sample is large enough for an application, give a confidence level for one realised estimate, or protect against a model that does not represent the data. A consistent rule may perform poorly in a modest sample. A procedure might also converge slowly, particularly when observations are noisy or dependent. The property is still valuable because it rules out estimators that remain systematically wrong or noisy even with arbitrarily much relevant information, but it cannot substitute for a finite-sample uncertainty check.
The phrase 'more data' also hides a choice about what counts as comparable data. A long return history can mix different market structures or measurement conventions. In such a case the sample mean might converge beautifully to an average across regimes while failing to answer the current decision-maker's intended question. That is a target-definition problem, not a mathematical contradiction of consistency. In an exam vignette, identify the parameter being estimated and the assumptions under which the observations are generated before applying a convergence argument. Avoid treating consistency as a general promise that collecting any observations fixes a weak design.
A compact decision sequence is useful under time pressure. Name the parameter and estimator. Check whether the expected estimate approaches the parameter as sample size increases. Check whether the variance of the estimator, rather than the variance of the population, tends to zero. If both hold, conclude consistency under the stated assumptions. If one fails, seek a counterexample or an alternative argument instead of relying on the sufficient test. For a concrete study route, the FRM programme page places this concept in the broader course, and the Omni methodology explains how to organise a review of calculation errors.
The most frequent mistake is to equate unbiased with consistent. The first-observation rule shows why unbiasedness can be too weak; the divisor-by-sample-size variance rule shows why finite-sample bias need not be fatal. A second mistake is to read the standard error of the mean as the variability of individual observations. A third is to infer exact accuracy from a limiting theorem. When you see 'consistent' in an FRM question, answer the convergence question: as the sample grows under the specified conditions, does the estimator concentrate on the true parameter?
Sources
- 2026 FRM Learning Objectives | GARP · garp.org · Retrieved
- Estimation and Inference | CFA Institute · cfainstitute.org · Retrieved
- Members Guide to Refresher Readings | CFA Institute · cfainstitute.org · Retrieved



