Machine Learning for Time Series Forecasting Using State Space Models

Jose M. Sanchez-Bornot; Roberto C. Sotero

doi:10.1007/978-3-031-48232-8_43

Machine Learning for Time Series Forecasting Using State Space Models

Jose M. Sanchez-Bornot, Roberto C. Sotero

Research output: Chapter in Book/Report/Conference proceeding › Chapter › peer-review

1 Citation (Scopus)

Abstract

State-space models (SSMs) are becoming mainstream for time series analysis because of their flexibility and retained explainability, as they model observations separately from the underlying processes dynamics. Critically, the use of multivariate autoregressive (MVAR) modeling to capture the processes dynamics allows to model the system evolution through time and better interpretation of latent variable interactions. However, three important challenges remain critically unsolved: model estimation in (i) large-scale scenarios or (ii) with very noisy data; and (iii) model selection. In this study, we explore a state-space alternating least squares (SSALS) algorithm for time series forecasting and demonstrate its application with simulated and real data. We also demonstrate how to solve model selection for MVAR-based SSMs with a novel cross-validation technique. Altogether, testing this methodology with time series forecasting is ideal to demonstrate its strengths and weaknesses, and appreciate its advantages compared to current methods.

Original language	English
Title of host publication	International Conference on Intelligent Data Engineering and Automated Learning
Publisher	Springer Nature
Pages	470-482
Number of pages	13
Volume	14404
ISBN (Electronic)	978-3-031-48232-8
ISBN (Print)	9783031482311
DOIs	https://doi.org/10.1007/978-3-031-48232-8_43
Publication status	Published online - 15 Nov 2023

Bibliographical note

Funding Information:
The authors are grateful for access to the Tier 2 HighPerformance Computing resources provided by the Northern Ireland High Performance Computing (NI-HPC) facility funded by the UK Engineering and Physical Sciences Research Council (EPSRC), Grant No. EP/T022175/1. RCS was supported by RGPIN2022-03042 from Natural Sciences and Engineering Research Council of Canada.

Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023.

Keywords

machine learning
state space models
time series analysis

Access to Document

10.1007/978-3-031-48232-8_43

Cite this

@inbook{88dd80f1609a48ffa551ee19efa937cc,

title = "Machine Learning for Time Series Forecasting Using State Space Models",

abstract = "State-space models (SSMs) are becoming mainstream for time series analysis because of their flexibility and retained explainability, as they model observations separately from the underlying processes dynamics. Critically, the use of multivariate autoregressive (MVAR) modeling to capture the processes dynamics allows to model the system evolution through time and better interpretation of latent variable interactions. However, three important challenges remain critically unsolved: model estimation in (i) large-scale scenarios or (ii) with very noisy data; and (iii) model selection. In this study, we explore a state-space alternating least squares (SSALS) algorithm for time series forecasting and demonstrate its application with simulated and real data. We also demonstrate how to solve model selection for MVAR-based SSMs with a novel cross-validation technique. Altogether, testing this methodology with time series forecasting is ideal to demonstrate its strengths and weaknesses, and appreciate its advantages compared to current methods.",

keywords = "machine learning, state space models, time series analysis",

author = "Sanchez-Bornot, {Jose M.} and Sotero, {Roberto C.}",

note = "Funding Information: The authors are grateful for access to the Tier 2 HighPerformance Computing resources provided by the Northern Ireland High Performance Computing (NI-HPC) facility funded by the UK Engineering and Physical Sciences Research Council (EPSRC), Grant No. EP/T022175/1. RCS was supported by RGPIN2022-03042 from Natural Sciences and Engineering Research Council of Canada. Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Switzerland AG 2023.",

year = "2023",

month = nov,

day = "15",

doi = "10.1007/978-3-031-48232-8_43",

language = "English",

isbn = "9783031482311",

volume = "14404",

pages = "470--482",

booktitle = "International Conference on Intelligent Data Engineering and Automated Learning",

publisher = "Springer Nature",

address = "Switzerland",

}

TY - CHAP

T1 - Machine Learning for Time Series Forecasting Using State Space Models

AU - Sanchez-Bornot, Jose M.

AU - Sotero, Roberto C.

N1 - Funding Information: The authors are grateful for access to the Tier 2 HighPerformance Computing resources provided by the Northern Ireland High Performance Computing (NI-HPC) facility funded by the UK Engineering and Physical Sciences Research Council (EPSRC), Grant No. EP/T022175/1. RCS was supported by RGPIN2022-03042 from Natural Sciences and Engineering Research Council of Canada. Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023.

PY - 2023/11/15

Y1 - 2023/11/15

N2 - State-space models (SSMs) are becoming mainstream for time series analysis because of their flexibility and retained explainability, as they model observations separately from the underlying processes dynamics. Critically, the use of multivariate autoregressive (MVAR) modeling to capture the processes dynamics allows to model the system evolution through time and better interpretation of latent variable interactions. However, three important challenges remain critically unsolved: model estimation in (i) large-scale scenarios or (ii) with very noisy data; and (iii) model selection. In this study, we explore a state-space alternating least squares (SSALS) algorithm for time series forecasting and demonstrate its application with simulated and real data. We also demonstrate how to solve model selection for MVAR-based SSMs with a novel cross-validation technique. Altogether, testing this methodology with time series forecasting is ideal to demonstrate its strengths and weaknesses, and appreciate its advantages compared to current methods.

AB - State-space models (SSMs) are becoming mainstream for time series analysis because of their flexibility and retained explainability, as they model observations separately from the underlying processes dynamics. Critically, the use of multivariate autoregressive (MVAR) modeling to capture the processes dynamics allows to model the system evolution through time and better interpretation of latent variable interactions. However, three important challenges remain critically unsolved: model estimation in (i) large-scale scenarios or (ii) with very noisy data; and (iii) model selection. In this study, we explore a state-space alternating least squares (SSALS) algorithm for time series forecasting and demonstrate its application with simulated and real data. We also demonstrate how to solve model selection for MVAR-based SSMs with a novel cross-validation technique. Altogether, testing this methodology with time series forecasting is ideal to demonstrate its strengths and weaknesses, and appreciate its advantages compared to current methods.

KW - machine learning

KW - state space models

KW - time series analysis

UR - https://doi.org/10.1007/978-3-031-48232-8_43

UR - http://www.scopus.com/inward/record.url?scp=85177841188&partnerID=8YFLogxK

U2 - 10.1007/978-3-031-48232-8_43

DO - 10.1007/978-3-031-48232-8_43

M3 - Chapter

SN - 9783031482311

VL - 14404

SP - 470

EP - 482

BT - International Conference on Intelligent Data Engineering and Automated Learning

PB - Springer Nature

ER -