Work a reserve analysis, a rate indication, and a layer price. Live.

The engine speaks Excel. You send one formula and you get one value back, over plain HTTP, with no key and no account. Nothing about the calculation is stored after the response.

Start with the number every review opens on. AY2023 reported losses are 3,050,000 against 6,350,000 of earned premium. Press Run, or edit the formula first: the cell below is a live request to POST /v1/evaluate, not a screenshot of one.

The catalog holds the Excel library plus a deep actuarial one, called by bare name exactly as you would write them in a cell: RISK.VAR, TVAR, LDF.CDF, PREM.ONLEVEL, CRED.BLEND, and several hundred more.

Here are the year's 20 largest closed claims, as an array literal. Read a percentile over a truncated listing for what it is: the 75th percentile of the top-20 file, which over a 460-claim year sits near the 99th percentile of the whole claim file. The starter is that Value at Risk.

Off the triangle you selected age-to-age factors from 36 months: 1.07, 1.03, 1.01, then a 1.02 tail. The age-to-ultimate factor is their product, and LDF.CDF takes the whole column at once.

Run it, and you have the cumulative development factor that carries 36 months of reported losses to ultimate.

A selected factor is a judgement. Mack’s method gives the same chain-ladder reserve with a standard error attached, and the engine runs the real chainladder package rather than a reimplementation of it.

Library calls pass a fitted-model handle between functions: FIT takes a model and a triangle, GET pulls one fitted attribute, TOTAL sums it. The triangle here is raa, the Reinsurance Association of America’s published incurred triangle, so the answer is checkable against the package itself. Your own triangle goes in the same argument slot as a cell range.

Development answers what AY2023 will ultimately cost at AY2023 price levels. Trend carries that cost to the period the new rates will cover. The book runs a −1% frequency trend and a +5% severity trend, and they combine multiplicatively, not additively: (1 − 0.01) × (1 + 0.05) − 1 = 3.95%.

The trend period is derived, not asserted: the new rates take effect 7/1/2026 for one year of annual policies, so the average written date is 1/1/2027 and the average loss date is near 7/1/2027. From the AY2023 midpoint, 7/1/2023, that is 4.0 years.

Two rate changes were filed since AY2023 was earned: +5% effective 7/1/2024, +3% effective 7/1/2025. A loss ratio over unadjusted premium compares projected losses to prices no longer charged. PREM.ONLEVEL applies the cumulative rate-level index, oldest change first.

One assumption to state out loud: applying each change in full treats every historical dollar as written at the pre-change rate level. That is the extension-of-exposures ideal; when premium is only available in aggregate, the parallelogram method earns each change in geometrically and the factors shrink accordingly.

460 claims is a thin basis for a 0.6624 projected ratio. Limited fluctuation credibility sets the weight: Z = √(n / 1082), with 1,082 the Longley-Cook full-credibility standard for ±5% at 90% confidence under Poisson frequency.

The complement of credibility here is 0.685: the loss and LAE ratio underlying the current rates, trended to the same future period. The complement must live at the same cost and rate level as the experience, or the blend mixes two eras.

You have now run every link by hand: develop, load LAE, trend, on-level, weight by credibility. POST /v1/run holds the chain as one workbook: a template of cells, inputs binding named assumptions into cells, outputs naming the cells you want back.

The last row is the Werner & Modlin loss-ratio-method indication: the credibility-weighted ratio plus the 6.5% fixed expense ratio, divided by the variable permissible loss ratio (1 − 25% variable expense − 5% target profit), minus one. Run it and read indicatedChange: +5.0%.

A workbook binds meaning to coordinates: B20 is the credibility-weighted ratio because row 20 says so. TSRA — a Traceable Spreadsheet Reproducible Artifact — is the same computation written as named declarations instead: input for what the caller supplies (with defaults), let for named intermediate formulas, output for results, and func for reusable functions. It compiles to the same workbook /v1/run executes, so the numbers cannot differ between the two forms.

The program below prices the reinsurance layer the next chapter simulates: expected ceded loss plus a standard-deviation risk load, grossed up for expenses, with rate on line and payback period as functions any program in the bundle could reuse. This editor holds TSRA source, and Run posts it to POST /v1/run-tsra.

Where do 119,702 of expected ceded loss and its standard deviation come from? From this model. Each iteration draws the year’s large-claim count from Poisson(0.8), draws each claim from the lognormal fitted to the book’s five-year large-loss file (μ = 12.6, σ = 0.85), cedes each through LAYER.CEDED(loss, 250000, 500000), and sums. Twelve claim slots bound the draw: P(N > 12) is below 10⁻¹¹, so the truncation never binds. Note the fit is to five years, not to 2023 alone: the year’s own largest claim was 175,000, below the attachment, which is exactly the kind of year a Poisson(0.8) process produces.

The seed is what makes this an audit-grade tool rather than a demo: the same seed over the same workbook draws the same samples, and the seed used comes back in the response, so even an unpinned run can be replayed exactly.

Everything above was plain HTTP, which is why it reaches you wherever you work. The three surfaces below share one evaluator, so the number in a cell, the number your assistant quotes, and the number your script logs are the same number.