Salient's proprietary models are run on 3 separate timescales (weekly, monthly, and quarterly) targeting windows of predictability on S2S timescales from 2-52 weeks. See Forecast Models for additional information.
Salient models utilize ML/AI rather than physical equations to forecast weather, an approach that is optimized for sub-seasonal to seasonal (S2S) timescales—2 to 52 weeks. Physics-based weather models are effective for short-range forecasts but are limited on longer timescales by the computational demands of solving a large chaotic system. Deep learning methods can identify climate patterns that provide predictability without the constraints of physics-based solutions. Salient models are trained on a wide range of climate data spanning a 70-year historical record, augmented by additional data sources. Our models ingest novel data sources focused on the ocean and land surface, which provide the inertia of the climate system on seasonal timescales.
Traditional deterministic weather forecasts provide a single discrete forecast value for each time and location out to about two weeks with reasonable accuracy. Due to the chaos of the atmosphere, it is unrealistic to predict a single value for longer timescales.
Probabilistic forecasts provide a set of probabilities associated with all possible future outcomes, which is essential whenever uncertainty is irreducible, as is the case with weather forecasting***.*** This approach gives users a more complete understanding of future events' uncertainty, allowing for more informed decision-making.
Dynamical models of the atmosphere, also known as numerical weather prediction (NWP) models, are extremely complex and use supercomputers to solve the mathematical equations governing the physics and motion of the atmosphere. Traditional (NWP) models such as GEFS and ECMWF quantify uncertainty by generating ensemble forecasts. Ensemble forecasts generate multiple forecasts of possible scenarios by perturbing the initial conditions of the respective physical models to create a range of outcomes. Despite providing ensemble forecasts, NWP models tend to be biased and overconfident - similar to someone telling you that the chance of getting heads on a coin flip was 70%. Salient's AI models tackle these limitations by providing natively probabilistic forecasts that are calibrated by design in order to provide more reliable forecasts.
Calibration provides reliability. Â A reliable probabilistic forecast is one in which the forecasted event is observed at the frequency indicated by the probability of the forecast. In other words, when averaging across every forecast that predicts colder than normal conditions three weeks from now with a 60 percent probability, a colder-than-normal week should occur 60 percent of the time.
Salient applies a rigorous calibration methodology to physical models (GEFS and ECMWF) to improve the reliability of these forecasts, enhancing their practical applicability and usability.