Why this began
The research began in my own retirement cashflow work, where I encountered assumptions that were questionable and influencing the conclusions being presented to clients.
I started by rewriting the assumptions, but wanted to go further. What would a genuinely best-in-class retirement cashflow policy look like?
That became the Capital at Risk methodology. The resulting rules and technical manual were comprehensive, but applying them consistently in everyday practice was another matter. I developed an Execution Sheet to make the process usable, yet the mechanics of cashflow modelling still consumed time.
Take a simple question from a client: “What happens if I retire a year later?”
Even in conventional cashflow software, answering that question can involve changing inputs, creating another scenario, recalculating and comparing the results. Meanwhile, the client is sitting across the table waiting.
I developed Engine B to execute the methodology consistently, and HAL to make it accessible through conversation. The question can be asked naturally, the engine performs the calculation and the chart updates.
The adviser can spend more of the meeting discussing the financial implications with the client.
The FCA’s cashflow modelling findings (20 March 2024) set out the same concerns: input accuracy, justified assumptions, uncertainty and review of outputs.
Governance before calculation
A cashflow projection is only as good as the assumptions fed into it. The methodology settles those assumptions, and the evidence needed before any conclusion is allowed, before a model is run. It is four documents:
- Rules
- What must be true for a conclusion to stand
- Register
- The numbers: growth, stress, charges, inflation
- Execution Sheet
- How a practitioner runs a case
- Manual
- Why each rule exists, and the evidence behind it
The methodology distinguishes financial planning from formal grading. A Current projection shows where the client stands today and is not graded. Clients and advisers can develop and explore financial plans without a formal grade. An Advised plan, where an adviser recommends a strategy, is formally graded against the rules.
The methodology sets projections in today’s money, running to age 99, with investment stresses calibrated to the agreed risk level and their financial consequences assessed against the client’s objectives.
Making the methodology executable
A substantial technical manual cannot sit beside every adviser in every meeting. Engine B applies the rules in code.
The same inputs on the same engine version give the same answer. Every result carries the inputs, assumptions and controls that produced it.
When a conclusion is not supported, Engine B says why, and tells three cases apart: evidence is missing, no strategy has been set, or the circumstances do not justify a grade. It applies governed assumptions. It never invents client facts.
Engine B was tested against independently constructed oracles.
- Foundation Capital at Risk Methodology Sets the evidence requirements, assumptions, rules and decision controls
- implemented in
- Execution Engine B Runs the governed calculations and stress tests, and records the evidence
- structured inputsresults, evidence or HOLD
- Interface HAL Conversational AI and live chart. Asks, explains, shows
- conversation
- Judgement Client / Adviser Asks questions, supplies information, keeps the decision
Making it conversational
HAL takes a question as asked. It works out what is changing, asks for anything missing and has Engine B run it. Then the chart updates.
The language model never does the arithmetic. It handles the conversation; Engine B is the calculation authority.
HAL works two ways: on your own, or with an adviser in the room. Where there is an adviser, the adviser owns any recommendation.
Any reply backed by a calculation can be opened in the Calculation Inspector, which shows the inputs, assumptions and controls behind the number.
Does the gap matter here?
Not every missing piece of information prevents a financial conclusion. Where the methodology permits alternative readings, Engine B can test whether the uncertainty affects the decision.
If the relevant outcomes remain unchanged across those readings, the calculation can proceed where the rules permit, with the uncertainty recorded. If the outcomes differ, further information is required before a supported conclusion can be reached.
Some information is fundamental to the calculation itself. Without an established expenditure requirement, for example, retirement affordability cannot be properly assessed.
What does the methodology permit?
- Sufficient Engine B calculates and returns a supported conclusion
- Missing, but immaterial Proceed where the methodology permits, recording the uncertainty
- Missing and material HOLD. HAL asks the next necessary question
- No strategy, or no justified grade HAL explains the limitation. No conclusion is manufactured
Try HAL
There are two ways to start:
- New client: build a plan by conversation. HAL asks one question at a time. The only structured step is the interactive attitude-to-risk questionnaire.
- Demo A, B or C: explore a prepared fictional household.
Then ask what you would ask an adviser. Am I on track? What if I retire later, spend less or save more? HAL answers with the chart and a sentence. Where a reply rests on a calculation, open the Calculation Inspector to see the workings.
HAL is a research demonstration, not financial advice. You will be asked to confirm this before you begin.
What is established, and what is not
Engine B is tested in two ways. For determinism, independent runs on the same inputs and engine version must reproduce identical outputs and trace hashes. For conformance, its conclusions are checked against expected results constructed independently of the engine.
HAL and Engine B are tested together, end to end, before release.
Not claimed
- That any output is financial advice or a suitability assessment.
- That the methodology has been independently reviewed.
- That independent implementations have converged.
- That the economic assumptions are calibrated against external benchmarks.