Brian Head
10/09/2023, 3:48 PMCristian (Nixtla)
10/10/2023, 5:25 PMMQLoss
to output multiple quantiles, or with the DistributionLoss
class that assumes some underlying distribution. Both type of losses are viable options that will produce calibrated PI in most cases. This is the standard approach for deep-learning methods (for eg. refer to the DeepAR paper). Here are some tutorials: https://nixtla.github.io/neuralforecast/examples/uncertaintyintervals.html https://nixtla.github.io/neuralforecast/examples/longhorizon_probabilistic.html