School of Electrical and Computer Engineering, National Technical University of Athens
Machine Learning for time series forecasting: Lessons learned from the M4 Competition and ways forward
Machine Learning (ML) is widely used to solve a variety of complicated problems in signal processing, image recognition, and many other applications. More recently, it has been proposed as an alternative to traditional forecasting, introducing the means that could revolutionise the field. When lots of data are present, ML methods display many advantages over traditional ones. They can identify non-linear relationships, capture complex patterns and retain both short- and long-term dependencies. In this regard, ML has become very popular in applications that involve high-frequency data and numerous explanatory variables, such as energy and web-traffic forecasting. On the other hand, business forecasting is quite different in nature, especially when referring to low frequency, non-stationary and diverse data of limited sample. Training ML models under such circumstances becomes a challenging problem which requires innovative solutions to be effectively solved. The M4 Competition investigated such solutions by evaluating a total of 61 methods across 100,000 series. The winner was a hybrid method that mixed neural networks with exponential smoothing formulas, while the runner-up a cross-learning algorithm which optimally combined forecasts of various models. On the contrary, many ML methods failed to beat the simplistic benchmarks set by the organisers. This paper analyses the results of the M4 Competition to discuss the strengths and weaknesses of ML in time series forecasting, indicating when and why should we expect it to provide accurate results. It also presents best practices for developing accurate ML models and highlights the elements that can transform them into powerful forecasting tools. Some innovative ideas for exploiting the capacity of ML in forecasting in the future are finally discussed.
Evangelos Spiliotis is a Research Fellow at the Forecasting & Strategy Unit, School of Electrical and Computer Engineering (ECE), National Technical University of Athens (NTUA). He graduated from ECE at the NTUA in 2013 and got his PhD in 2017. His research interests are time series forecasting, decision support systems, optimisation, statistics, energy forecasting, energy efficiency and conservation. He has conducted research and development on tools for management support in many National and European projects. He was a co-organizer of the M4 Forecasting Competition.