Exploring the association between time series features and forecasting by temporal aggregation using machine learning - Empirical study

Aug 5, 2026

Upload by Himbo

When forecasting a total value over multiple future time periods, two temporal aggregation (TA) approaches are used: aggregated forecasting (AF) and aggregate data via non-overlapping temporal aggregation (AD). Practice often suggests matching aggregation frequency to decision needs, yet superiority may vary across situations with limited evidence on deciding indicators. An empirical experiment on monthly M4 competition series tests both approaches and builds feature-based supervised learning to predict whether AF or AD should be used.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages21

Reading Time53 min

AccessFree to download

Introduction

Temporal aggregation approaches for forecasting

Feature-based supervised learning framework

Machine learning models and performance measures

ORCA – Online Research @ Cardiff This is an Open Access document downloaded from ORCA, Cardiff University' s institution al r epo sitory: [http s://orc a . c ardiff. ac. uk/id/ e print/ 1 6 0 4 6 5 /](http s://orc a . c ardiff. ac. uk/id/ e print/ 1 6 0 4 6 5 /) This is the author’s version of a work that was submitted to / accepted for publication . Citation for final published version: Rost ami-Ta bar, Bah man and Mir ce tic, Dejan 2 0 2 3 . Exploring the association between time series feature s and forecasting by temporal aggreg ation using m achine learning. Neuro computing 5 4 8 , 1 2 6 3 7 6 . 1 0 . 1 0 1 6/j. neu com. 2 0 2 3 . 1 2 6 3 7 6 file Publishers p age : [http s:// doi.org/ 1 0 . 1 0 1 6 /j. neu com. 2 0 2 3 . 1 2 6 3 7 6](http s:// doi.org/ 1 0 . 1 0 1 6 /j. neu com. 2 0 2 3 . 1 2 6 3 7 6) Plea se note : Change s m ade a s a result of publishing processes such a s copy-editing, formatting and p age numbers m ay not be reflected in this version. For the definitive version of this publication, plea se refer to the published source. You are advised to consult the publisher’s version if you wish to cite this p aper. This version is being m ade available in accord ance with publisher policies. See [http://orca . cf. ac. uk/ policies. html](http://orca . cf. ac. uk/ policies. html) for u s age policies. Copyright and mor al right s for publications m ade available in ORCA are retained by the copyright holders. Exploring the association between time series features and forecasting by temporal aggregation using machine learning Bahman Rostami-Tabara, 􀀁 , Dejan Mirceticb, 1 a Cardiff Business School, 3 Colum Drive, CF10 3EU Cardif, United Kingdom b Institute for Artiﬁcial Intelligence Research and Development of Serbia, Fruskogorska 1, 21000 Novi Sad, Serbia | a r t i c l e i n f o | | a b s t r a c t | | --- | --- | --- | | Article history:
Received 28 June 2022
Revised 28 February 2023
Accepted 22 May 2023
Available online 30 May 2023 Communicated by Zidong Wang | | When a forecast of the total value over several time periods ahead is required, forecasters are presented with two temporal aggregation (TA) approaches to produce required forecasts: i) aggregated forecast (AF) or ii) aggregate data using non-overlapping temporal aggregation (AD). Often, the recommendation is to aggregate data to a frequency relevant to the decision the eventual forecast will support and then produce the forecast. However, this might not be always the best choice and we argue that both AF and AD approaches may outperform each other in different situations. Moreover, there is a lack of evidence on what indicators may determine the superiority of each approach. We design and execute an empirical experiment framework to ﬁrst explore the performance of these approaches using monthly time series of M4 competition dataset. We further turn the problem into a classiﬁcation supervised learning by constructing a database consisting of features of each time series as predictor and model class labelled as AF/AD as response/outcome. We then build machine learning algorithms to investigate the association between time series features and the performance of AF and AD. Our ﬁndings suggest that both AF and AD approaches may not consistently generate accurate results for every individual series. AF is shown to be signiﬁcantly better than AD for the monthly M4 time series, especially for longer horizons. We build several machine learning approaches using a set of extracted time series features as input to predict accurately whether AD or AF should be used. We ﬁnd out that Random Forest (RF) is the most accurate approach in correctly classifying the outcome assessed both by statistical measures such as misclassiﬁcation error, F-statistics, area under the curve, and a utility measure. The RF approach reveals that curvature, nonlinearity, seas_pacf, unitroot_pp, mean, ARCHM.LM, Coefﬁcient of Variation, stability, li

FAQs

What are the two temporal aggregation approaches compared in the study?

The study compares aggregated forecasting (AF) and aggregated data using non-overlapping temporal aggregation (AD). Both are designed to produce required forecasts for totals over multiple future periods.

How is the decision of AF vs AD modeled?

The work constructs a supervised learning dataset where each time series’ extracted features are predictors and the outcome labels which approach (AF or AD) performed better.

Which machine learning method and features drive the predictions?

Random Forest (RF) provides the most accurate classification of the better-performing approach, and feature importance highlights variables related to curvature, nonlinearity, seasonal PACF, unit roots, mean level, stability, and level shifts.

Aug 5, 2026

Upload by Himbo