Heterogeneous Forecast Adjustment

• Author: Bert De Bruijn,Philip Hans Franses
• DOI: http://doi.org/10.1002/for.2433
• Date: 01 July 2017
• Published date: 01 July 2017

Heterogeneous Forecast Adjustment

BERT DE BRUIJN AND PHILIP HANS FRANSES*

Econometric Institute, Erasmus School of Economics, PO Box 1738, NL-3000 DR Rotterdam, The

Netherlands

ABSTRACT

There is ample empirical evidence that expert-adjusted model forecasts can be improved. One way to potential

improvement concerns providing various forms of feedback to the sales forecasters. It is also often recognized that

the experts (forecasters) might not constitute a homogeneous group. This paper provides a data-based methodology

to discern latent clusters of forecasters, and applies it to a fully new large database with data on expert-adjusted

forecasts, model forecasts and realizations. For the data at hand, two clusters can clearly be identified. Next, the

consequences of having clusters are discussed. Copyright © 2016 John Wiley & Sons, Ltd.

key words forecast support system; sales forecasters; forecast accuracy; feedback; latent classes

INTRODUCTION

Sales forecasters often rely on the output of a forecast support system (FSS) when they create their own forecasts. A

typical situation is that an FSS delivers forecasts and that an individual forecaster modifies these as he or she sees fit.

Oftentimes, an FSS includes forecast algorithms that only include recent past sales data, while the forecaster with

domain-specific knowledge may believe that additional information can be useful and thus delivers an adjusted

forecast (see Goodwin, 2000, 2002). Due to the recent availability of novel databases including FSS-based forecasts,

managers’forecasts and actual realizations, more insights are gained as to how these adjusted forecasts perform, how

they are actually constructed and how they can be improved. Fildes et al. (2009) and Syntetos et al. (2009) marked the

start of this large data-based researchand Franses (2014) recently summarized various findings across a rangeof studies.

The main stylized facts concerning manually adjusted FSS forecasts seem to be that (i) FSS forecasts are almost invariably

adjusted by individual forecasters, that (ii) expert-adjusted forecasts are often not as accurate as the FSS forecasts and that (iii)

there are various ways to improve current approaches of adjusting FSS forecasts. One way to achieve improved adjusted fore-

casts is to provide feedback to the forecasters on their actual behavior and on their past track record. Legerstee and Franses

(2014) analyze the behavior of forecasters who deliver forecasts for monthly sales data, where they compare data before and

after the moment that experts received different kinds of feedback on their behavior and their task. They conclude that after

feedback the adjusted forecasts deviated less from the FSS forecasts and that their accuracy had improved substantially.

At the same time, various studies suggest that individuals who manually adjust model-based forecasts do not

constitute a homogeneous group. Boulaksil and Franses (2009) interviewed various forecasters of whom some say

that they never look at FSS forecasts when creating their own, and others say that they deviate only a little from

the FSS forecasts. In the earnings forecasting literature, there also appear to be one or more different types of behavior

of forecasters, depending on whether they want to deliver accurate forecasts or want to stand out with exceptional

quotes (see Clement and Tse, 2005; Jegadeesh and Kim, 2010). Similarly, in the macroeconomics literature there

are also examples of differing behavior across forecasters (see, for example, Lamont, 2002).

When there are different types of forecasters who behave differently when adjusting FSS forecasts, then feedback to these

different types of forecasters most likely should also be different, and this is an important premise in this paper. Hence, to

improve the overall quality of adjusted forecasts, one may wish to discern groups of individual forecasters with common be-

havior within the group and differing behavior across the groups. As records of the adjustment process typically do not exist

(Franses, 2014), the division of forecasters into various groups must be done using the available data. In this paper we there-

fore propose a method to disentangle groups of forecasters based on their actual behavior. Knowing the clusters can lead to

more tailor-made feedback, and we recommend such variants of feedback in our case study below. The database in our case

study has never been analyzed before and concerns the FSS forecasts, adjusted forecasts and realizations of sales of stock

keeping units (SKUs) of a very large Germany-based pharmaceutical company.

The outline of our paper is as follows. In the next section we discuss the database. In the third section we provide

our methodology to link the behavior of sales forecasters with their forecast performance, while allowing for latent

classes of individual forecasters with common behavior. In the fourth section we discuss the main results for our

*Correspondence to: Philip Hans Franses, Erasmus School of Economics, Rotterdam, The Netherlands.

E-mail: franses@ese.eur.nl

Journal of Forecasting,J. Forecast. 36, 337–344 (2017)

Published online 28 July 2016 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/for.2433

Copyright © 2016 John Wiley & Sons, Ltd.