Meta-learning for Forecasting Model Selection

Barak, Sasan and Crone, Sven and Kourentzes, Nikos (2021) Meta-learning for Forecasting Model Selection. PhD thesis, Lancaster University.

Text (2021BarakPhD)
2021BarakPhD.pdf - Published Version
Available under License Creative Commons Attribution-NonCommercial-NoDerivs.
Download (2MB)

Abstract

Model selection for time series forecasting is a challenging task for practitioners and academia. There are multiple approaches to address this, ranging from time series analysis using a series of statistical tests, to information criteria or empirical approaches that rely on cross-validated errors. In recent forecasting competitions, meta-learning obtained promising results establishing its place as a model selection alternative. Meta-learning constructs meta-features for each time series and trains a classifier on these to choose the most appropriate forecasting method. In the first part, this thesis studies the main components of meta-learning and analyses the effect of alternative meta-features, meta-learners, and base forecasters in the final model selection results. We investigate different meta-learners, the use of simple or complex base forecasts, and a large and diverse set of meta-features. Our findings show that stationarity tests, which identify the presence of unit root in time series, and proxies of autoregressive information, which show the strength of serial correlation in a series, have the highest importance for the performance of meta-learning. On the contrary, features related to time series quantiles and other descriptive statistics such as the mean, and the variance exhibit the lowest importance. Furthermore, we observe that using simple base forecasters is more sensitive to the number of groups of features employed as meta-feature and overall had worse performed. In terms of the choice of learners, classifiers with evidence of good performance in the literature resulted in the most accurate meta-learners. The success of meta-learning largely depends on its building components. The selection and generation of the appropriate meta-features remains a major challenge in meta-learning. In the second part, we propose using Convolutional Neural Networks (CNN) to overcome this. CNN have demonstrated breakthrough accuracy in pattern recognition tasks and can generate features as needed internally, within its layers, without intervention from the modeller. Using CNN, we provide empirical evidence of the efficacy of the approach, against widely accepted forecast selection methods and discuss the advantages and limitations of the proposed approach. Finally, we provide additional evidence that using meta-learning, for automated model selection, outperformed all of the individual benchmark forecasts.