Bitte verwenden Sie diesen Link, um diese Publikation zu zitieren, oder auf sie als Internetquelle zu verweisen: https://hdl.handle.net/10419/173659 
Erscheinungsjahr: 
2018
Schriftenreihe/Nr.: 
FAU Discussion Papers in Economics No. 02/2018
Verlag: 
Friedrich-Alexander-Universität Erlangen-Nürnberg, Institute for Economics, Nürnberg
Zusammenfassung: 
We present a comprehensive simulation study to assess and compare the performance of popular machine learning algorithms for time series prediction tasks. Specifically, we consider the following algorithms: multilayer perceptron (MLP), logistic regression, naïve Bayes, k-nearest neighbors, decision trees, random forests, and gradient-boosting trees. These models are applied to time series from eight data generating processes (DGPs) - reflecting different linear and nonlinear dependencies (base case). Additional complexity is introduced by adding discontinuities and varying degrees of noise. Our findings reveal that advanced machine learning models are capable of approximating the optimal forecast very closely in the base case, with nonlinear models in the lead across all DGPs - particularly the MLP. By contrast, logistic regression is remarkably robust in the presence of noise, thus yielding the most favorable accuracy metrics on raw data, prior to preprocessing. When introducing adequate preprocessing techniques, such as first differencing and local outlier factor, the picture is reversed, and the MLP as well as other nonlinear techniques once again become the modeling techniques of choice.
Dokumentart: 
Working Paper

Datei(en):
Datei
Größe
6.23 MB





Publikationen in EconStor sind urheberrechtlich geschützt.