The Weather Company's paradigm for producing forecast content is different from more traditional approaches in several significant ways. In traditional approaches, forecast guidance such as models, model output statistics and similar content are pre-generated and provided to human forecasters for review and improvement. Forecasters apply experience and wisdom based on current data, adjust the forecasts, and then periodically publish new forecast products. The process is characterized with typical time-scales of several hours, and with spatial fidelity and temporal frequency limited by that which humans can reasonably process, perhaps thousands of locations at several hour increments. In contrast, models today are routinely producing output at millions of locations, with output increments of tens of minutes, update frequencies of hourly in some cases, and from a plethora of model sources. Hence, traditional approaches are often significantly under sampling and underutilizing the input guidance data.
The foundation of the Weather Company's forecasting philosophy is that there is tremendous value in the full richness of the input forecast guidance and that a modern forecasting system should be able to retain as much of that richness in output forecast products. Therefore, a new paradigm has been adopted in which man and machine work in parallel (rather than serially) allowing the complete ensemble of input guidance data to be exploited at its native resolution and frequency, while retaining the wisdom of the human in guiding the end results. In this forecasting approach, referred to as Forecasts On Demand, the production has the following characteristics:
1. Forecasts are never pre-produced and set aside for consumption at a predetermined set of locations, but all forecast content is generated on the fly using intelligent integrations of the latest input guidance from models, nowcasting methods and humans. Forecasts are always produced precisely for the location required and using the latest input data at its native resolutions. 2. The foundation of the 1-15 day forecast content is a plethora of global and regional models, combined together using dynamically-adapting, machine-learning methods that continuously update as new models are available and learn as new verification arrives. 3. The foundation of the first few hours of the forecast is a combination of extrapolation nowcasting methods and rapid-update NWP, which is also continuously updated as new dependent input data becomes available. 4. Human influence on the forecasts is by an Over the Loop paradigm in which the forecasters continuously update guidance instructions which are applied when the forecast is generated at the time of client request. 5. Short-term warnings of life and property threatening weather from official sources are incorporated into the forecasts to ensure consistent public messaging in such situations.
An overall general description of the forecast system will be provided in this presentation. The companion papers mentioned above will provide details of these methods and results.
The Weather Company has built and deployed an operational Forecasts on Demand system that embodies these science principals and methods. The system is run using cloud-based computing and serves billions of requests per day worldwide. It has been in operation since 2013 and a variety of statistics on its performance will be shown. This includes comparative forecast accuracy statistics from independent third parties such as ForecastWatch which show that the forecast content is or is amongst the most accurate available.