Time series extrinsic regression algorithms for forecasting long time series with a short horizon
Time series extrinsic regression algorithms for forecasting long time series with a short horizon
Time series forecasting has a rich research history, with a variety of use cases across diverse domains. Many machine learning and deep learning forecasting algorithms approach the problem as regression by windowing. Given a univariate time series Y=(y 1,..,y T), a training set of input-output pairs (x i,y i) is constructed where the input window x i=(y i-p+1,..,y i) contains the most recent p observations, and the target variable z i is a future value y i+h, with h as the forecasting horizon. We call this approach time series forecasting regression (TSFR). Recently, an alternative type of problem, time series extrinsic regression (TSER), has been described. In TSER, each training instance is a complete time series X i⊂ℜ, assumed independent of the others, paired with an external response variable z i⊂ℜ. Unlike TSFR, the response is not a future point of the same series but an exogenous variable to be predicted from the overall dynamics of the series. An archive of 63 TSER datasets was released in 2024. Our aim is to bridge the gap between TSER and TSFR by providing a common framework for both. We reformat a selection of diverse series used in forecasting research as TSER problems through windowing and create train/test splits. We then compare state-of-the-art TSER algorithms on these data and benchmark them against standard statistical and deep learning approaches. We find that the pattern of results seen on the TSFR archive does not mirror that on the TSER data. We propose a simple reformulation that allows the best TSER algorithm to achieve performance equivalent to statistical forecasters.
Time series forecasting, time series machine learning, time series regression
149-164
Banwell, Alexander
4365a89b-d67d-4ae3-b75f-c35a489375ff
Bagnall, Anthony
d31e6506-2a00-4358-ba3f-baefd48d59d8
21 February 2026
Banwell, Alexander
4365a89b-d67d-4ae3-b75f-c35a489375ff
Bagnall, Anthony
d31e6506-2a00-4358-ba3f-baefd48d59d8
Banwell, Alexander and Bagnall, Anthony
(2026)
Time series extrinsic regression algorithms for forecasting long time series with a short horizon.
Lemaire, Vincent, Ifrim, Georgiana, Bagnall, Anthony, Malinowski, Simon, Tavenard, Romain and Schäfer, Patrick
(eds.)
In Advanced Analytics and Learning on Temporal Data - 10th ECML PKDD Workshop, AALTD 2025, Revised Selected Papers.
vol. 16255 LNAI,
Springer Cham.
.
(doi:10.1007/978-3-032-15535-1_10).
Record type:
Conference or Workshop Item
(Paper)
Abstract
Time series forecasting has a rich research history, with a variety of use cases across diverse domains. Many machine learning and deep learning forecasting algorithms approach the problem as regression by windowing. Given a univariate time series Y=(y 1,..,y T), a training set of input-output pairs (x i,y i) is constructed where the input window x i=(y i-p+1,..,y i) contains the most recent p observations, and the target variable z i is a future value y i+h, with h as the forecasting horizon. We call this approach time series forecasting regression (TSFR). Recently, an alternative type of problem, time series extrinsic regression (TSER), has been described. In TSER, each training instance is a complete time series X i⊂ℜ, assumed independent of the others, paired with an external response variable z i⊂ℜ. Unlike TSFR, the response is not a future point of the same series but an exogenous variable to be predicted from the overall dynamics of the series. An archive of 63 TSER datasets was released in 2024. Our aim is to bridge the gap between TSER and TSFR by providing a common framework for both. We reformat a selection of diverse series used in forecasting research as TSER problems through windowing and create train/test splits. We then compare state-of-the-art TSER algorithms on these data and benchmark them against standard statistical and deep learning approaches. We find that the pattern of results seen on the TSFR archive does not mirror that on the TSER data. We propose a simple reformulation that allows the best TSER algorithm to achieve performance equivalent to statistical forecasters.
Text
AALTD_TSER_for_forecasting
- Accepted Manuscript
More information
Published date: 21 February 2026
Keywords:
Time series forecasting, time series machine learning, time series regression
Identifiers
Local EPrints ID: 512350
URI: https://eprints.soton.ac.uk/id/eprint/512350
ISSN: 0302-9743
PURE UUID: c2301dd2-dc64-407f-878e-582b498b2c2a
Catalogue record
Date deposited: 25 Jun 2026 16:44
Last modified: 13 Sep 2026 15:49
Export record
Altmetrics
Contributors
Author:
Alexander Banwell
Author:
Anthony Bagnall
Editor:
Vincent Lemaire
Editor:
Georgiana Ifrim
Editor:
Anthony Bagnall
Editor:
Simon Malinowski
Editor:
Romain Tavenard
Editor:
Patrick Schäfer
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics