Does the kitchen‐sink model work forecasting the equity premium?

• Published date: 01 March 2022
• Author: Anwen Yin
• Date: 01 March 2022
• DOI: http://doi.org/10.1111/irfi.12352

ORIGINAL ARTICLE

Does the kitchen-sink model work forecasting

the equity premium?

Anwen Yin

A.R. Sanchez, Jr. School of Business, Texas

A&M International University, Laredo, Texas

Correspondence

Anwen Yin, A.R. Sanchez, Jr. School of

Business, Texas A&M International University,

Laredo, TX 78041, USA.

Email: anwen.yin@tamiu.edu

Abstract

We propose applying partial least squares (PLS) to estimat-

ing the previously considered ineffective multivariate regres-

sion model when forecasting the market equity premium

out-of-sample. First, PLS is a dimension reduction method

that effectively addresses the issue of multicollinearity prev-

alent among financial variables. Second, PLS constructs fac-

tors with the supervision of past equity premiums, resulting

in an explicit linkage between the forecasting target and PLS

components. Our empirical results show that the PLS-

estimated kitchen-sink model consistently and robustly out-

performs many competing alternatives, such as shrinkage

estimators and forecast combinations, by a statistically and

economically significant margin. Our analysis differs from

Kelly and Pruitt (2013) in factors such as data source, model

estimation and specification, and economicrationale.

KEYWORDS

dimension reduction, equity premium, forecast evaluation,

multicollinearity, partial least squares

JEL CLASSIFICATION

C53; C58; G11; G17

1|INTRODUCTION

Accurate forecasts of the equity premium play a critical role in diverse areas of empirical finance such as optimal

portfolio decisions, capital budgeting, and performance evaluation of investment funds managers (see, for example,

Ait-Sahali & Brandt, 2001; Avramov & Wermers, 2006; Barberis, 2000; Xia, 2001). Consequently, numerous studies

Received: 3 October 2019 Revised: 25 January 2021 Accepted: 15 April 2021

DOI: 10.1111/irfi.12352

© 2021 International Review of Finance Ltd.

International Review of Finance. 2022;22:223–247. wileyonlinelibrary.com/journal/irfi 223

have provided empirical evidences of both the in-sample and out-of-sample predictability for a multitude of financial

and economic variables forecasting the equity premium (see, for instance, Campbell, 1987; Campbell & Shiller, 1988;

Fama & French, 1988; Fama & French, 1989; Faria & Verona, 2018; Ferreira & Santa-Clara, 2011; Jiang, Lee, Mar-

tin, & Zhou, 2019; Li, Ng, & Swaminathan, 2013; Ma, Wen, Wang, & Jiang, 2019; Rapach, Ringgenberg, &

Zhou, 2016).

Given the large and constantly expanding set of predictors for the market equity premium, researchers might

reasonably employ a multivariate regression model comprising all the available variables, which is often termed the

“kitchen-sink”model in related literature. Intuitively, one would expect that the kitchen-sink model which harnesses

all the available information would lead to superior predictive performance. However, contrary to the aforemen-

tioned expectation, by undertaking a comprehensive analysis Goyal and Welch (2008) show that the kitchen-sink

model consistently fails to generate superior predictive gains compared with the simple historical average

benchmark.

In light of the documented failure of the kitchen-sink model, subsequent research has attempted to uncover the

possible predictive content of variables by adopting a “one-variable-at-a-time”approach. It involves estimating uni-

variate regressions, making forecasts separately, then aggregating all predictions via a given weighting scheme to

form a single forecast. Studies such as Rapach, Strauss, and Zhou (2010), Dangl and Halling (2012), and Pettenuzzo,

Timmermann, and Rossen (2014) follow such a strategy to support the predictability of the aggregate equity pre-

mium. Although pooling forecasts have proved successful in uncovering the predictive content of many variables,

recently, studies such as Kelly and Pruitt (2013) and Li and Tsiakas (2017) argue that the pooling of information

rather than forecasts may deliver better gains as it makes efficient use of all available information. This paper follows

the latter approach.

As such, we contribute to the literature of forecasting stock returns by proposing using the partial least squares

(PLS) to estimate the kitchen-sink predictive model as opposed to the ordinary least squares (OLS) estimation consid-

ered in Goyal and Welch (2008). Our empirical results show that not only does the PLS-estimated kitchen-sink model

outperform the simple yet difficult to beat historical average, but also it consistently generates superior statistical

and economic gains to investors who use its forecasts to guide optimal portfolio decisions. While we do not attempt

to make theoretical advances, we do focus on several important empirical issues which are often overlooked in prac-

tice: First, why does the kitchen-sink model perform the worst in Goyal and Welch (2008)? Next, can we uncover

the possible predictive prowess of the kitchen-sink model? Finally, which method works the best in terms of generat-

ing statistical and economic gains?

The closest study related to this paper is Kelly and Pruitt (2013). However, our analysis differs from Kelly and

Pruitt (2013) in several aspects. Table 1 summarizes the major differences between the analysis conducted in Kelly

and Pruitt (2013) and that performed in this article. First, we use the updated data maintained by Amit Goyal while

Kelly and Pruitt (2013) mainly source data from Ken French.

1

This renders our empirical results easily comparable

with those from articles such as Rapach et al. (2010), Pettenuzzo et al. (2014) and Li and Tsiakas (2017). Second, we

construct PLS factors from a set of aggregate financial and economic variables while Kelly and Pruitt (2013) primarily

build PLS components by condensing information from a number of disaggregated valuation ratios. Third, we opti-

mally select the number of PLS components used when forecasting the equity premium at each out-of-sample step

via cross-validation while Kelly and Pruitt (2013) arbitrarily use only one PLS component for ease of interpretation

and comparison. Fourth, Kelly and Pruitt (2013) estimate regressions using an adapted PLS procedure termed

“three-pass regression filter”while we estimate the kitchen-sink model via the standard PLS, which can be easily

implemented with established packages of programming languages or statistical software. Finally, we compare PLS

regressions with a wealth of alternative models, many of which have been previously shown effective. In contrast,

Kelly and Pruitt (2013) largely compare PLS forecasts with those from univariate models which have been shown to

be inferior in Goyal and Welch (2008).

We begin by showing that the strong, persistent, and seemingly time-varying sample correlations among predic-

tors in our data might lead to the failure of the OLS-estimated kitchen-sink model. As a result, we suggest using the

224 YIN