Tutorials › CAS Exam 8 - Advanced Ratemaking › Separating Process Noise from Risk Differences

Home  /  Tutorials  /  CAS Exam 8 / Evaluating Experience Rating

CAS Exam 8 · Evaluating Experience Rating · Tutorial 572 of 989

Separating Process Noise from Risk Differences

Separate persistent risk differences from process noise and see how repeated, comparable observations change the weight placed on experience.

Advanced 12 min read

What You'll Learn

  • Define persistent risk effects and period-specific process noise in a repeated-observation model
  • Decompose total variation into between-risk and within-risk components
  • Relate the correlation of repeated observations to persistent risk differences
  • Calculate how averaging repeated observations reduces process noise
  • Apply exposure-weighted credibility when periods have different exposures
  • Recognize when caps, floors, or changing measurement bases invalidate a simple decomposition

Why Observed Experience Varies

An experience modification can change from one renewal to the next for two different reasons. The insured's underlying risk may be persistently higher or lower than the rating-plan baseline; alternatively, claims may fluctuate randomly even when the underlying risk has not changed. These are distinct sources of variation. Treating every change as a change in underlying risk can overreact to a noisy period, while treating all variation as noise can obscure durable differences between risks.

To separate the two, work with observations measured on a comparable scale: for example, an adjusted loss-cost index or an uncapped modification signal. The observations must use consistent exposure, coverage, trend, and loss-adjustment conventions. A reported modification that has been clipped by a floor or cap may not preserve the variation in its underlying signal, so a simple additive variance model may not apply directly to the clipped values.

Definition: In a repeated-observation model, write the comparable observation for risk \(i\) in period \(t\) as \(X_{it}=\mu+\theta_i+\varepsilon_{it}\). Here \(\mu\) is the portfolio baseline, \(\theta_i\) is risk \(i\)'s persistent deviation from that baseline, and \(\varepsilon_{it}\) is period-specific process noise. Assume \(E[\theta_i]=0\), \(\operatorname{Var}(\theta_i)=\tau^2\), \(E[\varepsilon_{it}]=0\), and \(\operatorname{Var}(\varepsilon_{it})=\sigma^2\). Assume the process errors are independent across periods and risks and independent of the persistent effects. The variance \(\tau^2\) measures variation in underlying risk; \(\sigma^2\) measures random variation in an individual period.

Under this model, a single period combines both sources of variation:

$$ \operatorname{Var}(X_{it})=\tau^2+\sigma^2. $$

Repeated observations of the same risk share \(\theta_i\), but have independent process errors. For two different periods \(t\ne s\), this gives:

$$ \operatorname{Cov}(X_{it},X_{is})=\tau^2. $$

Thus persistence appears as covariance across periods. Process noise contributes to each period's variance, but, under the independence assumption, not to the covariance between periods. This distinction is the basis of both variance-component estimation and repeated-experience credibility.

What Repeated Observations Tell Us

If the risk is observed for \(n\) equally weighted periods, its average is \(\bar X_i=\mu+\theta_i+\bar\varepsilon_i\). Averaging leaves the persistent effect unchanged, but reduces the process-noise variance from \(\sigma^2\) to \(\sigma^2/n\). Consequently:

$$ \operatorname{Var}(\bar X_i)=\tau^2+\frac{\sigma^2}{n}. $$

The share of the average's variance attributable to persistent risk is often called its reliability:

$$ Z_n=\frac{\tau^2}{\tau^2+\sigma^2/n}. $$

A higher \(Z_n\) means the observed average is more informative about the persistent risk effect relative to the baseline. It does not mean the risk has become intrinsically more variable or that the rating plan is more accurate in every respect.

Credibility of the Repeated Average

For a normal random-effects model, suppose \(\theta_i\sim N(0,\tau^2)\), and the independent process errors satisfy \(\varepsilon_{it}\sim N(0,\sigma^2)\). The posterior mean of the persistent level, conditional on the \(n\) observations, is a credibility-weighted average of the portfolio baseline and the risk's observed average.

Theorem: Under the stated normal, independent, equal-variance assumptions, \(E[\mu+\theta_i\mid X_{i1},\ldots,X_{in}]=\mu+Z_n(\bar X_i-\mu)\), where \(Z_n=\tau^2/(\tau^2+\sigma^2/n)\). If normality is not assumed, the same expression is the best linear predictor under the stated variance and independence assumptions, but it need not equal the full conditional expectation.

The weight follows directly from the covariance of the persistent level with the observed average. The average has mean \(\mu\), variance \(\tau^2+\sigma^2/n\), and covariance \(\tau^2\) with \(\mu+\theta_i\). For jointly normal quantities, the conditional mean is the mean of the target plus its covariance with the observation, divided by the observation's variance, times the observation's deviation from its mean. Substitution gives:

$$ E[\mu+\theta_i\mid \bar X_i] =\mu+\frac{\operatorname{Cov}(\mu+\theta_i,\bar X_i)} {\operatorname{Var}(\bar X_i)}(\bar X_i-\mu) =\mu+\frac{\tau^2}{\tau^2+\sigma^2/n}(\bar X_i-\mu). $$

The result explains why multiple comparable periods can improve the estimate: the process-noise contribution in the denominator declines as \(n\) grows. If \(\tau^2=0\), there are no persistent risk differences for the experience to identify, and the credibility weight is zero. If process noise is zero, the observed average identifies the persistent level exactly, and the weight is one.

Worked Examples

Worked Example: Credibility After Four Comparable Periods

A hypothetical plan uses a baseline index of \(\mu=1.00\). For a comparable uncapped index, assume \(\theta_i\sim N(0,0.09)\) and independent period errors \(\varepsilon_{it}\sim N(0,0.16)\), independent of \(\theta_i\). The four observed indices are 1.10, 1.40, 1.20, and 1.30. Find the estimated persistent index and explain how much of a single-period variance is process noise.

Method. Use the normal random-effects credibility formula. The specified variances are \(\tau^2=0.09\) and \(\sigma^2=0.16\); they are variances, not standard deviations.

First calculate the observed average:

$$ \bar X_i=\frac{1.10+1.40+1.20+1.30}{4} =\frac{5.00}{4}=1.25. $$

The average's process-noise variance is \(0.16/4=0.04\). Its total variance is \(0.09+0.04=0.13\), so its credibility is:

$$ Z_4=\frac{0.09}{0.13}=0.6923. $$

The average is \(1.25-1.00=0.25\) above baseline. The estimated persistent index is therefore:

$$ 1.00+0.6923(0.25)=1.1731. $$

For one period, total variance is \(0.09+0.16=0.25\). The process-noise share is \(0.16/0.25=0.6400\), or 64.00%. Check: the four-period estimate lies between the baseline 1.00 and observed average 1.25, as a credibility-weighted estimate should. The common mistake is using the single-period weight \(0.09/(0.09+0.16)=0.3600\) even though four periods were averaged; that fails to reduce process noise for repeated observations.

Worked Example: Estimating Variance Components From a Balanced Panel

In a hypothetical study, three risks each have three comparable annual indices. The observations are: Risk A, 0.80, 1.00, 1.20; Risk B, 1.00, 1.20, 1.40; Risk C, 1.20, 1.40, 1.60. Assume the random-effects model above with independent, equal-variance normal effects and errors. Estimate the process variance, persistent-risk variance by the balanced-panel method of moments, and the correlation between two periods for the same risk.

Method. Calculate the within-risk mean square to estimate \(\sigma^2\), then use the between-risk mean square to estimate \(\sigma^2+3\tau^2\). Each risk has \(n=3\) observations.

The risk averages are \(1.00\), \(1.20\), and \(1.40\); the grand average is \((1.00+1.20+1.40)/3=1.20\). Within each risk, deviations from its average are \(-0.20, 0, 0.20\). Each risk therefore has squared-deviation sum \(0.04+0+0.04=0.08\). The total within sum of squares is \(3(0.08)=0.24\), with \(3(3-1)=6\) degrees of freedom:

$$ MS_{\text{within}}=\frac{0.24}{6}=0.04, \qquad \widehat{\sigma^2}=0.04. $$

The sum of squared deviations of the three risk averages from the grand average is \(0.20^2+0^2+(-0.20)^2=0.08\). The between-risk sum of squares multiplies this by the number of periods, three, giving \(0.24\). With \(3-1=2\) degrees of freedom:

$$ MS_{\text{between}}=\frac{0.24}{2}=0.12. $$

For a balanced panel, \(E[MS_{\text{within}}]=\sigma^2\) and \(E[MS_{\text{between}}]=\sigma^2+n\tau^2\). Thus:

$$ \widehat{\tau^2}=\frac{0.12-0.04}{3}=0.0267. $$

Two observations on the same risk have covariance \(\tau^2\) and individual variance \(\tau^2+\sigma^2\). The estimated correlation is:

$$ \hat\rho=\frac{0.0267}{0.0267+0.04}=0.4000. $$

Check: the between-risk mean square, \(0.12\), exceeds the within-risk mean square, \(0.04\), producing a nonnegative persistent variance estimate. The common mistake is to set the between-risk mean square equal to \(\tau^2\); it also contains process variance, so the process component must be subtracted before dividing by \(n\).

Worked Example: Unequal Exposure Across Periods

A hypothetical account has observations from two periods with exposures \(e_1=1\) and \(e_2=3\), measured in the same exposure units. Its comparable indices are \(X_{i1}=1.40\) and \(X_{i2}=1.10\). Assume \(\mu=1.00\), \(\theta_i\sim N(0,0.04)\), and independent \(\varepsilon_{it}\sim N(0,0.36/e_t)\), independent of \(\theta_i\). Find the exposure-weighted average and the estimated persistent index.

Method. With error variance \(\sigma^2/e_t\), each period contributes precision proportional to its exposure. The exposure-weighted average has process variance \(\sigma^2/(e_1+e_2)\).

The weighted average is:

$$ \bar X_w=\frac{1(1.40)+3(1.10)}{1+3} =\frac{1.40+3.30}{4}=1.175. $$

Total exposure is \(4\), so the process variance of this average is \(0.36/4=0.09\). Its total variance is \(0.04+0.09=0.13\), giving:

$$ Z_w=\frac{0.04}{0.13}=0.3077. $$

The estimated persistent index is:

$$ 1.00+0.3077(1.175-1.00)=1.0538. $$

Check: the estimate lies between 1.00 and 1.175. The higher-exposure observation receives three times the weight of the lower-exposure observation. The common mistake is to take the unweighted average, \(1.25\), which gives the less-exposed period too much influence under the stated error-variance model.

Conditions, Interpretation, and Common Traps

The decomposition is useful only when the observations measure the same underlying quantity on a consistent basis. Changing coverage, exposure definition, trend treatment, loss-development basis, or included claim types can create apparent period variation that is neither the modeled persistent effect nor ordinary process noise. In those cases, standardize the observations or extend the model to represent the change before interpreting variance components.

The basic model also assumes that the persistent effect stays constant over the observation window and that process errors are independent across periods. If an underlying risk genuinely changes, or if weather, inflation, or another shared factor creates correlated period errors, covariance across periods need not equal \(\tau^2\). A time trend or common shock can be mistaken for persistent differences unless it is addressed explicitly.

  • Do not infer persistence from one noisy period. One observation contains both \(\tau^2\) and \(\sigma^2\); it cannot separate them by itself.
  • Do not confuse standard deviations with variances. The formulas add variances. If inputs are standard deviations, square them first.
  • Do not average unequal exposures equally without justification. When process variance falls in inverse proportion to exposure, the appropriate average is exposure weighted.
  • Do not apply an uncapped additive model blindly to clipped modifications. Floors and caps alter observed variation and can hide the size of the underlying signal.

Separating process noise from risk differences provides a disciplined interpretation of renewal-to-renewal movement: repeated comparable observations reduce the influence of random fluctuation, while a persistent deviation remains. This is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

For each problem, use the stated model and distinguish persistent variation from process noise.

  1. A hypothetical index has \(\tau^2=0.05\), \(\sigma^2=0.20\), and four equally weighted independent periods. Find the reliability of the four-period average and its variance. (Answer: \(Z_4=0.5000\); variance \(=0.1000\).)
  2. Under a normal random-effects model, \(\mu=1.00\), \(\tau^2=0.12\), and \(\sigma^2=0.27\). A risk's three-period average is 1.35. Find the credibility estimate of its persistent index. (Answer: \(Z_3=0.5714\); estimated index \(=1.2000\).)
  3. Assume \(\mu=1.00\), \(\tau^2=0.02\), and process-error variance \(0.64/e_t\). Two observations are 1.30 at exposure 2 and 1.10 at exposure 6. Find the weighted average and credibility estimate. (Answer: weighted average \(=1.1500\); credibility \(=0.2000\); estimate \(=1.0300\).)
  4. In a repeated-observation model, \(\tau^2=0.04\) and \(\sigma^2=0.06\). Find the variance of one observation and the correlation between two periods for the same risk. (Answer: variance \(=0.1000\); correlation \(=0.4000\).)
  5. A balanced panel has four observations per risk, with estimated within mean square \(0.09\) and between mean square \(0.21\). Use the method-of-moments formulas to estimate \(\sigma^2\) and \(\tau^2\). (Answer: \(\widehat{\sigma^2}=0.09\); \(\widehat{\tau^2}=(0.21-0.09)/4=0.0300\).)

Previous

Next

Back to Tutorials
Clinical Biostats