Model Uncertainty, Explained

Model uncertainty is the gap between a model's output and the truth it tries to represent. It comes from several sources at once: inputs that are estimated rather than measured, such as wall construction or system efficiency; methods that simplify reality, such as standard occupancy patterns or steady-state heat flow; weather that differs from the standard year; and future behaviour that cannot be known, such as how occupants will heat a home after retrofit. A model result without an uncertainty statement implies more precision than modelling can deliver, and that false precision is itself a problem.

Uncertainty is not a flaw to be eliminated; it is a property to be managed. Some uncertainty can be reduced by better data, such as observing a wall or servicing a boiler, while some cannot, such as future weather or future occupancy. The practical question is therefore not how to make a model certain, but how to decide well despite uncertainty: which decisions are robust across the plausible range, which depend on a single uncertain input, and which should wait for better evidence.

Input uncertainty usually dominates for homes. A wall U-value taken from a table, a ventilation rate assumed rather than measured, or a heating efficiency estimated from age can each move a result substantially, and their combined effect can be larger than any single one. Method uncertainty adds a further layer: two approved methods can give different results for the same home, because they standardise differently, and neither is simply wrong. Behavioural uncertainty adds a third layer, because real heating patterns, hot water use and window opening differ from standard patterns in ways that vary by household and season.

Presenting uncertainty honestly means showing ranges, not just point figures. A result presented as a range, with the inputs that drive the range named, is more useful than a single figure with no indication of spread, because it lets the reader see whether the decision is sensitive to uncertainty. A measure that looks worthwhile across the whole range is a robust choice; a measure that looks worthwhile only at one end of the range is a gamble on that end being right.

To handle uncertainty, start by listing uncertain inputs and ranking them by influence, using sensitivity checks. Then ask, for each decision, whether it holds across the plausible range of the influential inputs. If it does, proceed, noting the range. If it does not, obtain better data for the input that decides the outcome, or choose a more robust option, or stage the work so that later stages can adapt to what is learned. Those three responses, better data, robust choice, staged work, cover most uncertainty problems.

Also distinguish uncertainty that matters from uncertainty that does not. A model may be uncertain about absolute annual demand while being quite certain about the ranking of options, because the same uncertain input affects all options similarly. In that case, the ranking can be trusted even where the absolute figure cannot, and decisions based on ranking are sound. Conversely, a model may be precise about a minor input while being uncertain about a major one, and polishing the minor input does not help.

By place, uncertainty handling is similar in principle across UK and EU systems, because it is a property of modelling rather than of one method. SAP-based advice and national-method advice both benefit from ranges and sensitivity notes, even though formal requirements for reporting uncertainty differ. Where a certificate is concerned, the official result remains a single standardised figure, and uncertainty information sits alongside it as advice.

For high-value or complex decisions, formal uncertainty analysis, varying many inputs together, may be worthwhile, while for typical homes, simple one-at-a-time sensitivity checks on the main uncertain inputs usually reveal the decision-relevant uncertainty. The method should match the stakes, not exceed them.

A common mistake is to demand a precise answer from a model that cannot give one, and then to choose on false precision. Another mistake is the opposite: to treat all uncertainty as equal, and to paralyse a decision that is robust across the plausible range. Both mistakes confuse uncertainty about a number with uncertainty about a decision.

A further mistake is to hide uncertainty by reporting only the most favourable end of a range. That practice turns uncertainty analysis into marketing, and it destroys the trust that honest ranges build. Report the full range, with its drivers, even where the range is wide.

Keep the uncertainty note, the sensitivity table and the decision reasoning together with the model. That file shows a later reader not only what was decided, but how uncertainty was handled in deciding it.

Questions to ask include: what is uncertain here, does the decision hold across the plausible range, and what evidence would narrow the uncertainty that matters. Those questions make uncertainty a tool rather than a worry.