Sequential Conjugate Updating Across Policy Periods

Why Update Period by Period?

An insurer may receive experience in stages: one policy period closes, its claims are recorded, and another period follows. Bayesian updating allows the posterior after the first period to become the prior for the next. This is convenient operationally, but it creates an important bookkeeping requirement: each period’s observations must enter exactly once.

For conjugate models, the period-by-period calculation can be checked against one update using all the accumulated sufficient statistics. This equivalence is valuable for reviewing calculations, combining experience summaries, and avoiding the common error of adding the same observations twice. The examples below use hypothetical data and assume that the same underlying risk parameter applies across the periods being combined.

Definition: A sequential conjugate update uses the posterior distribution after one period as the prior distribution for the next period, then updates it with that next period’s data. When the observations in all periods are conditionally independent given one fixed parameter, and the likelihood depends on the data through additive sufficient statistics, sequential updating gives the same posterior as one update using all periods together.

Gamma-Poisson Updating Across Exposure Periods

Let $\lambda$ be a claim frequency per unit of exposure. Assume that, conditional on $\lambda$, the number of claims $N_j$ in period $j$ is Poisson with mean $w_j\lambda$, where $w_j$ is known exposure. Counts from different periods are conditionally independent given the same $\lambda$. Use the shape-rate parameterization $\lambda\sim\operatorname{Gamma}(\alpha,\beta)$, with density proportional to $\lambda^{\alpha-1}e^{-\beta\lambda}$. In this parameterization, $E[\lambda]=\alpha/\beta$.

N_j\mid\lambda\sim\operatorname{Poisson}(w_j\lambda), \qquad \lambda\sim\operatorname{Gamma}(\alpha,\beta) \quad\Longrightarrow\quad \lambda\mid N_j\sim \operatorname{Gamma}(\alpha+N_j,\beta+w_j).

Thus each period adds its claim count to the shape and its exposure to the rate. After $k$ periods, the posterior is

\lambda\mid N_1,\ldots,N_k \sim \operatorname{Gamma}\left( \alpha+\sum_{j=1}^{k}N_j,\, \beta+\sum_{j=1}^{k}w_j \right).

Worked Example: Combining Claim Counts Without Double-Counting

A hypothetical risk has prior claim frequency $\lambda\sim\operatorname{Gamma}(2,100)$, using shape-rate notation. In period 1, exposure is 200 units and there is 1 claim. In period 2, exposure is 300 units and there are 2 claims. Find the posterior by sequential updating, verify it with a pooled update, and identify the result of a specified double-counting error.

Method and assumptions. Use a Gamma-Poisson model with the same frequency parameter in both periods, conditionally independent counts, and known exposures. The prior has shape 2 and rate 100.

Period 1. Add the first count to the shape and the first exposure to the rate: $\alpha_1=2+1=3$ and $\beta_1=100+200=300$. The posterior after period 1 is $\operatorname{Gamma}(3,300)$.

Period 2. Use $\operatorname{Gamma}(3,300)$ as the prior for period 2, and add only period 2’s data. The updated shape is $3+2=5$, and the updated rate is $300+300=600$. Therefore the final posterior is $\operatorname{Gamma}(5,600)$.

Pooled check. The total count is $1+2=3$, and total exposure is $200+300=500$. Updating the original prior once gives shape $2+3=5$ and rate $100+500=600$, again yielding $\operatorname{Gamma}(5,600)$. Its posterior mean frequency is $5/600=0.008333$ claims per exposure unit.

Common mistake and correction. Suppose a student starts with the period 1 posterior $\operatorname{Gamma}(3,300)$ and then adds the pooled count and exposure. That adds period 1’s data a second time: the erroneous shape is $3+3=6$, and the erroneous rate is $300+500=800$, giving $\operatorname{Gamma}(6,800)$. The sequential update must add only period 2’s statistics to the period 1 posterior. The correct result is $\operatorname{Gamma}(5,600)$.

Check. The sequential and pooled calculations agree at $\operatorname{Gamma}(5,600)$; the erroneous calculation has larger shape and rate because it counted period 1 twice.

Why the Sequential and Pooled Updates Agree

The equivalence follows from the likelihood’s dependence on the observations. For conditionally independent Poisson counts, the likelihood across $k$ periods is proportional, as a function of $\lambda$, to $\lambda^{\sum_j N_j}\exp(-\lambda\sum_j w_j)$. Multiplying this likelihood by the Gamma prior adds the total count to the prior shape and the total exposure to the prior rate.

\begin{aligned} \pi(\lambda\mid N_1,\ldots,N_k) &\propto \lambda^{\alpha-1}e^{-\beta\lambda} \prod_{j=1}^{k} \left(\lambda^{N_j}e^{-w_j\lambda}\right)\\ &\propto \lambda^{\alpha+\sum_j N_j-1} e^{-(\beta+\sum_j w_j)\lambda}. \end{aligned}

This is the kernel of a Gamma density with shape $\alpha+\sum_jN_j$ and rate $\beta+\sum_jw_j$. Sequential updating produces the same result because addition is associative: adding $N_1$, then $N_2$, gives the same shape as adding $N_1+N_2$ at once, and the same reasoning applies to exposure.

Record the starting prior.
Write its parameterization explicitly; here the Gamma parameters are shape and rate.

Summarize each new period.
For Gamma-Poisson updating, record the period’s claim count and exposure.

Update with new information only.
Add that period’s sufficient statistics to the current posterior parameters.

Check by pooling.
Add all period counts and exposures, then update the original prior once. The two parameter pairs should match.

The equivalence requires one fixed parameter shared across the periods and a correctly specified conditional likelihood. If the frequency parameter changes between periods, or if observations are dependent in a way not represented by the model, simply pooling the statistics need not be appropriate. The calculation also fails as a bookkeeping procedure if some data are omitted or counted twice.

Beta-Binomial Updating by Period

The same additive structure appears when each of $m_j$ independent opportunities in period $j$ produces either a claim or no claim, with common claim probability $p$. Let $Y_j$ be the number of claims in that period. If $p\sim\operatorname{Beta}(a,b)$, then $Y_j$ adds to the first parameter and the number of nonclaims, $m_j-Y_j$, adds to the second. This model describes Bernoulli opportunities; it is not the Poisson exposure model.

Worked Example: Updating a Claim Probability Over Two Periods

A hypothetical coverage has a prior claim probability $p\sim\operatorname{Beta}(3,7)$ per policy-year. In the first period, 20 independent policy-years produce 4 claims. In the second period, 30 policy-years produce 9 claims. Find the sequential posterior and its posterior mean, then check the result using pooled data.

Method and assumptions. Use a Beta-Binomial model with a constant probability $p$, conditionally independent Bernoulli outcomes, and fully observed claim/no-claim outcomes.

Period 1. There are 4 claims and $20-4=16$ nonclaims. The posterior is $\operatorname{Beta}(3+4,7+16)=\operatorname{Beta}(7,23)$.

Period 2. There are 9 claims and $30-9=21$ nonclaims. Updating the period 1 posterior gives $\operatorname{Beta}(7+9,23+21)=\operatorname{Beta}(16,44)$.

Pooled check and posterior mean. Across both periods there are $4+9=13$ claims and $16+21=37$ nonclaims. Updating the original prior gives $\operatorname{Beta}(3+13,7+37)=\operatorname{Beta}(16,44)$. The posterior mean is $16/(16+44)=16/60=0.2667$, to four decimal places.

Common mistake avoided. The second update uses only the second period’s 9 claims and 21 nonclaims. Adding all 13 claims and 37 nonclaims to $\operatorname{Beta}(7,23)$ would count the first period twice. The pooled data belong with the original prior, not with the already-updated posterior.

Sequential Updating for Exponential Claim Severity

Sequential updating also applies to severity. Suppose individual claim amounts $X$ are conditionally independent exponential observations with unknown rate $\theta$, so $f_X(x\mid\theta)=\theta e^{-\theta x}$ for $x\geq0$. Give $\theta$ a Gamma prior in shape-rate form, $\operatorname{Gamma}(\alpha,\beta)$. For a period with $n_j$ fully observed claim amounts and total loss $T_j$, the posterior is $\operatorname{Gamma}(\alpha+n_j,\beta+T_j)$. The data needed for this update are the number of claims and their total amount.

Worked Example: Updating an Exponential Severity Rate Across Periods

A hypothetical insurer models individual claim amounts, measured in thousands of dollars, as exponential with unknown rate $\theta$ per thousand dollars. The prior is $\theta\sim\operatorname{Gamma}(2,4)$, in shape-rate notation. Period 1 has three observed claims of 40, 60, and 100 thousand dollars. Period 2 has two observed claims of 30 and 70 thousand dollars. Find the final posterior for $\theta$ and the posterior expected amount of one future claim.

Method and assumptions. Use the exponential likelihood and a Gamma prior. Assume a common rate across periods, conditionally independent claims, and exact, fully observed claim amounts; no truncation or censoring is present.

Period 1 sufficient statistics. The claim count is $n_1=3$. The total loss, in thousands of dollars, is $T_1=40+60+100=200$. The period 1 posterior is $\operatorname{Gamma}(2+3,4+200)=\operatorname{Gamma}(5,204)$.

Period 2 sufficient statistics. The claim count is $n_2=2$, and the total loss is $T_2=30+70=100$. Updating the period 1 posterior gives shape $5+2=7$ and rate $204+100=304$. Thus $\theta\mid\text{data}\sim\operatorname{Gamma}(7,304)$.

Pooled check. There are $3+2=5$ claims with total loss $200+100=300$ thousand dollars. Updating the original prior gives $\operatorname{Gamma}(2+5,4+300)=\operatorname{Gamma}(7,304)$.

Posterior expected severity. Conditional on $\theta$, an exponential claim has mean $1/\theta$. For a Gamma distribution in shape-rate form, $E[1/\theta]=\beta/(\alpha-1)$ when $\alpha>1$. Therefore the posterior expected amount of one future claim is $304/(7-1)=304/6=50.6667$ thousand dollars, or $50,666.67, to the nearest cent.

Common mistake avoided. The posterior mean of the rate is $7/304$, whose reciprocal is $304/7$. That is not the posterior expected severity: expectation does not commute with taking a reciprocal. The correct calculation is $E[1/\theta\mid\text{data}]=304/(7-1)$.

Conditions, Interpretation, and Exam Traps

The practical advantage of sequential updating is that a posterior can be maintained as periods close, without retaining every individual observation when the sufficient statistics are available. For the models here, those summaries are simple: claim count and exposure for Gamma-Poisson frequency; successes and failures for Beta-Binomial frequency; and claim count and total loss for exponential severity with a Gamma rate prior.

The arithmetic is straightforward, but model and parameterization checks matter. In particular:

Keep the same parameter. Combining periods assumes that they inform a shared frequency, probability, or severity parameter. A changed risk regime may require a different model.
Use the correct sufficient statistics. Poisson updating adds exposure to the Gamma rate; Beta-Binomial updating adds nonclaims to the second Beta parameter; exponential-severity updating adds total loss to the Gamma rate.
Do not mix parameterizations. The Gamma formulas here use shape and rate, not shape and scale. A scale parameter is not added to exposure or total loss in these formulas.
Count observations once. When the current distribution is already a posterior, add only the newly observed period’s statistics.
Respect the likelihood assumptions. The exponential severity calculation uses exact, uncensored claim amounts. Data observed under truncation or censoring require a likelihood that reflects that observation process.

A useful final check is to calculate the posterior both ways whenever pooled summaries are available: update period by period, then update the original prior with the accumulated sufficient statistics. Matching parameters confirm the bookkeeping for a shared-parameter conjugate model. Disagreement signals a likely arithmetic, parameterization, or double-counting error.

Key takeaway: Sequential conjugate updating is valid when each period contributes conditionally independent information about the same parameter. Add each period’s sufficient statistics once; the resulting posterior must match a single update using their totals.

Check Your Understanding

For each problem, state the posterior parameters and show the period-by-period additions before checking against pooled statistics.

A claim frequency has prior $\lambda\sim\operatorname{Gamma}(1.5,80)$. Two periods have exposures 120 and 180, with 2 and 1 claims, respectively. Find the final posterior. (Answer: $\operatorname{Gamma}(4.5,380)$.)
A claim probability has prior $p\sim\operatorname{Beta}(2,5)$. Period 1 has 12 opportunities and 3 claims; period 2 has 8 opportunities and 1 claim. Find the final posterior. (Answer: $\operatorname{Beta}(6,21)$.)
Claim amounts in thousands of dollars are exponential with unknown rate $\theta$, and $\theta\sim\operatorname{Gamma}(2,5)$. Period 1 has two claims totaling 80; period 2 has three claims totaling 120. Find the posterior for $\theta$. (Answer: $\operatorname{Gamma}(7,205)$.)
A frequency prior is $\operatorname{Gamma}(2,100)$. Period 1 records 1 claim at exposure 200, and period 2 records 2 claims at exposure 300. A student updates period 1, then adds the pooled count and exposure. Give the correct posterior and the student’s erroneous posterior. (Answer: correct $\operatorname{Gamma}(5,600)$; erroneous $\operatorname{Gamma}(6,800)$.)
For the severity model in Question 3, calculate the posterior expected amount of one future claim, in thousands of dollars. (Answer: $205/(7-1)=34.1667$ thousand dollars, to four decimal places.)

What You'll Learn