Inside the signal

Every weather station.
Every hour.

Thagorus ingests data from 14,000+ NOAA and ECMWF weather stations across North America. Temperature, humidity, precipitation, wind, UV index, pressure gradients, and derived biometeorology indices — all normalized to a consistent spatiotemporal grid.

Weather is not
a single number

Temperature alone explains almost nothing. Thagorus constructs compound features: the interaction of humidity and temperature that drives “feels-like” discomfort. The UV trajectory over a 5-day window that predicts sunscreen demand. The wind chill delta from last week that triggers coat purchases.

These features are generated per-category based on domain-specific meteorological research.

Separating weather
from everything else

The hard problem: did sales rise because of weather, or because of a promotion, a holiday, a competitor’s stockout, or seasonality? Thagorus uses structural causal models with instrumental variable estimation to isolate the weather-attributable component.

Weather is exogenous — it can’t be caused by your marketing. That makes it a natural instrument for causal identification.

demand = f(weather) + g(promos)
        + h(season) + ε

Thagorus isolates f(weather)
from the confounders.

Every market has
its own weather story

A cold snap in Phoenix is not the same as a cold snap in Chicago. Consumer response to identical atmospheric conditions varies dramatically by geography, acclimation, and local culture.

Thagorus fits separate response curves for each DMA, using hierarchical Bayesian estimation to borrow strength across similar markets while respecting local variation.

Forward-looking,
not rear-view

Historical weather data tells you what happened. Thagorus integrates ECMWF ensemble forecasts to tell you what’s about to happen. The 10-day atmospheric trajectory is mapped through your market’s specific response curves to produce a demand forecast that incorporates weather effects already baked in.

Ensemble spread becomes forecast uncertainty — wider ensemble disagreement means wider confidence intervals on demand.

One endpoint.
Everything you need.

A single REST call returns the weather-adjusted demand forecast, confidence intervals, recommended budget moves, backtest results, and break conditions for any market and category combination.

GET /v1/forecast
  ?market=DMA-501
  &category=sunscreen
  &horizon=7d

→ demand_lift: +12.4%
→ confidence: [8.1%, 16.7%]
→ recommendation: +$450 Meta
→ backtest_accuracy: 0.89

Every recommendation
shows its work

Attached to every API response: a backtest window showing how this recommendation would have performed historically, uncertainty bounds from the ensemble, and explicit break conditions — the scenarios under which you should ignore us.

If we’re not confident, we say so. If conditions change faster than our model can track, we flag it. The proof bundle is the contract between our model and your decision.