Effects of winsorization: The cases of forecasting non‐GAAP and GAAP earnings

• Author: Ruby Brownen‐Trinh
• Date: 01 January 2019
• DOI: http://doi.org/10.1111/jbfa.12365
• Published date: 01 January 2019

DOI: 10.1111/jbfa.12365

Effects of winsorization: The cases of forecasting

non-GAAP and GAAP earnings

Ruby Brownen-Trinh

Department of Accounting and Finance, School

of Economics, Finance and Management,

University of Bristol, Bristol, UK

Correspondence

RubyBrownen-Trinh,Department of Accounting

andFinance, School of Economics, Finance and

Management,University of Bristol, The Priory

RoadComplex, Priory Road, Clifton, Bristol BS8

1TU,UK.

Email:Ruby.Brownen-Trinh@bristol.ac.uk

JELClassification: G10, G11, G17, M41

Abstract

This study examines how the winsorization procedure affects the

performance of regression-based earnings forecasting models. I find

that the impact is multifaceted and depends principally on three fac-

tors: the level of data errors in the tails, the characteristics of firms

affected by the process, and the use of scaling. Fora non-GAAP earn-

ings yield specification, where data input errors exist, winsorization

changes the information set in a non-systematic way and helps to

improve the performance of regression-based forecasts, especially

when the least squares estimator is employed. However,for a non-

GAAP earnings per share specification, with fewer data input errors

found in the tails of the distribution, winsorization has a particularly

strong effect on very large companies, lowering the economic value

of earnings predictions. I observe similar results for corresponding

GAAP earnings specifications. Robust estimators, such as least abso-

lute deviation, high breakdown-point and Theil-Sen, appear to be a

more effective solution than winsorization. Their earnings forecasts

consistently yield significant positive abnormal returns across non-

GAAP and GAAP earnings specifications.

KEYWORDS

earnings forecasts, influential observations, robust regression,

scaling, stock returns, winsorization

1INTRODUCTION

Sets of observations which have been de-tailed by over-vigorous use of a rule for rejectingoutliers are inappro-

priate, since they are not samples.

Tukey(1960)

While econometric studies warn against the use of a winsorization process that replaces sample values aboveor below

a given percentile of the sample distribution with the values at the respective percentiles, the majority of empiri-

cal accounting studies employ this process (Leone, Minutti-Meza, & Wasley,2017). Extreme observations/outliers in

cross-sectional data used in these studies lead to biased coefficient estimates and heteroscedastic regression errors

(Barth & Kallapur, 1996), and winsorization appears to be a simple and convenient solution. In cases where outliers

J Bus Fin Acc. 2019;46:105–135. wileyonlinelibrary.com/journal/jbfa c

2018 John Wiley & Sons Ltd 105

106 BROWNEN-TRINH

occur due to shocks or data entry errors, winsorization helps to remove the effect of these observations (Leone et al.,

2017).However, when extreme values are just reflections of the cross-sectional variation in firm characteristics such as

firm size or profitability,this procedure risks systematically altering the data and any economic inferences for a subset

of firms that may form important parts of investment strategies.Hence, it risks affecting the efficiency and usefulness

of coefficient estimates.

This study seeks to shed light on these issues by addressing three questions. First, what is the extent and impact of

data input errors in regression-based forecasts of earnings? Second, what is the impact of the winsorization process on

the statistical performance of regression estimators? Third, how does winsorization affect the investment usefulness

of earnings predictions?

I use the earnings forecast setting for several reasons. First, earnings forecasts are a key determinant of equity

value (Ohlson, 1995; Ohlson & Juettner-Nauroth, 2005) and as such are important to investors in portfolio forma-

tion (Frankel& Lee, 1998; Hou, van Dijk, & Zhang, 2012). Although most investors rely on financial analysts’ forecasts

(Brown, Hagerman, Griffin, & Zmijewski, 1987), manystudies find that these are frequently biased (see, e.g., Bradshaw,

Richardson, & Sloan, 2001; Dichev & Tang,2009; Frankel & Lee, 1998). Therefore, a great deal of research has been

devoted to the developmentof bias-free regression-based forecasts. These forecasting models frequently rely on win-

sorization to reduce the effect of observations with extreme values (e.g., Harris & Wang, 2013; Hou et al., 2012; So,

2013). Hence, while claiming to outperform the forecasts of financial analysts in terms of accuracy,their results may

be limited to a specific sample and potential distortions by the winsorization process are largely ignored. Second, the

use of earnings forecasts in the pricing of stocks and in portfolio formation (see, e.g., Black, Christensen, Ciesielski,

& Whipple, 2018; Bradshaw & Sloan, 2002; Bradshaw,Christensen, Gee, & Whipple, 2018) allows me to look beyond

conclusions offered by existing studies on winsorization, such as Leone et al. (2017), byalso considering the impact of

winsorization on economic values.

Tobegin, I examine the authenticity of archival Generally Accepted Accounting Principles (GAAP) and non-GAAP

earnings data.1I manually check earnings figures in 10-K reports and find that the total GAAP earnings data down-

loaded from Compustat are highly reliable. Meanwhile,for non-GAAPearnings per share downloaded from the I/B/E/S

database, the veracityof 49% of the data in the tails of the distribution is questionable, being more than double the cor-

responding GAAP earnings per share. Here, the role of winsorization might serve different purposes and it might have

different effects. Fornon-GAAP earnings that claim to consist of recurring items, winsorization might help remove data

input errors, while for GAAP earnings, it might help to remove non-recurring items.

Toprovide some insights into the nature of the data in the tails, I carry out a further investigation of the companies

whose GAAP and non-GAAP earnings are likely to be replaced by winsorized values in a cross-sectional regression for

both unscaled and scaled earnings (namely, total earnings, earnings per share and earnings yield). I find that, for the

total earnings and for the earnings per share specifications, the upper tails of earnings distributions consist of genuine

earningsfigures of many important corporations, such as General Motors, Berkshire Hathaway, GeneralElectric, Exxon

Mobil Corporation,and IBM Corporation, all of which play a major role in capital market investment due to their promi-

nence in typical portfolios. In these cases, replacing reported accounting figures with winsorized values that are more

“acceptable”introduces statistical bias and potentially misleading information about large and economically important

companies.Earnings forecasts, therefore, mayhave less economic value even if they appear to have low forecast errors.

However,scaling by market capitalization changes the distribution of earnings. Here, winsorization of companies in the

tails appears to be non-systematic and the impact of winsorization on economic values is less likely to be serious.

I employ both GAAP and non-GAAP earnings forecasts to formally examine the effect of winsorization, with par-

ticular focus on the non-GAAP measure because of its availability, importance and relevance to investorsand other

stakeholders(see, e.g., Bentley, Christensen, Gee, & Whipple, 2018; Black et al., 2018; Bradshaw & Sloan, 2002; Brown

& Sivakumar, 2003; Hoogervorst, 2016; Wieland, Dawkins, & Dugan, 2013). I use both unscaled and scaled earnings

1Non-GAAP (“Street”) earnings numbers are the figures announced by corporations in their press releases and trackedby analyst estimate clearinghouse

services(Bradshaw & Sloan, 2002). They contain only the continuing component of GAAP earnings (Brown, Call, Clement, & Sharp, 2015).