Tuesday, 14 January 2020: 2:15 PM
In recent years, various segments of the Weather Enterprise have worked to expand the use of weather forecasts from single-value, deterministic predictions to probabilistic representations that expose not only the expected value but a range of reasonably-likely outcomes. At The Weather Company, an IBM Business, we view probabilistic forecasts as a way to convey weather forecast information in a richer, more complete way, enabling consumers and businesses to make better decisions. Snowfall accumulation is perhaps the forecast parameter best suited to probabilistic representations since consumers have to-date embraced deterministic snowfall forecasts that are expressed as an expected range. In this talk, we will discuss a new probabilistic snowfall capability that was introduced to consumers in Feb. 2019 via the iOS version of The Weather Channel mobile application. A number of science and engineering challenges were overcome to enable this global, on-demand solution including how to supply probabilistic information at a global scale, ensuring that the deterministic and probabilistic snowfall forecasts are consistent, and that the probabilistic snowfall forecasts are better-calibrated and more skillful than snowfall forecast distributions derived directly from the supporting multi-model ensemble model outputs. Forecasts are calibrated using a combination of logistic regression with a point-mass at 0, and construction of the continuous portion of the snowfall density using a gamma distribution via Bayesian Model Averaging.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner