A Machine Learning Approach for Forecasting with Limited Data and for Distant Time Horizons

Abstract

Time series forecasting has attracted the attention of the machine learning (ML) community to produce accurate forecasting models that address the limitations of classical methods. A large part of ML research focuses on innovative algorithms, but another important area is transitioning ML to industry settings. The objective of this Thesis is to apply ML in realistic scenarios by devising methods that make practical, usable forecasts and models.

We focus on three areas that contribute to more practical forecasts. First, we improve the problem formulation of multi-step ahead forecasting by including a notion of an offset to create a more customizable forecasting window. A comparative analysis across three datasets shows that at further out horizons, two models that include the notion of an offset consistently outperform their original counterparts. Secondly, we simulate a scenario where an ML model does not have access to the immense amount of training data normally necessary to train deep neural networks. We use transfer learning with a weight sharing algorithm and observe that it improves all seven target model accuracies even after they have accumulated two weeks of their own data. However, the input data to transfer knowledge from must be chosen wisely to avoid negative transfer. Finally, we address the challenge of deploying a safe and robust ML model by outlining key features in a time series forecasting library and compare 13 notable, actively maintained libraries, finding a wide variability in the features included in the libraries. A surprisingly low number of libraries include a benchmarking system, and around half of the libraries provide a pre/post-processing engine that would allow for modular ML models in a pipelining manner for easy deployment.