Why Match Moments?
A parametric model summarizes a distribution using a small number of parameters. When data are available, the method of moments estimates those parameters by equating quantities calculated from the sample to the corresponding theoretical quantities under the model. The method is often direct, but a solution to the equations is not automatically unique, finite, or admissible under the model.
The method complements likelihood-based estimation: it uses the model’s moments rather than the probability of the observed sample. Moment equations can be easy to solve, but they may fail to identify every parameter, or may produce estimates outside the parameter space. The checks are part of the estimation, not an optional afterthought. This treatment is aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.
The first raw sample moment is the sample mean. The second raw sample moment is not the sample variance: it is the average of the squared observations. When a model is described using its mean and variance, convert those quantities to raw moments using \(E[X^2]=\operatorname{Var}(X)+(E[X])^2\). Alternatively, one can match central moments, provided the chosen sample and theoretical moments correspond consistently.
For the usual exactly identified method, the number of equations equals the number of unknown parameters. A valid result must satisfy the equations and the model’s parameter restrictions. If fewer independent equations are available, the parameters may not be identifiable. If additional moments are imposed, the resulting system can be inconsistent; the conventional method does not require matching every available sample moment.
Solving and Checking Moment Equations
The central task is to invert the model’s moment relationships. If the equations have one solution in the allowed parameter space, the method gives a unique finite estimate. If there are several solutions, moment matching alone does not choose between them. If there is no admissible solution, the equations do not produce a valid method-of-moments estimate for that specification.
Use the observations and the stated divisor \(n\) to calculate the raw sample moments needed for the model.
State the assumed distribution and its parameters, then express each required moment as a function of those parameters. The moment must be finite for the equation to be used.
Set sample moments equal to theoretical moments and solve the system, keeping track of possible roots.
Substitute the estimates back into the equations and verify that they meet all model restrictions, including positivity or probability bounds.
Worked Applications
Worked Example: A One-Parameter Uniform Model
A hypothetical claims analyst models positive repair costs, in thousands of dollars, as uniformly distributed on \([0,\theta]\), where \(\theta>0\). Four recorded costs are \(1,2,3,4\). Estimate \(\theta\) by the method of moments and check whether the estimate is compatible with the observed maximum.
Method and model. For \(X\sim\operatorname{Uniform}(0,\theta)\), the density is \(1/\theta\) on \([0,\theta]\). Its first moment is \(E[X]=\int_0^\theta x(1/\theta)\,dx=\theta/2\). The one-parameter method matches this first moment to the sample mean.
Calculate the sample mean. \(\bar{x}=(1+2+3+4)/4=10/4=2.5\) thousand dollars.
Solve the moment equation. \(2.5=\theta/2\), so \(\hat{\theta}=5\) thousand dollars.
Check. The estimate is finite and satisfies \(\hat{\theta}>0\). The largest observed cost is \(4\), which is no greater than the fitted upper endpoint \(5\). Substitution gives \(E[X]=5/2=2.5\), matching the sample mean. The method avoids treating the sample maximum as the moment estimate: the upper endpoint is obtained from the model’s mean relationship.
Worked Example: Two Parameters in a Normal Model
A hypothetical portfolio’s annual claim amounts, in thousands of dollars, are modeled as independent observations from \(X\sim N(\mu,\sigma^2)\), with unknown \(\mu\) and \(\sigma^2>0\). The observed values are \(2,4,6,8\). Find the method-of-moments estimates.
Method and theoretical moments. The model has \(E[X]=\mu\) and \(E[X^2]=\mu^2+\sigma^2\). Match these two raw moments to the first two raw sample moments.
First raw sample moment. \(m'_1=(2+4+6+8)/4=20/4=5\), so \(\hat{\mu}=5\).
Second raw sample moment. \(m'_2=(2^2+4^2+6^2+8^2)/4=(4+16+36+64)/4=120/4=30\). The second equation is \(30=\hat{\mu}^2+\hat{\sigma}^2\), so \(\hat{\sigma}^2=30-5^2=30-25=5\).
Check. The variance estimate is positive, and the fitted theoretical moments are \(E[X]=5\) and \(E[X^2]=25+5=30\), as required. Thus the estimates are \(\hat{\mu}=5\) thousand dollars and \(\hat{\sigma}^2=5\) squared thousand dollars. The moment variance uses divisor \(n=4\); it is not the unbiased sample variance, which uses divisor \(n-1\).
Worked Example: Checking Identification and Probability Bounds
A hypothetical discrete severity model assigns probability \(p\) to a fixed amount \(\theta>0\) and probability \(1-p\) to zero, where \(0<p\leq1\). Four recorded amounts, in thousands of dollars, are \(0,0,2,4\). Use the first two raw moments to estimate \(p\) and \(\theta\), and verify that the parameters are admissible.
Method and theoretical moments. Under this model, \(E[X]=p\theta\) and \(E[X^2]=p\theta^2\). The two unknowns require two moment equations.
Calculate the sample moments. \(m'_1=(0+0+2+4)/4=6/4=1.5\). \(m'_2=(0^2+0^2+2^2+4^2)/4=(0+0+4+16)/4=20/4=5\).
Solve for the parameters. The equations are \(p\theta=1.5\) and \(p\theta^2=5\). Since \(m'_1>0\), divide the second equation by the first to obtain \(\hat{\theta}=5/1.5=10/3\). Then \(\hat{p}=1.5/(10/3)=1.5(3/10)=0.45\).
Check. The estimates satisfy \(\hat{\theta}>0\) and \(0<\hat{p}\leq1\). Substitution gives \(\hat{p}\hat{\theta}=0.45(10/3)=1.5\), and \(\hat{p}\hat{\theta}^2=0.45(100/9)=5\). The solution is unique for these positive sample moments: their ratio fixes \(\theta\), after which the first equation fixes \(p\). Dividing is justified here because \(m'_1=1.5\ne0\).
Why the Equations Work—and When They Do Not
Moment matching is justified by a direct correspondence: if the assumed model has parameter \(\boldsymbol{\theta}\), its population moments are \(E_{\boldsymbol{\theta}}[X^r]\). The sample moments are data-based estimates of those population quantities. Setting them equal gives equations for the parameters; it does not assert that every sample moment must exactly equal its population counterpart in a random sample.
For example, in the normal model, the second raw moment satisfies \(E[X^2]=\operatorname{Var}(X)+(E[X])^2=\sigma^2+\mu^2\). This identity explains why matching \(m'_2\) gives \(\hat{\sigma}^2=m'_2-\hat{\mu}^2\). For any data, \(m'_2-(m'_1)^2\) is the sample variance with divisor \(n\), so it cannot be negative. That guarantees a nonnegative value in this particular calculation, but the model requires a strictly positive variance; data with all observations identical give zero and do not yield an estimate in the stated parameter space \(\sigma^2>0\).
Uniqueness depends on the moment relationships, not merely on having the same number of equations and parameters. The equations must identify the parameters over the allowed parameter space. A nonlinear system can have more than one root, and a root may violate a parameter restriction. Finiteness also matters: a formula involving division by a sample moment is undefined if that moment is zero. Before reporting estimates, substitute them back into the equations and check their ranges.
Common Exam Trap: Using the Unbiased Sample Variance
In the normal example, the sum of squared deviations from the sample mean is \((2-5)^2+(4-5)^2+(6-5)^2+(8-5)^2=9+1+1+9=20\). The method-of-moments variance estimate is \(20/4=5\), because the second raw sample moment uses divisor \(n\). A common incorrect calculation uses the unbiased sample variance \(20/(4-1)=20/3\), then reports \(\hat{\sigma}^2=20/3\) as the moment estimate. The correction is to use the moment equations as defined: \(m'_2=30\), \(\hat{\mu}=5\), and \(\hat{\sigma}^2=m'_2-\hat{\mu}^2=30-25=5\). The unbiased sample variance is a different estimator, not the result of this moment-matching calculation.
A second trap is accepting algebraic roots before checking the model. In the two-point example, \(p\) must be a probability and \(\theta\) must be positive. A solution with \(p>1\), \(p\leq0\), or \(\theta\leq0\) would fail the stated model even if it solved the equations algebraically.
Check Your Understanding
For each problem, form the required moment equations and check the parameter restrictions before accepting the result.
- A hypothetical loss is modeled as uniform on \([0,\theta]\), with \(\theta>0\). Three observed losses, in thousands of dollars, are \(3,5,7\). Find the first-moment estimate of \(\theta\). (Answer: \(\hat{\theta}=10\) thousand dollars.)
- Observations \(1,2,3,6\) are modeled as \(N(\mu,\sigma^2)\), with \(\sigma^2>0\). Find the estimates from the first two raw sample moments. (Answer: \(\hat{\mu}=3\), \(\hat{\sigma}^2=3.5\).)
- A discrete model assigns probability \(p\) to \(\theta>0\) and \(1-p\) to zero, where \(0<p\leq1\). For observations \(0,0,3,3,3\), use the first two raw moments to estimate both parameters. (Answer: \(\hat{\theta}=3\), \(\hat{p}=0.6\).)
- A uniform model has unknown endpoints \(a<b\), with mean \((a+b)/2\) and variance \((b-a)^2/12\). If the first two matched moments give mean \(10\) and variance \(12\), find the endpoints. (Answer: \(\hat{a}=4\), \(\hat{b}=16\).)
- A model is \(X\sim N(\theta,2)\), with unknown real \(\theta\). The sample moments are \(m'_1=2\) and \(m'_2=5\). Does one parameter value match both moments exactly? (Answer: No; the first moment requires \(\theta=2\), which gives theoretical second raw moment \(2^2+2=6\), not \(5\).)