Approach

---

There are two ways to misread a model: treat it as a complete solution, or dismiss it as a simplification. In both cases, value is lost.

A model is a way of seeing a system, not solely a way of measuring it. What it includes and what it omits are choices, and those choices shape the answer before any data is incorporated. 

Purpose matters too: not every model built to explain a system will predict it, and those that are optimised for prediction can fail badly outside the conditions to which they were calibrated.

We build our models so that these choices are visible and can be challenged. You should be able to interrogate any model you are asked to rely on, including ours.  Some of our models are open source, others are client-specific. Where a model is licensed, we supply the documentation, assumptions and diagnostics needed to validate it independently.

Models frame uncertainty; they do not remove it. We consider four forms:

Structural — what the model leaves out, and the choices its data collectors and developers made. We use ensemble approaches to expose this, because more data doesn't always help.

Parametric — what is unknown about the model’s inputs. We quantify this to show whether another survey or data stream would actually change the decision.

Stochastic — variability inherent in the process. We show the implications of this, pricing it rather than giving the impression it can be removed.

Scenario — which futures are considered plausible, and which courses of action are open. We evaluate what each intervention would achieve and what it would cost, even under circumstances you cannot control. 

Models differ greatly in the extent to which they have been confronted with evidence. Some are parameterised from literature values or expert judgement and never tested against observation.

We use state-of-the-art Bayesian and machine-learning methods to calibrate against the historical record, clarifying which parameters were fitted, which were assumed, and on what basis.

We actively look for where our models fail. Where the setting allows, we submit forecasts before the event so they can be scored against what really happens. For rare events, we test outside the conditions the model was calibrated on, even against situations never yet recorded. 

We also ask what didn't happen, and why. Might it happen next time?

A model incorporates a perspective on a system, drawing on what was known when it was built. As time passes, new observations arrive, exposure changes hands, one season behaves unlike anything in the record, and the science moves on.  

We treat modelling, uncertainty quantification and decision-making as a cycle rather than three steps that finish. Each pass sharpens what is known and reopens what has changed.

When licensing our models, we include ingestion or new data, recalibration, validation and support for the life of the licence. Your team accesses the figures and the visuals they need. Our clients choose a model that stays up to date — not permission to use one that doesn’t.