Estimated summer comfort, with the unknowns visible.

Looking for the practical version? Understand the summer comfort check · Sommerkomfort auf Deutsch · Check an address →

Beta model summer-peak-auto-beta-4 · 6 September 2026. The index concerns summer overheating only: sustained hot spells, short heat peaks and hot nights. Higher scores mean less modeled overheating. It is not an indoor measurement, weather forecast, energy certificate or guarantee of comfort.

Automatic, from the building onward

Once an address resolves to a residential building, public data supplies the inputs. No floor, bedroom, blinds or ventilation questionnaire is used. Previously saved apartment answers do not change the score. The building register’s coordinate anchors the climate sample, so nearby address pins for the same building use the same location. This is a building estimate; we cannot identify the actual room or floor of every apartment.

The citywide baseline tests all eight compass directions with equal influence. Exact windows and façade shading remain unknown. One-storey buildings test direct roof exposure; other buildings test both roof-exposed and non-roof-exposed apartments. Unknown windows, blinds, insulation, thermal mass and window opening remain explicit scenarios. A missing construction year uses the older-stock envelope set, including upgraded and poorly insulated alternatives.

Three simple comfort levels

The property view compares a building with the fixed reference population of residential buildings in the City of Zürich. It shows Less prone to overheating, Typical for Zürich, or More prone to overheating. These are relative comparisons, not indoor comfort guarantees.

Even a relatively better building can overheat during a hot spell. Low confidence and the possible range remain visible. The range reflects tested assumptions, not a calibrated confidence interval.

Presentation version summer-overheating-relative-3-v1 uses the unrounded internal calculation. The lower and upper third boundaries of the initial city reference population define the three levels. Every building has equal weight; tied calculations remain together. These boundaries stay fixed during routine data refreshes. Changes to the model require a new, explicitly versioned reference population.

The original absolute bands put all 34,149 eligible buildings in “Less comfortable”, despite different modeled burdens. The relative scale makes those differences visible without changing the heat calculations or claiming a measured improvement in comfort. Old shared snapshots retain their original scale.

The underlying indication and uncertainty

The score is an early indication to support apartment decisions. Every comfort level is accompanied by low confidence; the brief and comparison also show the possible range of levels. Broad ranges and extreme scenarios no longer hide the score. Both integration examples show Less comfortable, with a possible range from Less comfortable to Mixed comfort. Their underlying indices are 10 and 5, but that small numerical difference is not presented as a meaningful advantage.

The interval is a scenario range: the minimum and maximum scores across the tested building and operation assumptions. It is not a calibrated 90% or 95% confidence or prediction interval, and it does not quantify all structural model error. We do not know the probability that a real apartment’s score lies inside it. A wider range indicates more sensitivity to unknown apartment details; a narrow range alone does not establish accuracy.

The inherited simplified thermal model produces extreme upper temperatures in some scenarios. We retain those uncapped calculations internally, flag the score as especially uncertain and avoid publishing them as room-temperature predictions. The 40-point range and 50°C scenario thresholds are internal confidence flags, not rejection gates or health limits. The visible range does not make those extreme scenarios validated. No matched residential building or usable climate coverage still means no score; source failures can be retried without entering apartment details.

Sources and what they measure

Heat balance and assumptions

The inherited one-resistance, one-capacitance model balances envelope and ventilation heat transfer, solar/internal gains and effective thermal storage per square metre of floor area. Four quarter-hour exponential updates are made per hourly weather sample. Its single temperature state is an air-like screening approximation: it does not separately resolve air, radiant surface temperature or humidity. It is not a validated operative-temperature comfort model.

The daytime and nighttime spatial anomalies relative to Zürich/Fluntern are interpolated across the day and transferred onto the 2023 weather sequence. Each anomaly is capped at ±6 °C and its strength varies from 0.5 to 1. This transfer is an assumption, not an observed correction. Sun geometry and measured horizontal radiation supply approximate vertical gains with a fixed diffuse fraction of 0.4 and direct-normal cap of 950 W/m². Surrounding shade has assumed transmission of 0.65 or 1. Trees, full 3D shadows, balconies and courtyards are not explicitly simulated.

All following parameters are uncalibrated model scenarios, not measured building attributes. Wall U-values are 0.3, 0.8 or 1.4 W/(m²K) for older/unknown stock, and 0.25, 0.45 or 0.8 for construction from 2000 onward. Wall area/floor area is assumed 0.6. Window U-value is 1.2 for wall U below 0.5, otherwise 2.6. Effective storage is 35 or 65 Wh/(m²K); glazing/floor-area ratio is 0.15 or 0.25; solar transmittance is 0.55. Blinds have transmission 0.2 or 1. Air changes are 0.35 in the background, or 4 when the usable-purge scenario can open windows during 19:00–08:00 and outside air is cooler.

Roof-exposed scenarios test roof U-value 0.25 or max(0.4, wall U), with both heat transfer and a simplified solar term of GHI × roof U × 0.025. Roof area/floor area is implicitly one. Internal gains are 2 W/m² at night and 4 otherwise. Room height is assumed 2.7 m. These simplifications and unconstrained parameter combinations contribute to the extreme outcomes identified by the diagnostic confidence flags. A more complete thermal model and reference-simulation benchmarks are required; public-data availability alone does not fix model accuracy.

Peak score calculation

For every contiguous 72-hour post-warmup window, calculate positive temperature excess above 26 °C. Daytime is 07:00–22:00 CEST (45 hours per window); nighttime is 22:00–07:00 (27 hours). Let D and N be the respective mean excesses. Window burden B = 0.4D + 0.6N. Each scenario takes its maximum B across the 49 eligible windows. Cooler days elsewhere cannot dilute that peak.

Within each floor/direction group, take the median scenario burden. Take the median of those group medians, then calculate 100 × exp(−B/3). The relative comparison uses this continuous value; the technical absolute index is rounded to five points. This prevents different numbers of roof scenarios from weighting floors accidentally. Groups have equal influence, not population-frequency weights. The range is the minimum and maximum individual scenario score, rounded outward to five points. It is not a confidence interval, nor a guarantee that all real apartments lie inside it. A score of 100 would mean no modeled excess in this test, not guaranteed comfort.

The model also records the highest hourly temperature, the hottest complete nine-hour night and the longest consecutive run above 26 °C. Incomplete nights at the bundle edges are excluded; duration can be censored at the end of the test. Raw temperature diagnostics remain internal while unvalidated. Any tested hourly peak ≥30 °C or complete-night mean ≥27 °C is retained as a heat flag. It does not cap a relative category: the absolute heat warning is shown separately. These flags, the 26 °C reference, 72-hour window, weights and exponential scale are beta design choices, not SIA, Minergie, CIBSE or medical compliance criteria.

Accuracy and the next validation work

Temperature error and ranking accuracy are unknown. No ±°C, percentage-accuracy or probability-of-comfort claim is supported. Small score differences are inconclusive when scenario ranges overlap. Passing software tests establishes calculation consistency, not actual indoor comfort.

Next work: audit public-data coverage across a stratified building sample; calibrate outdoor boundary conditions against Zürich’s public urban sensors using held-out stations and heat events; benchmark a richer thermal model against reference simulations; and validate against consented bedroom/living-room observations. Pilot measurements train and test the shared model, and are not required from anyone checking a property. Geometry/shading improvements and future-summer stress tests remain planned. Measurements will improve the model and help calibrate future uncertainty intervals; perfect accuracy is not a prerequisite for this clearly labeled directional beta.

Only public location and building identity go to the heat API. Shared comparisons freeze the displayed score, scenario range, low-confidence label and model version. Previously saved personal heat answers remain stored locally for compatibility and are ignored by the automatic calculation.