Every method, stated plainly enough to be argued with.

Losses are arranged into a triangle: accident years down the side, development age across the top. Each cell holds cumulative losses for that accident year at that age. The most recent accident year has the fewest development periods, which is what gives the array its triangular shape.

Age-to-age link ratios are computed as the ratio of cumulative losses at one age to the age before. A selected development factor for each age interval is then applied cumulatively to produce a factor to ultimate, and applying that to the latest diagonal gives projected ultimate losses. The difference between projected ultimate and reported-to-date is the IBNR.

Three selection methods are offered and shown side by side. The volume-weighted average — the sum of losses at the later age divided by the sum at the earlier age, restricted to years present at both — gives more weight to larger accident years and is the standard selection. The simple average treats each year equally, which is more responsive to recent change and noisier. The median is the most robust to a single distorting year.

The selection is not a detail. Changing it moves the answer, sometimes materially, which is why the platform requires you to choose it explicitly and records the choice in the audit vault rather than defaulting silently.

A triangle covering five or six years does not capture development beyond its final observed age, and casualty lines in particular continue to develop for years after that. A tail factor extends the projection past the end of the observed data.

The platform requires a rationale in writing whenever a tail factor other than 1.000 is applied. A tail factor is one of the easiest ways to move a feasibility conclusion without appearing to change anything, so it is treated as an assumption requiring justification rather than a slider.

Chain ladder has a specific weakness: it multiplies whatever is currently reported. In an immature accident year, where very little is reported, a large development factor amplifies a small and unstable number, and the result can swing wildly.

Bornhuetter-Ferguson addresses this by blending in an a priori expectation. Expected ultimate losses are estimated independently — typically from an expected loss ratio applied to premium, or from a loss rate applied to exposure — and IBNR is computed as that expectation multiplied by the proportion of development still to come. The estimate becomes the sum of what is actually reported and that expected unreported amount.

The practical consequence is that BF is more stable for immature years and chain ladder is more responsive for mature ones. The two are always shown side by side rather than blended into a single number, because the divergence between them is diagnostic information in its own right. A large divergence tells you the data is thin and the conclusion is fragile.

Frequency — claims per unit of exposure — and severity — average cost per claim — are trended separately, because they behave differently and they respond to different management action. Frequency responds to loss control. Severity responds to medical inflation, wage inflation, and litigation environment, none of which a safety programme reaches.

Both are fitted by log-linear regression against year, which estimates a constant compound rate of change. The R-squared of each fit is reported. Where the fit is weak, the platform says so and treats the trend as unreliable rather than applying it as though it were established.

Exposure adjustment happens before trending. Without it, a growing business looks like a deteriorating one: five years of rising loss totals can conceal a flat or improving loss rate per unit of exposure.

Aggregate retained loss is simulated using a collective risk model. Each iteration draws a claim count from a frequency distribution, draws a severity for each claim from a severity distribution, applies the per-occurrence retention to each claim, and sums the retained amounts. Ten thousand iterations produce a distribution of possible annual outcomes.

Severity is modelled lognormally. That choice is not arbitrary: casualty severity is right-skewed and strictly positive, with a small number of very large claims dominating the tail, and the lognormal captures that shape while remaining tractable. The parameters are derived by method of moments from the observed mean and an assumed coefficient of variation.

Frequency is modelled as Poisson when the observed variance is close to the mean, and as negative binomial when the variance materially exceeds it. That test matters. Using Poisson on overdispersed data understates volatility, which understates required capital — the more dangerous of the two possible errors.

Where per-claim detail is unavailable, a severity coefficient of variation of 2.0 is assumed. This is deliberately conservative and always disclosed. Understating volatility understates the capital the captive needs, so the assumption errs toward requiring more rather than less.

Every simulation stores its random seed. A number in a board report that cannot be reproduced is not evidence of anything, so the seed is recorded and re-running the engine with it reproduces the figures exactly.

Indicated premium is built as the loss provision at a chosen confidence level, plus an expense load built line by line, plus any fronting fee computed on gross premium. The expense load is never applied as an unexplained percentage of premium.

Funding at the expected level means roughly a one-in-two chance of the year exhausting the premium, which is why regulators generally expect a margin above it. The platform reports the expected, 75th, 90th, and 95th percentile levels.

A second table translates common risk margin multipliers — 1.10, 1.25, 1.50 times expected — into the confidence level each actually buys on your own simulated distribution. Practitioners routinely quote a multiplier without stating what confidence it corresponds to, and the answer is frequently below the 75th percentile. Showing it turns a comfortable convention into a number a board can interrogate.

Five-year projections are produced under three scenarios: base case, moderately adverse, and severe adverse. The severe adverse case is the one that earns the study its fee. A base-case projection tells a board what it hopes will happen; the severe case tells it whether the company survives being wrong.

Premium-to-surplus leverage is tested against domicile expectations, and any year where the projection falls below minimum capital or exceeds the leverage ceiling is flagged with the capital call required to remedy it. Capital calls are shown as explicit line items rather than netted quietly into surplus.

It cannot substitute for a signed actuarial opinion. Most domiciles require a feasibility study signed by a qualified actuary as part of the licence application, and no credentialed actuary has reviewed or certified the output of this platform. It uses the same methods and shows all of its working; it does not provide the signature.

It cannot fix thin data. With three or four accident years and a handful of claims a year, the projections carry wide uncertainty no matter which method is used. The platform reports that plainly rather than presenting a fragile number with false confidence.

It cannot assess whether your particular arrangement constitutes insurance for federal tax purposes. That is a legal determination on your specific facts, and it belongs to tax counsel.

It cannot know what your broker knows about your specific market. Where the commercial market is genuinely mispricing your risk downward, retaining that risk is a worse trade, and no model can see that from loss history alone.

It cannot make the decision. The platform narrows the question and shows its working; a person still answers it. Every development factor, tail factor, confidence level and threshold is a selection someone made, visible and editable, recorded with the reasoning given at the time. A conclusion you cannot interrogate is not analysis.