Noēsis / Approach

How a number earns its place in a board pack

Every figure in Noēsis passes through four recorded review states, every model shows its workings, and every decision is made and signed by a person.

Four review states

A figure in a board pack carries an implied claim: that someone checked it, and that the committee approved it. In most institutions that claim cannot be tested after the event. Noēsis makes it explicit. Every figure set, model and report carries one of four states, and the state is shown wherever the item appears.

○ Draft

Prepared, not yet checked.

◉ Verified

Checked against source by an independent reading.

◐ Reviewed

Reviewed by an analyst or the risk function.

● Approved

Approved by the investment committee, ALCO or board.

Draft. The item has been prepared, from an extract, a source document or an analyst's model, and has not been checked. Drafts can be worked on freely. They cannot be included in a committee or board pack.

Verified. The figures have been checked against source by an independent reading. For company financials, that means a first extraction and a separate blind second reading of the source page, with every difference resolved by a person. For liquidity data, it means the extract has been reconciled to the general ledger within the tolerance your finance team sets, with every break cleared or explained.

Reviewed. An analyst or the risk function has reviewed the item: the assumptions, the method and the result, not only the arithmetic. The reviewer is named and the review is dated.

Approved. The investment committee, ALCO or the board has approved the item. The approval records the committee, the meeting and the decision.

Three rules hold throughout. An item cannot skip a state. The person who prepared an item cannot verify or approve it. And any change to an approved item creates a new version that starts again at Draft, while the approved version is kept unchanged.

Models estimate; people decide

Noēsis uses two kinds of model, and we are precise about which is which.

Stated models. Maturity ladders, stress scenarios, early-warning indicators and valuation models are stated models. Their logic and assumptions are written down and shown to the user. Behavioural assumptions, such as the renewal rate on maturing deposits or the prepayment rate on a lease book, are set and approved by your institution. Stated models do not use machine learning.

Machine learning. We use machine learning in two places today: extracting figures from source documents, and making the independent second reading that checks them. In both, it does clerical work that people then check. It does not produce early-warning signals, model values, ratings, limits or recommendations.

This follows a principle familiar from the BCBS Principles for Sound Liquidity Risk Management and Supervision and from supervisory expectations on model risk more widely: the assumptions behind a risk measure should be documented, reviewed and understood by the people who rely on it.

Every model is versioned

Each model and each assumption set carries a version number, a date and an approver. A report records the versions it used. If ALCO changes the deposit run-off assumption in March, a stress test from January can still be reproduced exactly as it was presented, and the change is visible in the history.

Ranges, not points

A model value is an estimate. We show it as a range, together with the assumptions that move it most and what would have to change for the view to change. A survival horizon under stress, a funding gap, a model value for a company: each is presented with its range. A single number suggests a precision that the inputs do not support.

Back-testing before any accuracy claim

We make no claim about the accuracy of any model until it has been back-tested. For liquidity, that means comparing assumed outflows and renewals with what happened, on your data, over a period long enough to mean something. For data extraction, it means measuring the error rate found by people in the sample audit. Where results exist, we will publish the method alongside them.

Human sign-off

Ratings, target prices, limits, scenarios and approvals are set by your committees, never by a model. Noēsis records each decision: who made it, in which forum, on what date, and on which version of the figures. The software provides the evidence and keeps the record. The judgement stays with the people accountable for it.

Talk to us about a pilot

We are selecting a small number of institutions for three-month pilots, with no fee for the pilot.

Request a briefing