Skip to main content
FRM Part 1 · Topic overview

Quantitative Analysis

Quantitative Analysis supplies the probability, estimation, regression and time-series tools used to measure risk. Match the method to the data, state its assumptions and interpret uncertainty before trusting a forecast.

The topic at a glance

  • 01Probability

    How it works: Use the correct conditioning group and distinguish a density from a probability.

    Watch for: Dividing conditional defaults by the entire population.

  • 02Estimation

    How it works: Separate population variability from uncertainty in an estimated mean.

    Watch for: Treating standard deviation and standard error as interchangeable.

  • 03Models

    How it works: Inspect residuals, stability and performance on unseen data.

    Watch for: Choosing a model only for its in-sample fit.

Original worked example

How precise is the sample mean?

There are 64 independent observations. Population standard deviation is 12 and the sample mean is 50. Use a normal approximation with known population variance.

  1. STEP 1

    Standard error = 12/√64 = 1.5.

  2. STEP 2

    A 95% interval uses approximately 1.96 × 1.5 = 2.94 on either side.

  3. STEP 3

    Compute 50 - 2.94 and 50 + 2.94.

The interval is [47.06, 52.94]. It describes estimation uncertainty, not the range containing 95% of individual observations.

Illustrative assumptions and calculations for learning. Not observed market data or an official exam item.

How the ideas fit together

Before you start: Algebra, percentages and the difference between a sample and a population.

Choose the random variable before the formula

Start with the outcome: a default indicator, a count of defaults, a return, or a waiting time. Then specify its possible values and dependence assumptions. Expectation summarizes probability-weighted outcomes; variance summarizes dispersion around that expectation. Conditional probability changes the information set. These distinctions feed directly into credit-loss models and market-risk forecasts.

Separate estimation, testing and prediction

An estimate provides a fitted value, a confidence interval describes uncertainty about a parameter, and a hypothesis test asks whether data are inconsistent with a specified null model. Regression adds explanatory variables; forecasting adds a future horizon. A prediction interval must account for a new outcome’s randomness as well as estimated coefficients. Check residuals and evaluate forecasts on data unavailable when the model was fitted.

Check dependence before scaling

The square-root sample-size formula assumes independent observations. Serially correlated returns can make the effective information smaller than the observation count suggests. Similarly, square-root-of-time volatility scaling needs stable variance and appropriate dependence assumptions. A correct substitution into an unsuitable formula still produces an unreliable risk number.

Work through the topic

Choose an appropriate statistical tool and interpret its output before applying it to a risk model. The distribution and regression lessons provide the more detailed calculations.

Choose the method from the question and data

Begin with the object being measured. A default indicator has two outcomes, a number of defaults is a count, and a daily return can take positive or negative values. The variable’s support constrains the model: a positive-only distribution cannot directly represent a return that can be negative. Next specify what information is available and whether observations can reasonably be treated as independent. These choices precede the formula.

The question also determines the output. “What fraction defaults?” asks about a probability or proportion. “What is average loss?” asks about an expectation. “How uncertain is the estimated mean?” asks about sampling variability. “How large can losses be at a stated confidence level?” asks about a quantile. Different answers can all be correct for the same dataset because they describe different features of it.

For example, a sample of loan losses can have many zeros, a positive mean and a large upper-tail quantile. Reporting only the mean would hide the tail, while calling the quantile an expected loss would misstate its meaning. Write one sentence defining the target quantity and its units. That sentence is a useful safeguard when a question presents more numerical inputs than the chosen method actually needs.

A method-selection map for risk questions

Describe possible outcomes
Tool to start with: Random variable and distribution
Key qualification: Support, probabilities and dependence
Estimate an unknown quantity
Tool to start with: Estimator and standard error
Key qualification: Sampling design and data quality
Challenge a stated claim
Tool to start with: Hypothesis test
Key qualification: Null, alternative and significance level
Predict a new outcome
Tool to start with: Forecast and prediction interval
Key qualification: Information available before the outcome

Separate dispersion, estimation uncertainty and interval meaning

Standard deviation describes variability in observations. Standard error describes variability in an estimator across repeated samples. If independent observations have a known standard deviation of 12, averaging 64 of them gives a mean with standard error 1.5. The individual observations do not suddenly become less variable because the mean is estimated more precisely. Confusing the two scales can make a forecast interval much too narrow.

The interval [47.06, 52.94] in the opening example comes from a procedure with approximately 95% coverage under the stated model. In the frequentist interpretation, repeated sampling would produce intervals that cover the fixed population mean about 95% of the time. It is not a statement that 95% of future observations lie inside those bounds. When population variance is estimated, the appropriate reference distribution and assumptions must also be reconsidered.

Increasing sample size reduces sampling uncertainty only to the extent that the observations add information. Repeated measurements from one strongly connected episode may contribute less than independent observations from varied conditions. A larger sample can also preserve a systematic selection bias. Check the population represented by the data and its relevance to the decision before treating a smaller standard error as proof of a better risk estimate.

Interpret a test without turning it into a certainty claim

A hypothesis test compares data with a specified null model. Suppose the null mean is 50, the observed sample mean is 53 and the known-variance standard error is 1.5. The standardized statistic is (53 - 50)/1.5 = 2. Under a two-sided standard-normal test, the p-value is about 4.55%, so the null would be rejected at a 5% significance level. The result depends on those stated assumptions and the chosen alternative.

The p-value is a probability of a statistic at least as extreme under the null, not the probability that the null is true. Statistical significance also differs from economic importance. An estimated effect can be very precisely measured yet too small to change a decision after costs. Conversely, an economically important exposure can be estimated imprecisely in a small or noisy sample.

Testing many relationships and reporting only the strongest one creates selection problems. If a model was chosen after trying many versions on the same data, its apparently strong fit needs fresh evaluation. Reserve information that was not used to choose the model, record the search process and compare against a simple benchmark. Those practices address the decision process around the statistic, rather than asking a single p-value to establish everything.

Respect time when evaluating a forecast

A forecast must use information available when it would actually have been produced. Randomly splitting a time series can accidentally train on observations later than the outcomes being predicted. A chronological holdout or rolling evaluation better represents many risk applications. Publication delays matter too: a variable dated in March may not have been known to a forecaster until April.

Separate a forecast’s average error, its variability and its tail misses. A model can have a small average error because positive and negative errors cancel while still producing very poor risk decisions. For a VaR model, coverage and exception timing matter; for a mean forecast, magnitude and direction of errors may be more relevant. Select an evaluation criterion that matches the purpose of the forecast.

A change in the data-generating environment can make old relationships less useful. Monitor residual behavior and revisit the model when evidence changes, while avoiding automatic refitting after every unusual observation. The regression lesson develops coefficient interpretation and diagnostic examples. The model-governance lesson then connects those findings to validation, limits on use and accountable remediation.

Should every risk variable be modeled as normal?

No. Check support, skewness, tail behavior and the mechanism generating the variable. A normal approximation can be useful for some sums or estimators under suitable conditions, but it is not automatically a good description of a default indicator, a positive exposure or a portfolio with nonlinear payoffs.

What is the most useful order for learning these tools?

Understand random variables and conditional probability, then expectation, variance and covariance. Use those to study estimation and testing, and then regression and forecasts. This is a dependency-based study sequence: it helps explain the formulas rather than promising that a fixed number of study days establishes exam readiness.

Try a changed assumption

In the known-variance example, increasing independent observations from 64 to 256 changes the standard error by how much?

Show answer and reasoning

It halves, from 12/√64 = 1.5 to 12/√256 = 0.75. Four times as many independent observations give twice the precision on this standard-error scale, not four times.

Original learning check. No email or account needed.

What should I learn before regression?

Conditional probability, expectation, variance, covariance and estimation. These explain what the coefficients and residual uncertainty mean.

Sources and scope

Links checked 2026-10-10. GARP establishes exam scope; the technical references support the methods identified below. These are selected study notes and original examples, not complete curriculum coverage or official exam questions.

Prepared by MiloPrep editorial team. MiloPrep is independent of GARP. References do not imply endorsement or independent review of these notes.