An Operational, Purely Data-Driven Seasonal Forecast System

We will share high-level analysis from a purely data-driven seasonal forecast system. The forecast system is fully operational and has higher deterministic and probabilistic skill than an ensemble of physics-based seasonal forecast systems, including the ECMWF SEAS5 system. The data-driven model also requires five orders of magnitude less compute than a physics-based system and can be run on a single workstation.

We have found that limiting non-physical outputs relies on a careful curation of training data, an appropriate loss function, and a physics-informed neural network architecture. Each of these aspects is necessary but insufficient on its own to produce a reliable, stable seasonal forecast system. The design process that led to the most success was informed by physical and scientific reasoning through all stages and relied only on advanced ML techniques when more basic methods were insufficient. For this reason, we have found that tightly coupling scientific and ML engineering expertise is necessary for successfully developing fully data-driven forecast systems.

This forecast system is being used operationally in private markets and has undergone extensive testing and validation, both internal to our company and by external parties. The scientific methodology has also been reviewed and approved by the National Science Foundation on two separate occasions.