| Vose Software

Industry: Insurance and Reinsurance
Product: ModelRisk
Application: Claims Risk Assessment


Reserving the Tail: Probabilistic Claims-Risk Modeling for a Multi-Line P&C Portfolio

A reserve-strengthening surprise is the kind of event that ends a CFO's quarter and triggers a regulator letter. It is also exactly what an actuarial best-estimate plus a fixed risk margin cannot predict — because both the chain-ladder point estimate and the margin sit on the assumption that next year's frequency and severity behave like a smoothed average of past years. The Monte Carlo question is different: across the distribution of plausible accident-year outcomes, where in the right tail does the booked reserve sit, and how much capital does the gap between the mean and P99.5 actually consume?

Annual aggregate loss — $1.2B-premium portfolio

The shape above is the whole story. A multi-line P&C carrier with $1.2B of annual gross written premium across personal auto, homeowners, and commercial liability rebuilt its portfolio claims-risk model in ModelRisk, and the simulated aggregate loss runs to a mean near $852M with a 99.5% tail above $1.36B. The driving need was a unified, transparent reserving-risk view that the reserve actuary, the chief risk officer, and the Solvency II model owner could read off a single output — replacing three siloed spreadsheets that had been disagreeing with each other for years.

Frequency and severity, line by line

For each line of business the model is a compound frequency-severity distribution, with parameters from the carrier's own closed-file data:

Line Frequency Severity (LogNormal)
Personal auto NegBin(20, mean 12,500) μ = ln 11.5K, σ = 1.10
Homeowners NegBin(20, mean 3,400) μ = ln 28K, σ = 1.35
Commercial liability NegBin(20, mean 820) μ = ln 115K, σ = 1.60

Two distribution choices matter. First, frequency is Negative Binomial, not Poisson — claim counts in real portfolios are over-dispersed because shared shocks (storm seasons, inflation regimes) drive several files at once, and the variance-to-mean ratio is consistently above 1.5. Modelling counts as Poisson, the most common in-house simplification, systematically understates count volatility. Second, severity is LogNormal because claim sizes are bounded below by zero, right-skewed, and span several orders of magnitude — but the σ rises sharply for the longer-tail lines, with commercial liability needing σ ≈ 1.6 to fit historical large-loss frequency. Where individual losses can exceed two or three standard deviations of LogNormal-implied scale, the severity body is supplemented with a peaks-over-threshold Generalized Pareto fit so the tail is not undersold.

Correlation is where the diversification illusion breaks

The reserve actuary's instinct, and the deterministic model's assumption, is that the three lines are largely independent — and so portfolio capital scales as the square root of the sum of per-line variances. The data does not support that. The realised inter-LoB correlation matrix, with two external risk drivers added, looks like this:

Where diversification is not as diversified as it looks

Personal auto and homeowners correlate at 0.45 because weather shocks and inflation regimes hit both. Homeowners and the weather index correlate at 0.58 — almost a co-movement. A model that ignores these correlations underprices the diversification benefit precisely when the carrier needs that benefit most: in stressed years when correlations rise toward 1. The portfolio engine therefore couples the per-LoB severity draws through a Gaussian copula calibrated to the matrix above, then samples a common inflation shock per trial that scales every line's severity body.

What the deterministic best estimate misses

The deterministic best estimate fixes on the mean and the median, which sit close together at roughly $850M. The Monte Carlo simulation, with 80,000 trials of the full compound model, says the same chart carries far more information than that single number:

The simulation's mean is $852M and its median $837M. The 99th percentile is $1.30B — roughly 1.5× the mean — and the 99.5th percentile, the Solvency II SCR anchor, runs to $1.36B, with an expected shortfall beyond it of $1.46B. The probability that aggregate loss exceeds 1.4× the median (about $1.17B) is 3.8%, a low-frequency but capital-defining event. The point estimate was not wrong about the average; it was silent about the shape, and the shape was where the capital question lived.

Development risk: the cone the chain-ladder hides

Even within an accident year, ultimate loss is uncertain because of development. Mack-style chain-ladder gives a single curve to ultimate; the Monte Carlo version gives a cone:

IBNR development cone for accident year 2024

The P10–P90 cone narrows as development matures, but at end of development year 1 the P90 ultimate is roughly 18% above the booked best estimate. The cone is the actuary's defence when a year develops adversely — the explicit prior probability of an 18% strengthening was 10%, not zero, and it was on the page from day one.

What moves the 99.5% number

Sensitivity ranking on the P99.5 ultimate loss identifies where the reserving team should spend its next dollar of analytical effort:

What moves the 99.5% reserve number

Severity tail σ is by far the largest mover, which directs the next data project to better fit of the right tail on the longest-tail line (commercial liability). Inter-LoB correlation ranks second — every five-point change in average pairwise correlation moves the SCR meaningfully, which is what makes correlation calibration a board-level analytical investment rather than a footnote. Frequency over-dispersion ranks third, validating the choice of Negative Binomial over Poisson as a structural decision rather than a stylistic one.

What changed

  • Booked reserves restated by line of business. Total uplift was 3.2% across the book — small as a percentage, but it removed roughly $17M of reserving-risk capital that had previously been held centrally as undifferentiated buffer.
  • Reinsurance retention restructured. The aggregate stop-loss attachment was lowered from $620M to $580M because the new view of the P99 made the existing retention demonstrably under-protective on the commercial-liability tail.
  • Solvency II SCR contribution model rebuilt with explicit per-line and per-driver attribution, replacing a black-box vendor formula. The regulator's questions in the next ORSA dialogue were answered directly off the simulation.
  • A single source of truth. The reserve actuary, the CRO, and the Solvency II owner now read the same numbers off the same engine, and the cross-team reconciliation meetings that used to consume a week per quarter were replaced by a fifteen-minute output review.

ModelRisk Functionality Used

  • Negative Binomial frequency and LogNormal severity per line of business, with σ values fitted off closed-file data and the option to substitute a GPD tail above the large-loss threshold for the longest-tail lines.
  • Gaussian copula linking the three LoB severities to a calibrated correlation matrix, capturing the diversification breakdown stressed years actually exhibit.
  • Mack chain-ladder development extension producing the P10–P90 cone that travels on every accident-year reserve memo.
  • Sensitivity ranking that placed severity-tail σ and correlation strength at the top, reshaping the year's analytical agenda.
  • Direct SCR attribution — the same simulation produces per-line contributions to the 99.5% portfolio number, replacing a vendor black-box with a transparent in-house view that satisfies regulator and rating-agency dialogue.
  • Excel-native authoring, so reserve actuaries iterate in the workbook they already own, and the auditor can open the file and trace every input to its source.

Reserving risk is not the best estimate — it is the shape of the distribution around the best estimate, and the way that shape stretches when correlations rise in a stressed year. Monte Carlo is what makes that shape visible, and once visible, it is what allows capital, reinsurance, and reserve adequacy to be set against the distribution rather than against the mean.