Revisiting the Holt-Winters' Additive Method for Better Forecasting

Revisiting the Holt-Winters' Additive Method for Better Forecasting

Seng Hansun (Universitas Multimedia Nusantara, Tangerang, Indonesia), Vincent Charles (University of Buckingham, Buckingham, UK), Christiana Rini Indrati (Universitas Gadjah Mada, Yogyakarta, Indonesia) and Subanar (Universitas Gadjah Mada, Yogyakarta, Indonesia)
Copyright: © 2019 |Pages: 15
DOI: 10.4018/IJEIS.2019040103

Abstract

Time series are one of the most common data types encountered by data scientists and, in the context of today's exponentially increasing data, learning how to best model them to derive meaningful insights is an important skill in the Big Data and Data Science toolbox. As a result, many researchers have dedicated their efforts to developing time series analysis methods to predict future values based on previously observed values. One of the well-known methods is the Holt-Winters' seasonal method, which is commonly used to capture the seasonality effect in time series data. In this study, the authors aim to build upon the Holt-Winters' additive method by introducing new formulas for finding the initial values. Obtaining more accurate estimations of the initial values could result in a better forecasting result. The authors use the basic principle found in the weighted moving average method to assign more weight to the most recent data and combine it with the original initial conditions found in the Holt-Winters' additive method. Based on the experiment performed, the authors conclude that the new formulas for finding the initial values in the Holt-Winters' additive method could give a better forecasting when compared to the traditional Holt-Winters' additive method and the weighted moving average method in terms of the accuracy level.
Article Preview
Top

1. Introduction

Forecasting can be defined as the prediction of future events based on foreknowledge acquired through a systematic process or intuition (Soyiri & Reidpath, 2013). Yi, Ke, and Junde (2014) described forecasting as an art and science to predict future events through historical data and mathematical models. It uses only past information for the variable being forecasted, under the assumption that the observed trend and seasonality will persist (Dong, Sigrin, & Brinkman, 2017). The correct understanding and implementation of forecasting methods could be useful in many fields. Many private and public organizations use forecasting to inform a number of decisions (Fye, Charbonneau, Hay, & Mullins, 2013). For example, a forecasting model that could accurately predict the direction of the stock market on the following day would be extremely valuable and profitable (Evans, 2003; Diebold & Yilmaz, 2012). Excellence in sales forecasting can amplify a firm’s financial health and gratify customers and employees (Moon, Mentzer, Smith, & Garver, 1998). At a larger scale, better forecasting techniques for electricity price forecasts have become a vital input to energy companies’ decision-making mechanisms and effective risk management (Weron, 2014; Christensen, Hurn, & Lindsay, 2012). Time series are one of the most common data types encountered by data scientists and, in the context of today’s exponentially increasing data, learning how to best model them to derive meaningful insights is an important skill in the Big Data and Data Science toolbox. A discussion on Data Science and Big Data is beyond the purposes of the present paper, but for some general information on the concepts and recent developments or topics of interest, the interested reader can refer to the works of Charles and Emrouznejad (2018), Charles and Gherman (2013, 2018), Charles, Tavana, and Gherman (2015), and Emrouznejad and Charles (2018), among others.

To get a better forecasting result, different time series analysis methods have been developed in time. Some researchers used conventional methods, such as the moving average and exponential smoothing models, which can be found in the works of Grebenkov and Serror (2014), Klinker (2011), Papailias and Thomakos (2012), and Goodwin (2010), while some others used soft computing methods, such as fuzzy logic, neural networks, and genetic algorithms (see, e.g., Stevenson & Porter, 2009; Hansun, 2012; Faraway & Chatfield, 1998; Oancea & Ciucu, 2013; and Chai, Chuek, Mital, & Tan, 1997). Some researchers even tried to combine two or more methods to create new hybrid forecasting methods, as the ones that can be found in the works of Garcia, Figueroa, Lourdes, Vasquez, Vega, and Hernandez (2014), Thakur, Kumar, and Tiwari (2015), Nasseri, Asghari, and Abedini (2008), Draidi and Labed (2015), Popoola, Ahmad, and Ahmad (2004), Hassan, Jaafar, Samir, and Jilani (2012), and Ferdinandoa, Pasila, and Kuswanto (2010).

Complete Article List

Search this Journal:
Reset
Open Access Articles: Forthcoming
Volume 16: 4 Issues (2020): 1 Released, 3 Forthcoming
Volume 15: 4 Issues (2019)
Volume 14: 4 Issues (2018)
Volume 13: 4 Issues (2017)
Volume 12: 4 Issues (2016)
Volume 11: 4 Issues (2015)
Volume 10: 4 Issues (2014)
Volume 9: 4 Issues (2013)
Volume 8: 4 Issues (2012)
Volume 7: 4 Issues (2011)
Volume 6: 4 Issues (2010)
Volume 5: 4 Issues (2009)
Volume 4: 4 Issues (2008)
Volume 3: 4 Issues (2007)
Volume 2: 4 Issues (2006)
Volume 1: 4 Issues (2005)
View Complete Journal Contents Listing