PP: Shape and time distortion loss for training deep time series forecasting models

Problem: time series forecasting

Challenge: forecasting for non-stationary signals and multiple future steps prediction

?? how to deal with non-stationary datasets??

Introduction

one-step prediction problem VS multi-step prediction;

multi-step forecasting requires to accurately describe time series evolution. 

limitation of the euclidean loss(MSE): in non-stationary context;