Reservoir

The reserve, and every reason for it.

Reserving software for property and casualty actuaries. Every figure traces back to its source, and every change is on the record.

Request trial access Arrange a walkthrough Guided demo, coming soon

Go to source

Select a figure and press Ctrl+[. Reservoir walks it back to the method, the factors and the cells it came from.

Below, Reservoir’s own trace of one cell. Public CAS Schedule P data; State Farm Mut Grp is a dataset name, not a client.

135,525

Selected ultimate, workers compensation, accident year 1997

Selected methodChain ladder, incurred
Latest incurred × factor to ultimate56,003 × 2.41996
Selected link ratios from 12 months, then the tail1.7034 × 1.1628 × 1.0758 × 1.0407 × 1.0286 × 1.0156 × 1.0181 × 1.0075 × 1.0064 × 1.0119
Each ratioVolume-weighted, all years; exponential tail fitted from 36 months

The screen you work in

Reservoir's Ratios view in the dark theme: link ratios for workers compensation, eight averages and a fitted curve, and the selected factors with an exponential tail. — Ratios: link ratios coloured against the selection, eight averages and a fitted curve, the selected factors and the tail. Click a ratio to exclude it; the reason is recorded.

Checked against the papers

Before asking you to trust its numbers, Reservoir reproduces the published ones.

Published result	Paper	Reservoir
Chain-ladder IBNR, RAA data Mack (1994)	52,135	52,135
Mack standard error, RAA Mack (1994)	26,909	26,909
Chain-ladder reserve, Taylor–Ashe Mack (1993)	18,680,856	18,680,856
Mack standard error, Taylor–Ashe Mack (1993)	2,447,095	2,447,095
ODP prediction error, RAA England & Verrall (2002)	18,193	18,195
One-year CDR, root MSEP Merz & Wüthrich (2008)	81,080	81,081
LDF method, Weibull reserve Clark (2003)	21,214,761	21,214,728

Rounded as the papers print them. The ODP difference is the paper’s iterative fit against an exact one; the Clark difference is where two maximum-likelihood searches stop, 0.0002% apart. Merz–Wüthrich and Clark use their papers’ own data.

Methods

Chain ladder, paid and incurred
Bornhuetter–Ferguson
Cape Cod
Benktander
Expected loss ratio
Mack standard error
ODP, analytic and bootstrap
One-year claims development result
Munich chain ladder
Clark growth curves
Berquist–Sherman
Average cost per claim
GLMs with calendar-year trend
Back-testing

Tested against the published worked examples where they exist. Triangles or claim-level data from CSV, Excel or Google Sheets; IFRS 17 and Solvency II outputs; an Excel add-in.

On the record

Built for the reviewer, the auditor and the regulator who will ask how the number was reached.

Every change is a signed command; any tampering shows.
Preparer, reviewer and approver sign separately, with multi-factor authentication.
A report is final only when the team server has verified its sign-offs.
The engine is deterministic: the same figures on any machine, desktop or browser.
The assistant only proposes; each change is previewed and applied when you accept it.

The first of three

Reservoir is built to feed pricing and capital. Tributary and Highwater, in development, run on the same engine and the same set of assumptions.

The reserve risk a capital model needs is already computed: bootstrap ODP, Mack, and the one-year view of Merz–Wüthrich.
Correlation and aggregation across lines, ready to hand on.
One audit trail and one sign-off will cover all three.
One licence per user will cover the whole suite.

Try Reservoir.

The browser trial runs on your computer; data you upload never leaves it. Or book thirty minutes and we will take you through it on our data or yours.

Request trial access Arrange a walkthrough info@ponamus.com