Probabilistic and Summary Forecasting, and some Pitfalls in Forecasting Practice

25 June 2023

Talk given at

  • International Symposium on Forecasting (ISF) 2023, 24-28/6/2023 Charlottesville, Virginia, USA
  • The second part of the talk was also given at DABI Summit 2022 (a Meta internal event), Menlo Park, CA, USA, December 2022

The talk was given with Slawek Smyl. Slawek covered in the first part his novel ideas about probabilistic and summary forecasting, I covered in the second part common pitfalls in forecasting practice.


The talk will cover some pitfalls of forecasting in practice, and then go into details of applications for probabilistic and the related and less usual summary forecasting. Probabilistic forecasting means providing a full distribution. It can be done in a “statistical way” - assuming and fitting parameters of a distribution, or in the “ML way” by outputting a required number of quantiles. The latter can be embodied as a system that learns to output a number of a priori agreed quantiles or a dynamic one, in which during training, it learns to output any quantile and delivers requested quantiles during serving, but the list of needed quantiles is supplied only during serving. The setup for summary forecasting is to have highfrequency data, but make decisions in low-frequency, and long forecasting horizons, e.g., having hourly or daily data and forecasting a year or two ahead. Difficult long-term forecasting in high-frequency is usually not needed, as customers would aggregate it into some lowfrequency periods like months or quarters anyway. So instead, in summary forecasting, we directly forecast the periodic summaries.


 Download pdf

« Interpretability and Causal Inference for Global Time Series Forecasting Methods | Short open problem talk: Hierarchical summary forecasting »