What ships

• Loss Table — the simulated loss distribution of ONE instrument: one loss ratio per simulated year, in the simulation's own year order. Every summary figure is derived, never stored — mean loss, attachment frequency, 99% VaR and TVaR, and the exceedance curve at 1-in-10 to 1-in-250.
• Optimization Run — one optimization of a book: the question (book, method, budget, candidates) and the answer (what to deploy, and what that does to the book's RoRAC) stored together, so a run is evidence rather than a screenshot.
• Workbench — the studio at /PortfolioOptimization/Studio: every book on the mesh with its invested notional and RoRAC, every loss table, every past run. Derived live; stores nothing.

The model, in four lines

• Book loss per simulated year = Σ amount × loss ratio. Every table in a book comes from the same simulation, so year i is one jointly realised year — a sample-count mismatch is refused by name, never silently mis-paired.
• RoRAC = (spread income − expected loss) ÷ 99% TVaR. A book with no measurable tail reads 0, never infinity.
• Marginal RoRAC = the extra net return a slice buys per extra unit of tail capital. A slice that costs no extra capital is quoted on the candidate's own terms instead.
• Greedy budget = deploy in slices, each to whichever candidate raises the book's RoRAC most, respecting every cap, stopping when nothing improves.

Generic by construction

An instrument is a node PATH with a loss distribution.

• The held book is read from the Portfolio module's nodes generically, by property name.
• The portfolio-marginal view is pushed into an ILS analysis by value and node path.
• Neither module is referenced in C#, and neither is required: the engine works for anything with a loss (or return) distribution per instrument.

Contents