Student-t Predictive Claim Costs

Why Future Claim Costs Have a Student-t Distribution

A predictive distribution describes a future observation while accounting for uncertainty about the parameters that generated it. Under a normal model with both an unknown mean and an unknown variance, that uncertainty makes the predictive distribution heavier-tailed than a normal distribution with fixed parameters. The resulting distribution is Student-t.

For claim-cost modeling, this distinction matters: the predictive distribution is for a future claim cost, not merely for the unknown mean cost. It includes both ordinary claim-to-claim variation and uncertainty about the parameters. This tutorial is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Model and notation: Let \(X_1,\ldots,X_n\) be observed claim costs, conditionally independent and identically distributed given \(\mu\) and \(\sigma^2\), with \(X_i\mid\mu,\sigma^2\sim N(\mu,\sigma^2)\). Use the normal-inverse-gamma prior \[ \mu\mid\sigma^2\sim N\left(m_0,\frac{\sigma^2}{\kappa_0}\right), \qquad \sigma^2\sim\operatorname{InvGamma}(\alpha_0,\beta_0), \] where \(\kappa_0,\alpha_0,\beta_0>0\). The inverse-gamma density is proportional to \((\sigma^2)^{-\alpha_0-1}\exp(-\beta_0/\sigma^2)\). For the observed data, write \(\bar{x}=n^{-1}\sum_i x_i\) and \(S=\sum_i(x_i-\bar{x})^2\). There is no interest discounting or mortality assumption in this claim-cost model.

The posterior remains normal-inverse-gamma. The updated parameters, which summarize the information in the prior and observations, are

\kappa_n=\kappa_0+n,\qquad m_n=\frac{\kappa_0m_0+n\bar{x}}{\kappa_n},\qquad \alpha_n=\alpha_0+\frac n2,

\beta_n=\beta_0+\frac S2+ \frac{\kappa_0n}{2\kappa_n}(\bar{x}-m_0)^2.

Here \(m_n\) is the posterior mean of \(\mu\), and the last term in \(\beta_n\) accounts for disagreement between the prior location and the observed sample mean. The predictive calculation below uses this posterior; it does not treat \(m_n\) and the variance as known fixed parameters.

Deriving the Predictive Distribution

Let \(X_{n+1}\) be one future claim cost from the same population, with the same conditional normal model as the observed costs. First condition on the posterior variance. Since \(\mu\mid\sigma^2,x_1,\ldots,x_n\) is normal with mean \(m_n\) and variance \(\sigma^2/\kappa_n\), integrating over \(\mu\) gives

X_{n+1}\mid\sigma^2,x_1,\ldots,x_n \sim N\left(m_n,\sigma^2\left(1+\frac{1}{\kappa_n}\right)\right).

The extra \(1/\kappa_n\) term is the uncertainty about the unknown mean. Now integrate over the posterior inverse-gamma distribution of \(\sigma^2\). The normal-inverse-gamma mixture yields a Student-t distribution. Under the location-scale convention, \(t_\nu(m,s)\) has location \(m\), degrees of freedom \(\nu\), and scale \(s\); its variance, when it exists, is \(s^2\nu/(\nu-2)\), not \(s^2\).

Theorem: The posterior predictive distribution of one future observation is \[ X_{n+1}\mid x_1,\ldots,x_n\sim t_{\nu}(m_n,s_n), \qquad \nu=2\alpha_n,\qquad s_n^2=\frac{\beta_n(\kappa_n+1)}{\alpha_n\kappa_n}. \] Equivalently, \(s_n^2=(\beta_n/\alpha_n)(1+1/\kappa_n)\). The predictive mean is \(m_n\) when \(\alpha_n>1/2\). Its variance exists when \(\alpha_n>1\), and then \[ \operatorname{Var}(X_{n+1}\mid x_1,\ldots,x_n) =s_n^2\frac{\nu}{\nu-2} =\frac{\beta_n(\kappa_n+1)}{\kappa_n(\alpha_n-1)}. \]

For justification, write the conditional predictive normal variance as \(c\sigma^2\), where \(c=1+1/\kappa_n\). The posterior density of \(\sigma^2\) is proportional to \((\sigma^2)^{-\alpha_n-1}\exp(-\beta_n/\sigma^2)\). Multiplying this by the conditional normal density and integrating over \(\sigma^2>0\) leaves a density in \(x\) proportional to

\left[\beta_n+\frac{(x-m_n)^2}{2c}\right]^{-(\alpha_n+1/2)} \ \propto\ \left[1+\frac{(x-m_n)^2}{2\alpha_n\,s_n^2}\right]^{-(2\alpha_n+1)/2}, \qquad s_n^2=\frac{\beta_n c}{\alpha_n}.

This is the Student-t density with \(2\alpha_n\) degrees of freedom, location \(m_n\), and scale \(s_n\). The heavier tails arise because \(\sigma^2\) is uncertain and is being averaged over, rather than held fixed. As \(\alpha_n\) grows, the degrees of freedom increase and the predictive tails become less heavy relative to a normal distribution.

Using Location, Degrees of Freedom, and Scale

A Student-t predictive result is most useful when all three parameters are reported and interpreted. The location \(m_n\) is the center of the predictive distribution and, when the mean exists, its expected future claim cost. Degrees of freedom \(2\alpha_n\) govern tail heaviness. Scale \(s_n\) measures spread in the Student-t parameterization, but it is not the predictive standard deviation.

For a central predictive interval with probability \(1-\gamma\), let \(t_{\nu,1-\gamma/2}\) denote the \(1-\gamma/2\) quantile of the standard Student-t distribution with \(\nu\) degrees of freedom. Then the interval is

\left[m_n-t_{\nu,1-\gamma/2}s_n,\ m_n+t_{\nu,1-\gamma/2}s_n\right].

A normal likelihood permits negative values even if claim costs cannot be negative. The Student-t predictive distribution does too. It is appropriate only when the normal claim-cost model is a reasonable approximation for the data and the purpose at hand. It is not automatically a suitable model for strongly skewed or strictly positive costs.

Worked Example: Predicting a Future Cost With Symmetric Data

A hypothetical insurer models individual claim costs in dollars. The observed costs are \(8,10,12\). The prior parameters are \(m_0=10\), \(\kappa_0=1\), \(\alpha_0=2\), and \(\beta_0=2\). Find the posterior predictive distribution for one future cost and its mean and variance. No interest discounting or mortality assumption is used; the conditional severity model is normal with the stated normal-inverse-gamma prior.

Method and data summaries. Here \(n=3\), \(\bar{x}=(8+10+12)/3=10\), and \(S=(8-10)^2+(10-10)^2+(12-10)^2=4+0+4=8\).

Update the parameters. The posterior precision multiplier is \(\kappa_n=1+3=4\). The posterior location is \(m_n=(1\cdot10+3\cdot10)/4=10\). The shape parameter is \(\alpha_n=2+3/2=3.5\). The scale parameter is \[ \beta_n=2+\frac82+\frac{1\cdot3}{2\cdot4}(10-10)^2=2+4+0=6. \]

Form the predictive distribution. The degrees of freedom are \(\nu=2(3.5)=7\), and the squared scale is \[ s_n^2=\frac{6(4+1)}{3.5\cdot4}=\frac{30}{14}=\frac{15}{7}\approx2.1429. \] Thus \(s_n=\sqrt{15/7}\approx1.4639\), and \(X_4\mid x_1,x_2,x_3\sim t_7(10,1.4639)\). The mean is \(10\). The variance is \((15/7)\cdot7/(7-2)=3\), so the predictive standard deviation is \(\sqrt3\approx1.7321\).

Check. The variance formula using posterior parameters gives \(6(5)/(4(3.5-1))=30/10=3\), agreeing with the Student-t variance formula. The common error of calling \(s_n^2=15/7\) the predictive variance would omit the factor \(\nu/(\nu-2)=7/5\).

Worked Example: A 95% Predictive Interval

For a hypothetical portfolio, claim costs in dollars are modeled conditionally as normal, with the same normal-inverse-gamma assumptions and no interest discounting or mortality. The observed costs are \(16,20,24\). Use prior parameters \(m_0=20\), \(\kappa_0=2\), \(\alpha_0=3\), and \(\beta_0=12\). Find the predictive distribution and a 95% central predictive interval.

Compute the data summaries. We have \(n=3\), \(\bar{x}=(16+20+24)/3=20\), and \(S=(16-20)^2+(20-20)^2+(24-20)^2=16+0+16=32\).

Update. The parameters are \(\kappa_n=2+3=5\), \(m_n=(2\cdot20+3\cdot20)/5=20\), and \(\alpha_n=3+3/2=4.5\). Also, \[ \beta_n=12+\frac{32}{2}+\frac{2\cdot3}{2\cdot5}(20-20)^2=12+16=28. \]

Therefore \(\nu=2(4.5)=9\), and \[ s_n^2=\frac{28(5+1)}{4.5\cdot5} =\frac{168}{22.5}=7.4667, \qquad s_n\approx2.7325. \] The predictive mean is \(20\). The 95% interval uses \(t_{9,0.975}\approx2.262\). Its margin is \(2.262(2.7325)\approx6.1809\), giving \([13.82,26.18]\) dollars when rounded to the nearest cent.

Check and trap. The interval is centered at the posterior location, \(20\), and has equal margins. Using \(2.262\) times the predictive standard deviation instead of the Student-t scale would be incorrect: the quantile multiplies \(s_n\), while the predictive standard deviation is \(s_n\sqrt{9/7}\approx3.101\). The final interval is for a future individual cost, not for the unknown mean \(\mu\).

Worked Example: Prior-Data Disagreement and Predictive Variance

A hypothetical insurer has two observed costs, \(30\) and \(34\) dollars. The conditional model is normal with unknown mean and variance and a normal-inverse-gamma prior having \(m_0=28\), \(\kappa_0=2\), \(\alpha_0=3\), and \(\beta_0=18\). There is no interest or mortality adjustment. Find the posterior predictive mean, scale, degrees of freedom, and variance for one future cost.

Summarize and update. Here \(n=2\), \(\bar{x}=(30+34)/2=32\), and \(S=(30-32)^2+(34-32)^2=8\). Thus \(\kappa_n=2+2=4\), \(m_n=(2\cdot28+2\cdot32)/4=30\), and \(\alpha_n=3+2/2=4\).

The posterior variance parameter is \[ \beta_n=18+\frac82+\frac{2\cdot2}{2\cdot4}(32-28)^2 =18+4+\frac12(16)=30. \] The final term is \(8\), reflecting the gap between prior and observed centers. The predictive degrees of freedom are \(\nu=2(4)=8\), and \[ s_n^2=\frac{30(4+1)}{4\cdot4}=\frac{150}{16}=9.375, \qquad s_n\approx3.0619. \] Hence the predictive mean is \(30\) dollars and the variance is \(9.375(8/6)=12.5\) dollars squared.

Check. Separating sources of uncertainty gives \(E[\sigma^2\mid\text{data}]=\beta_n/(\alpha_n-1)=30/3=10\) and \(\operatorname{Var}(\mu\mid\text{data})=E[\sigma^2\mid\text{data}]/\kappa_n=10/4=2.5\). Their sum is \(10+2.5=12.5\), matching the predictive variance. Ignoring the prior-data disagreement would incorrectly omit \(8\) from \(\beta_n\).

Exam Traps and Interpretation

The predictive distribution is wider than the conditional claim distribution with parameters treated as known because it includes uncertainty about the mean and variance. In particular, the \(1+1/\kappa_n\) factor in the scale comes from uncertainty about \(\mu\), while the Student-t tail behavior comes from mixing over \(\sigma^2\). Both features are part of the predictive result.

Confusing scale and standard deviation. For \(\nu>2\), predictive standard deviation is \(s_n\sqrt{\nu/(\nu-2)}\), not \(s_n\).
Using \(\alpha_n\) as the degrees of freedom. Under the stated inverse-gamma parameterization, the degrees of freedom are \(2\alpha_n\).
Using the posterior mean distribution as the predictive distribution. A distribution for \(\mu\) describes uncertainty in the mean, whereas \(X_{n+1}\) also has claim-to-claim variation.
Omitting the prior-data disagreement term. The contribution \(\kappa_0n(\bar{x}-m_0)^2/(2\kappa_n)\) belongs in \(\beta_n\).

Always state the inverse-gamma parameterization before using the formula: a rate parameterization or a different definition of \(\beta\) changes the expressions. Also check whether a requested predictive variance exists; it requires \(\alpha_n>1\), equivalently \(\nu>2\).

Check Your Understanding

For each hypothetical claim-cost model, use the normal-inverse-gamma parameterization defined above. No interest discounting or mortality assumption applies. Report dollar amounts to the nearest cent when an interval is requested.

Observed costs are \(7,9\), with \(m_0=8,\ \kappa_0=2,\ \alpha_0=3,\ \beta_0=4\). Give the predictive degrees of freedom, location, scale, and variance. (Answer: \(\nu=8\), location \(8\), scale \(1.25\), variance \(25/12\approx2.0833\) dollars squared.)
Observed costs are \(48,52,54\), with \(m_0=50,\ \kappa_0=1,\ \alpha_0=2,\ \beta_0=6\). Give the predictive degrees of freedom, location, squared scale, and variance. (Answer: \(\nu=7\), location \(51\), squared scale \(40/7\), variance \(8\) dollars squared.)
Observed costs are \(8,12\), with \(m_0=10,\ \kappa_0=2,\ \alpha_0=4,\ \beta_0=8\). Find the central 90% predictive interval, using \(t_{10,0.95}=1.8125\). (Answer: \([6.86,13.14]\) dollars.)
One cost of \(15\) is observed, with \(m_0=10,\ \kappa_0=1,\ \alpha_0=1,\ \beta_0=2\). Find the predictive degrees of freedom, location, squared scale, and variance. (Answer: \(\nu=3\), location \(12.5\), squared scale \(8.25\), variance \(24.75\) dollars squared.)
Observed costs are \(4,6,8,10\), with \(m_0=5,\ \kappa_0=2,\ \alpha_0=3,\ \beta_0=10\). Find the predictive mean, degrees of freedom, squared scale, and variance. (Answer: mean \(19/3\), \(\nu=10\), squared scale \(238/45\), variance \(119/18\) dollars squared.)

What You'll Learn