The history of weather forecast models, up until very recently, has been dominated by physical models i.e. models that use physical equations such as thermodynamics, kinetics etc. A new genre of weather model has recently emerged, harnessing the rapidly expanding artificial intelligence (AI) technology, in particular focusing in on machine learning (ML). By training models on past weather data it is possible to run predictions/forecasts, just like the typically used physical models, but at a fraction of the computational power.
MetDesk are using the framework of the FourCastNet ML weather model to bring this new technology to Trading Weather.
Trained on ERA5 re-analysis data, generally recognised to be a best estimate of past weather conditions globally, the ML model produces forecasts (given a physically modelled starting point at t=0) to sit alongside models such as the ECMWF and GFS.
Beyond a typical deterministic run, MetDesk have created an AI ensemble, with 51 members, using specifically chosen perturbations for the initial conditions to provide an idea of confidence/ spread in potential solutions.
Key Info:
MDai50 (Sub-Seasonal Range)