Why Simulate Portfolio Price Outcomes?
A price plan produces uncertain outcomes when customers may renew or leave and losses vary from one period to another. A single expected value can summarize the average result, but it does not show how often a plan might produce a weak contribution or a low retention level. Monte Carlo analysis uses repeated simulated portfolio outcomes to estimate both expected results and the distribution around them.
The unit of analysis is a simulated renewal period for the portfolio. In each trial, the model generates customer renewal outcomes and any other uncertain quantities included in the model, such as losses. Those outcomes are aggregated into portfolio premium, retention, and contribution. The trial must use a coherent set of outcomes: for example, a policy that does not renew should not generate renewal premium or renewal-period losses in that trial.
For a portfolio with \(M\) policies, let \(I_{ij}=1\) if policy \(i\) renews in trial \(j\), and \(I_{ij}=0\) otherwise. If \(p_i\) is the premium for a renewing policy, then trial premium is \(P_j=\sum_{i=1}^{M} I_{ij}p_i\), and the number retained is \(K_j=\sum_{i=1}^{M} I_{ij}\). A count-based retention rate for a fixed renewal cohort is \(R_j=K_j/M\). Other retention measures, such as premium-weighted retention, are also possible, but the chosen definition must be stated and applied consistently.
For illustration, define trial contribution as premium less variable expenses, fixed expenses, and losses. If variable expenses are a fraction \(e\) of premium and fixed expense is \(f\) per retained policy, then \(C_j=P_j-eP_j-fK_j-L_j\), where \(L_j\) is the total loss in that trial. This definition is a modeling choice: an analysis should match the expense and loss definitions to the financial measure it intends to estimate. Premium, expenses, and losses must also refer to consistent coverage periods and accounting bases.
From Trial Outcomes to Portfolio Summaries
The simulated mean of an outcome is its arithmetic average across trials. For example, the estimated mean contribution is \(\bar{C}=N^{-1}\sum_{j=1}^{N}C_j\). The estimated probability of negative contribution is the fraction of trials in which contribution is below zero. An empirical quantile gives a further view of the downside: under the nearest-rank convention, sort the \(N\) values from smallest to largest and select the value at rank \(\lceil qN\rceil\) for quantile level \(q\).
The mean and the downside summaries answer different questions. The mean estimates average performance under the model. The probability of a negative result estimates how often the simulated result falls below a specified threshold. A lower quantile describes a point in the simulated distribution, not a guaranteed minimum or a forecast that a particular outcome will occur.
Worked Example: Summarizing Four Portfolio Trials
A hypothetical three-policy renewal cohort is simulated under one price plan. In each trial, the table gives total premium, number retained, and total loss. Variable expense is 12% of premium, and fixed expense is $20 per retained policy. Calculate mean premium, mean count-based retention, mean contribution, and the empirical probability of negative contribution.
Method. Calculate contribution within each trial using \(C_j=P_j-0.12P_j-20K_j-L_j\). Then average the trial-level results. This method preserves the relationship among premium, retention, and losses within each simulated outcome.
| Trial | Premium | Retained | Loss | Contribution calculation | Contribution |
|---|---|---|---|---|---|
| 1 | $900 | 3 | $510 | $900 − $108 − $60 − $510 | $222 |
| 2 | $700 | 2 | $650 | $700 − $84 − $40 − $650 | −$74 |
| 3 | $1,000 | 3 | $980 | $1,000 − $120 − $60 − $980 | −$160 |
| 4 | $800 | 2 | $520 | $800 − $96 − $40 − $520 | $144 |
Mean premium is \((900+700+1{,}000+800)/4=3{,}400/4=\$850\). The mean number retained is \((3+2+3+2)/4=10/4=2.5\), so mean count-based retention is \(2.5/3=0.8333\), or 83.33%.
Mean contribution is \((222-74-160+144)/4=132/4=\$33\). Two of the four trial contributions are negative, so the empirical probability of negative contribution is \(2/4=0.5000\). As a check, aggregating the trial-level components gives mean variable expense \(0.12(850)=\$102\), mean fixed expense \(20(2.5)=\$50\), and mean loss \((510+650+980+520)/4=\$665\). Thus \(850-102-50-665=\$33\), agreeing with the direct average.
Common mistake avoided: Do not calculate contribution for one trial using the premium from another trial, or omit fixed expense in only some trials. Each row is a complete simulated portfolio outcome, and its contribution must use that row’s own premium, retained count, and losses.
Why the Simulated Mean Estimates an Expected Outcome
Suppose the simulation produces independent, identically distributed trial outcomes \(C_1,\ldots,C_N\), each with finite mean \(E[C_j]=\mu_C\) and variance \(\operatorname{Var}(C_j)=\sigma_C^2\). These conditions mean that trials are generated under the same model and do not share random draws in a way that makes their outcomes dependent. By linearity of expectation,
Therefore, the sample mean is unbiased for the model’s expected contribution under those conditions. Independence also gives
Since \(\sigma_C\) is usually unknown, estimate it with the sample standard deviation \(s_C\), calculated using denominator \(N-1\). The estimated Monte Carlo standard error of the mean is \(s_C/\sqrt{N}\). It measures simulation noise in estimating the model’s expected contribution. It does not measure whether the model assumptions, selected rates, or customer behavior are correct.
Worked Example: Estimating the Precision of Mean Contribution
In a hypothetical simulation, five independent trials produce contributions of $80, −$20, $120, −$60, and $30. Calculate the estimated mean, sample standard deviation, Monte Carlo standard error, and an approximate normal-based 95% interval for the model’s mean contribution. Treat the five contributions as a random sample from a distribution with finite variance.
Method. Use the sample mean and sample standard deviation, then divide the latter by the square root of the number of trials. All amounts in the calculations below are dollars.
The mean is \((80-20+120-60+30)/5=150/5=\$30\). Deviations from the mean are \(50,-50,90,-90,0\). Their squared values sum to \(2{,}500+2{,}500+8{,}100+8{,}100+0=21{,}200\). The sample variance is \(21{,}200/(5-1)=\$5{,}300^2\), and the sample standard deviation is \(\sqrt{5{,}300}=\$72.80\) to the nearest cent.
The estimated Monte Carlo standard error is \(72.80/\sqrt{5}=\$32.55\) to the nearest cent. Using the normal approximation, the interval is \(30\mathbin{\pm}1.96(32.55)\), or \(30\mathbin{\pm}63.80\). Its endpoints are approximately −$33.80 and $93.80. Because only five trials were run, this interval is a rough diagnostic, not a reliable precise interval. More trials would generally reduce simulation error. The observed fraction of negative contributions is \(2/5=0.4000\), but that estimate is also very imprecise with only five trials.
Common mistake avoided: Dividing the sample standard deviation by \(N\) understates the estimated standard error. The standard error of the sample mean uses \(s_C/\sqrt{N}\), not \(s_C/N\).
Modeling Renewal and Loss Uncertainty Together
A useful simulation represents the uncertain quantities that drive the outcome rather than applying a single portfolio adjustment after the fact. A policy’s renewal indicator can be generated using its modeled renewal probability. Conditional on renewal, a loss outcome can then be generated from the loss model used for the analysis. If losses, renewals, or other policy outcomes are dependent, the simulation should preserve that dependence. For example, generating each policy independently when the model specifies shared conditions can understate the frequency of portfolio-wide adverse outcomes.
The following example treats trial records as generated outcomes from a stated hypothetical model. It shows how policy-level outcomes roll up into the portfolio; it does not imply that four trials are enough for a stable estimate of the underlying distribution.
Worked Example: Combining Renewal and Loss Outcomes
A hypothetical cohort has two policies. Policy A has a $600 renewal premium and renews with probability 0.80; conditional on renewal, its simulated loss is either $400 or $700. Policy B has a $1,000 renewal premium and renews with probability 0.60; conditional on renewal, its simulated loss is either $500 or $1,300. A policy that does not renew has no premium or loss in the renewal period. Four simulation trials yield the outcomes shown below. Variable expense is 10% of premium and fixed expense is $25 per retained policy. Calculate mean premium, mean retention rate, mean contribution, and the empirical 25th percentile of contribution using nearest rank.
Method. Aggregate the renewal indicators and losses in each trial, then calculate contribution with the stated expense formula. The modeled renewal probabilities describe how outcomes are generated; the realized simulation records, rather than those probabilities alone, determine these four empirical summaries.
| Trial | A outcome | B outcome | Premium | Retained | Loss | Contribution |
|---|---|---|---|---|---|---|
| 1 | Renews; $400 loss | Renews; $500 loss | $1,600 | 2 | $900 | $1,600 − $160 − $50 − $900 = $490 |
| 2 | Renews; $700 loss | Does not renew | $600 | 1 | $700 | $600 − $60 − $25 − $700 = −$185 |
| 3 | Does not renew | Renews; $1,300 loss | $1,000 | 1 | $1,300 | $1,000 − $100 − $25 − $1,300 = −$425 |
| 4 | Renews; $400 loss | Renews; $1,300 loss | $1,600 | 2 | $1,700 | $1,600 − $160 − $50 − $1,700 = −$310 |
Mean premium is \((1{,}600+600+1{,}000+1{,}600)/4=\$1{,}200\). Mean retained count is \((2+1+1+2)/4=1.5\), giving mean count-based retention \(1.5/2=0.7500\). Mean contribution is \((490-185-425-310)/4=-430/4=-\$107.50\).
The sorted contributions are −$425, −$310, −$185, and $490. For \(N=4\) and \(q=0.25\), nearest rank selects rank \(\lceil0.25(4)\rceil=1\), so the empirical 25th percentile is −$425. The estimate is highly dependent on this small sample; it is not a stable estimate of the model’s 25th percentile.
Common mistake avoided: A nonrenewing policy contributes neither renewal premium nor renewal-period loss here. Counting the loss for a policy that left, while excluding its premium, would mix incompatible exposure definitions and distort the result.
Interpreting Results and Avoiding Misleading Precision
Monte Carlo results are conditional on the model and inputs used to generate the trials. A large number of trials can make estimates more stable with respect to random simulation, but it cannot correct an inaccurate renewal model, an incomplete loss model, an inconsistent expense basis, or an omitted source of dependence. Accordingly, report both the number of trials and the outcome definitions, and distinguish simulation precision from broader uncertainty about the assumptions.
Portfolio summaries should also be computed in the order that matches the question. For a fixed cohort of \(M\) policies, the average trial-level count retention rate equals average retained count divided by \(M\). If the denominator varies by trial, however, average trial-level retention rates and total retained divided by total eligible are generally different summaries. The analyst should state which one is intended rather than treating them as interchangeable.
- Use a consistent time basis: pair renewal-period premium with expenses and losses from that same period.
- Calculate contribution trial by trial: use each trial’s own premium, losses, and retained count before summarizing.
- Label downside statistics precisely: state the threshold for a downside probability and the quantile convention for an empirical percentile.
- Do not overinterpret small simulations: rare outcomes and tail quantiles require enough simulated trials to be represented meaningfully.
- Separate two kinds of uncertainty: Monte Carlo error is variability from the finite set of random trials; model uncertainty concerns whether the model and assumptions represent the situation adequately.
Monte Carlo analysis supports comparison of price plans by making the distribution of modeled portfolio outcomes visible, not just the average. The results are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus when the simulation is tied to a clearly defined renewal cohort, consistent financial measures, and appropriately interpreted estimates.
Check Your Understanding
For each problem, use the stated outcome definitions and keep calculations within each trial before summarizing.
- A two-policy cohort has three simulated trials with premiums of $500, $700, and $600, and retained counts of 2, 1, and 2. What are mean premium and mean count-based retention? (Answer: Mean premium $600; mean retention \(5/6=0.8333\), or 83.33%.)
- Two simulated outcomes have premium and loss pairs of ($1,200, $900) and ($800, $750). Variable expense is 15% of premium, with fixed expense of $30 per trial. Calculate each contribution, mean contribution, and empirical probability of negative contribution. (Answer: Contributions $90 and −$100; mean contribution −$5; negative-contribution probability 0.5000.)
- Four simulated contributions are $40, −$10, −$30, and $80. Find mean contribution and empirical probability of negative contribution. (Answer: Mean contribution $20; negative-contribution probability 0.5000.)
- A fixed four-policy cohort has simulated retained counts of 4, 3, and 3 across three trials. Find the average trial-level count-based retention rate. (Answer: \((1+0.75+0.75)/3=0.8333\), or 83.33%.)
- Four independent simulated contributions, measured in thousands of dollars, are 2, 4, 6, and 8. Find the sample mean, sample standard deviation, and estimated Monte Carlo standard error. (Answer: Mean $5,000; sample standard deviation \(\sqrt{20/3}=\$2,582\) approximately; standard error \(\sqrt{5/3}=\$1,291\) approximately.)