A machine learning dynamic switching approach to forecasting when there are structural breaks

Jeronymo Marcondes Pinto and Jennifer Castle

Economics Series Working Papers from University of Oxford, Department of Economics

Abstract: Forecasting economic indicators is an important task for analysts. However, many indicators suffer from structural breaks leading to forecast failure. Methods that are robust following a structural break have been proposed in the literature but they come at a cost: an increase in forecast error variance. We propose a method to select between a set of robust and non-robust forecasting models. Our method uses time-series clustering to identify possible structural breaks in a time series, and then switches between forecasting models depending on the series dynamics. We perform a rigorous empirical evaluation with 400 simulated series with an artificial structural break and with real data economic series: Industrial Production and Consumer Prices for all Western European countries available from the OECD database. Our results show that the proposed method statistically outperforms benchmarks in forecast accuracy for most case scenarios, particularly at short horizons.

Keywords: Machine Learning; Forecasting; Structural Breaks; Model Selection; Cluster Analysis (search for similar items in EconPapers)
Date: 2021-10-13
New Economics Papers: this item is included in nep-big, nep-cmp, nep-ecm, nep-ets, nep-for and nep-ore
References: Add references at CitEc
Citations:

Downloads: (external link)
https://ora.ox.ac.uk/objects/uuid:d44fdba5-3436-4052-9cc4-2b2386b157bd (application/pdf)

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:oxf:wpaper:950

Access Statistics for this paper

More papers in Economics Series Working Papers from University of Oxford, Department of Economics Contact information at EDIRC.
Bibliographic data for series maintained by Anne Pouliquen ( this e-mail address is bad, please contact ).