Distribution-free prediction intervals for time series forecasting

Trustworthy machine learning (ML) systems should not only predict what they know, but also what they don’t know. However, in real-life applications, it often happens that ML systems are not able to quantify their uncertainty in a correct way. In this talk Prof. Waegeman will discuss various methods to quantify uncertainty in regression and time series forecasting settings. In the first part of the talk, he will present recent benchmarking results for methods that estimate prediction intervals in the classical regression setting with i.i.d. data. In the second part of the talk, he will elaborate on extensions of these methods in the time series forecasting setting. Compared to the i.i.d. setting, estimation of prediction intervals in time series forecasting can be very challenging, especially in the presence of autocorrelation and non-stationarity.