Why I build in the open

The locked libraries

I have spent some twenty years building actuarial models for life and health insurers — six of them inside FIS, working on Prophet itself, and the rest leading model migrations and transformations for insurers in the UK, Europe, the US and Asia, under Solvency II, IFRS 17, US GAAP and NAIC RBC. I mention the geography and the frameworks only to establish one fact: I have seen a great deal of modelling, most of it good, nearly all of it locked.

The best work of my profession sits inside black boxes: vendor engines whose cores no customer may read, and in-house models whose assumptions retire with the people who wrote them. An actuary’s duty is to challenge a model, yet the artefacts the profession depends on resist examination as a condition of their licence. I have spent my career on both sides of that wall, and the wall never stopped bothering me.

The dream and the excuse

The dream was old, and I doubt it was mine alone: modelling engines the profession could open, read and validate. Methodology written to citation grade. Code that can be read alongside the papers it encodes. Validation evidence that demonstrates rather than asserts. Not a vendor — the thing a practising actuary wishes had existed the first time he had to sign off a number he could not fully trace.

The excuse was cost. Building one production-grade engine was measured in vendor-team-years, and a single actuary’s evenings could not touch it. The dream stayed a dream because the arithmetic said so.

The excuse ran out

Frontier AI systems have collapsed that arithmetic. The economics of the collapse are argued in a companion piece, Build vs buy after the cost collapse; here I want only the personal claim. A domain expert directing these systems can now produce, alone, artefacts that once required a vendor team. The scarce input was never engineering hours. It was the judgement to specify what an engine must do, to recognise when its output is wrong, and to know in advance what evidence a validator will demand — and twenty years of that judgement was the part I already had.

I do not ask this to be taken on trust. There are two worked examples:

• The Gnanam ESG: seven stochastic models across interest rates, equity and credit, eight methodology documents, and a regression harness of 155 pytest files locking every output to a bit-identical fingerprint; the interest-rate calibration reprices the swaption surface to roughly 1e-9 basis points. Built in evenings and weekends.
• Gnanam ALM, now in development: an asset-projection and ALM engine for the Bermuda regime’s Scenario-Based Approach — nine prescribed scenarios, a Hull–White lattice for callable bonds, every run sealed into a bit-identical, self-contained evidence bundle, its published benchmark checked against an independent spreadsheet oracle. Built the same way.

One author. No team, no funding round, no organisation behind the curtain. The transformation this essay describes is not a prediction about the industry; it is a fact about two repositories.

Why free, why open

Three reasons, in ascending order of importance to me.

First, the honest commerce. The models are free under Apache-2.0, ungated, nothing withheld. The commercial work is the fitting — your curves, your products, your reporting basis, your control framework — delivered through my service company, Agaram Actuarial Limited. I would rather say that plainly than pretend the openness is charity. It is also self-interest of the best kind: an engine criticised in public becomes a better engine, and a better engine is a better shopfront.

Second, the usefulness. The profession’s bottleneck was never calculation; it is trust. Validation dominates the true cost of any model, and a black box puts a ceiling on what validation can buy. An open engine removes the ceiling: your own actuaries can read it, run it, and break it before they rely on it, and a supervisor can run it as an independent benchmark. A challenger model that can be inspected is itself an act of validation, and that is the sharpest use I hope these engines are put to.

Third, the debt. This profession gave me twenty interesting years. Actuaries share what they know — it is built into how we are trained and examined — yet our artefacts have never been shared the way our papers are. These engines are the return contribution: something a student can learn from, a validator can reproduce, a colleague can take apart and improve. If a stranger’s issue report makes a model better, the system is working.

What I am asking of you

Nothing gated, nothing to sign up for. Take the engines: clone a repository, run the harness, reproduce a published number, and then disbelieve me with evidence in hand. I measure success in the profession’s currency — an independent reproduction, a benchmark run, a cited methodology, a fitting enquiry — not in stars and forks.

The dream was never to build a vendor. It was to build the models I wished had existed and hand them to the community that taught me. You are welcome to check the workmanship.