The persistence forecast effect in time-series predictions

  • Huseyin Burak Akyol

Student thesis: Doctoral ThesisDoctor of Philosophy (PhD)

Abstract

Over the past few years, time-series forecasting has been put into practice for a variety of purposes across various fields. It is regarded as crucial for many types of organisations and applications, as effective decision-making processes and intelligent autonomous systems depend heavily on predictions of the future. Consequently, time-series forecasting has drawn the attention of academic researchers and industry professionals towards its robust implementation and then reliable evaluation. However, due to the many practical challenges and issues involved, these tasks are not always straightforward and achievable. This thesis provides a formal definition and in-depth investigation of an important form of bias that has so far been overlooked in the literature and thus works towards more robust time-series prediction models and their reliable assessments.

The bias, undermining the quality of models and skewing the predictions in a systematic way, arises when the underlying time-series data lacks regularity and consistency. When it occurs, it is observed that forecasting outputs systematically approximate one of the most recently observed values used in the input feature set, resulting in a series of predicted values that is almost identical to the series of observed values but is continuously delayed by a few steps in time. However, this behaviour cannot be detected by the current accuracy assessment methods, which ultimately leads to overconfidence in forecasting models and prediction outputs.

Therefore, with the objective of guarding time-series models and predictions against the bias to achieve robust models and reliable predictions, this thesis provides a formal definition of the bias, explores its characteristics in greater detail, establishes the factors causing the bias, evaluates its potential negative implications, proposes a novel method for quantitative detection of the bias, and finally, investigates the prevalence of the bias in the literature and discusses how the bias may invalidate the outcomes of previously published works. Moreover, it presents experimental studies using real-world domestic electricity consumption time-series data to demonstrate the bias and the implementation of the proposed method within a realistic setting.
Date of Award12 Jul 2023
Original languageEnglish
Awarding Institution
  • The University of Bristol
SupervisorDaniel Schien (Supervisor) & Chris W Preist (Supervisor)

Cite this

'