With 700+ participants and 50+ speakers from around the world, we are thrilled to say that this year's online event was a success! Better still, all sessions are now available to watch on YouTube.
This panel session promises to offer an interesting look at what the best forecasting methods are. The panellists will draw on the history of forecasting competitions to compare, contrast and combine machine learning and statistical methods in forecasting.
This stream will be a valuable addition to anyone's OR62 timetable, particularly those who work with machine learning methods, statistical analyses or otherwise use forecasting.
The quest for greater forecasting accuracy: Perspectives from Statistics and Machine Learning
The NN3 competition, which took place between 2007 and 2008, brought together computational intelligence and statistical methods in one of the first robust comparisons of statistical and machine learning approaches to forecasting in a competition setting.
The larger, more renowned M3 competition had only a single neural network contestant, which performed rather poorly. Since then, the M4 competition expanded on its predecessor to incorporate all major forecasting methods, including a larger number of submissions from the machine learning and computational intelligence domain and hybrid approaches.
While the results of the much smaller NN3 competition were encouraging and suggested some potential for machine learning methods, the much larger and more recent M4 competition was particularly striking in its observation of the poor performance of ‘pure’ machine learning methods.
However, greater forecasting accuracy could be achieved by more complex methods, and it was a hybrid approach using both statistical and machine learning properties which stole the show. It was noteworthy that all top performers of the competition used a combination of forecasts. However, many questions remain as to the future of machine learning versus statistics in forecasting and the role of forecast combinations in this quest for greater forecast accuracy.
In this panel discussion, we discuss the past, present and future of machine learning and statistics, beginning with the lessons learnt from over a decade of forecasting competitions. We will also discuss issues in forecast selection and forecast combination, which continues to rank well across all competitions and is a common approach in practice.
This panel session promises to be very interactive, it features time for each panellist to speak, time for moderator questions and time for audience questions.
Iqast - Intelligent Forecasting Software
University of Skovde
University of Birmingham