How TerraCube works, all the way down.

At the top level it's four steps. You give us an address; we read the best public science at that exact point; we turn it into things a homeowner actually cares about — a climate type, heating and cooling costs, hazard signals and an overall grade — and hand you a plain report you can trace all the way back to its sources. Nothing is invented in between.

Different questions are answered by different, purpose-built datasets. We never stretch one dataset to cover a question it wasn't built for. The full list, with links, lives on the Our data page; the ones that drive the science are:

• Today's climate normals — CHELSA v2.1. A peer-reviewed global climatology at roughly 1 km resolution, baseline 1981–2010. Source of annual mean temperature, seasonality and annual precipitation at your point.
• Climate type — Köppen-Geiger. The climate-class map (baseline 1991–2020) that names a location's climate and shows how that class is projected to shift.
• The future — NASA NEX-GDDP-CMIP6. Statistically downscaled CMIP6 projections for future 30-year windows (2041–2070 and 2071–2100) under low, medium and high emissions pathways. We use multiple models from this collection (see §4).
• Heating & cooling — building-stock surveys. US energy surveys (RECS / CBECS) and the EU's measured-calibrated TABULA typologies, combined with the climate above to estimate a home's heating and cooling demand and cost.
• This week's weather — NOAA & ECMWF. Public numerical weather models, refreshed each model cycle, for the 7-day forecast and near-term hazards. This is live forecast data; everything else above is a fixed, preprocessed climate baseline.

“Today’s climate” is always a 30-year average, because that is what separates climate from weather. A subtlety worth stating plainly: not every dataset uses the exact same 30 years, so a report can carry more than one “present” baseline at once. We label each number with the window it actually describes rather than forcing them to look identical:

A ~decade of offset between the 1981–2010 and 1991–2020 baselines is worth roughly a tenth to a few tenths of a degree of warming — small, but real, and we don't hide it by pretending the datasets are on the same clock.

A model is only as trustworthy as its track record. We compared TerraCube's present-day climatology against ERA5 reanalysis (via the free Open-Meteo archive) — an independent, observation-constrained dataset built by a different institution with a different method — at 30 globally distributed points, on the same 1981–2010 baseline so the comparison is fair. The full, re-runnable harness and per-point results are in the repository.

What those two words mean. Typical error (MAE) is how far off we are on an average place — smaller is better. Correlation is whether we get the pattern right (do warmer places come out warmer?) on a 0-to-1 scale, where 1 is perfect. The full table below adds bias (do we run consistently warm or cold?) and RMSE (which punishes big misses harder).

In plain terms: our temperature baseline sits within about half a degree of an independent reference and ranks places almost perfectly (correlation 0.998). Precipitation agrees far less tightly — that is expected and honest, because rainfall is harder to model and the reference's rainfall is itself model-generated, especially in monsoon and mountain regions. We report both so you can weight them accordingly. A few degrees of disagreement in steep terrain is largely the reference's coarser resolution, not our error; the harness documents each source of expected disagreement.

No single climate model is “the” answer — each makes slightly different assumptions, so they disagree, and that disagreement is real information. Rather than quietly pick one model and present its output as certainty, TerraCube runs an ensemble of climate models from the NEX-GDDP-CMIP6 collection at your point and reports two things for every future value:

• The central estimate — the median across the models, which is more robust than any one model alone.
• The spread — the range the models cover, shown as an uncertainty band. A wide band means the science is genuinely less settled for that place and variable; a narrow band means the models largely agree.

You also choose the emissions pathway (roughly: how much the world curbs greenhouse gases), because that choice, not model disagreement, is the single biggest lever on the later part of this century. The band describes model uncertainty within a pathway; switching pathways shows how much our collective choices still matter. Every projection in the report and API carries its central value and its spread — never a bare number pretending to a precision it doesn't have.

A home's heating and cooling demand is driven by how far, and how long, outdoor temperature sits away from a comfortable indoor target. The standard measure is degree-days: add up, over a year, how many degrees below a heating base (heating degree-days) or above a cooling base (cooling degree-days) the temperature runs.

We compute degree-days from local monthly-mean temperature (CHELSA-based for today; the downscaled projection for the future) using Thom's method, which accounts for the fact that daily temperatures swing above and below the monthly mean — a naïve monthly calculation badly undercounts near the base temperature. Two calibration choices matter and are made explicit rather than hidden:

• Regional base (balance-point) temperatures. The temperature at which a typical home needs neither heating nor cooling isn't a single global constant — it varies by region and differs for heating versus cooling. We use region-appropriate bases rather than one worldwide default.
• Day-to-day variability. Thom's method needs an estimate of how much daily temperatures spread around the monthly mean; we vary that with latitude and season rather than assuming one fixed value everywhere.

The resulting degree-days are combined with building-stock survey data and local energy prices to translate into an estimated heating/cooling cost. It is a well-grounded estimate for a typical home of its type and age — not a metered audit of your specific building.

Some risks are reported as proxies — indicators derived from climate variables rather than a full physical hazard model. Proxies are genuinely useful for “is this getting worse here, and roughly how much?” but they have limits we state up front:

• Heat stress is derived from temperature (and, where available, humidity) statistics — a strong indicator of trend and relative severity, but not a substitute for a local public-health heat-health warning.
• Wildfire and drought signals are built from climate indices (dryness, heat, seasonality). They capture whether conditions are trending more fire- or drought-prone; they do not model fuel, vegetation, ignition or local suppression, so they aren't a parcel-level fire rating.
• Flood is not a proxy we fake. True flood risk needs hydrology and elevation modeling we don't yet run, so we don't invent a flood score — consistent with the honest-gaps promise.

Every hazard signal on the site is labelled with what drives it, so you can tell a modeled quantity from a proxy indicator at a glance.

Each value in a report states, on the page, which dataset, baseline window and — for projections — which scenario and model ensemble produced it, along with the model spread. The same attribution is attached in the API response, so a number never travels without its source. If you can't trace a number, you shouldn't have to trust it.