Does a lot help a lot? Forecasting stock returns with pooling strategies in a data‐rich environment

• Author: Fabian Baetje
• DOI: http://doi.org/10.1002/for.2475
• Date: 01 January 2018
• Published date: 01 January 2018

RESEARCH ARTICLE

Does a lot help a lot? Forecasting stock returns with pooling

strategies in a data‐rich environment

Fabian Baetje

Department of Economics, Leibniz

Universität Hannover, Hannover, Germany

Correspondence

Fabian Baetje, Department of Economics,

Leibniz Universität Hannover,

Königsworther Platz 1, D‐30167 Hannover,

Germany

Email: baetje@gif.uni‐hannover.de

Abstract

A variety of recent studies provide a skeptical view on the predictability of stock

returns. Empirical evidence shows that most prediction models suffer from a loss

of information, model uncertainty, and structural instability by relying on low‐

dimensional information sets. In this study, we evaluate the predictive ability of var-

ious lately refined forecasting strategies, which handle these issues by incorporating

information from many potential predictor variables simultaneously. We investigate

whether forecasting strategies that (i) combine information and (ii) combine individ-

ual forecasts are useful to predict US stock returns, that is, the market excess return,

size, value, and the momentum premium. Our results show that methods combining

information have remarkable in‐sample predictive ability. However, the out‐of‐sam-

ple performance suffers from highly volatile forecast errors. Forecast combinations

face a better bias–efficiency trade‐off, yielding a consistently superior forecast per-

formance for the market excess return and the size premium even after the 1970s.

KEYWORDS

factor models, forecast combination, model uncertainty, principal components, return predictability

1|INTRODUCTION

The classical capital asset pricing model is often expanded by

additionally considering returns to a size portfolio (SMB), to

a value portfolio (HML), and to a momentum portfolio

(MOM) as risk factors (Carhart, 1997; Fama & French,

1993). The resulting four risk factors can usually be seen as

risk premia compensating investors for holding risky assets.

Due to the important role of these risk factors and their

related premia, the question arises: Are these risk premia

predictable by macroeconomic determinants?

We focus on the macroeconomic aspect because one may

argue that in the last instance the riskiness of firms is to a great

extent determined by the firms' exposure to macroeconomic

risks. This line of research has quite some history for equity

premium prediction and has developed established indicators,

such as the short‐term interest rate, the credit and the term

spread (see, e.g., Ang & Bekaert, 2007; Fama & French,

1989; Keim & Stambaugh, 1986), the inflation rate (e.g.,

Campbell & Vuolteenaho, 2004; Fama & Schwert, 1977;

Nelson, 1976), stock market volatility (investigated by Guo,

2006), or the consumption–wealth ratio provided by Lettau

and Ludvigson (2001), to name just a few. Most of the empir-

ical studies (e.g., Campbell & Shiller, 1988; Cochrane, 2008;

Lewellen, 2004) use valuation ratios such as the dividend

yield, the price–earnings ratio, or the book‐to‐market ratio,

which should serve as proxies for expected business condi-

tions, as mentioned by Campbell and Diebold (2009).

Although stock return predictability was accepted for a

long time, Goyal and Welch (2008), Timmermann (2008),

Griffin, Ji, and Martin (2003), and Griffin and Lemmon

(2002), among others, show that commonly used predictors

perform poorly in an out‐of‐sample (OOS) setting, especially

during more recent decades, which cast doubt on the empiri-

cal linkage between macroeconomic fundamentals and stock

returns. More recently, Rapach and Zhou (2013) provide a

survey of further approaches attempting to improve the fore-

cast performance. Campbell and Thompson (2008), Ferreira

Received: 5 October 2015 Revised: 30 November 2016 Accepted: 25 March 2017

DOI: 10.1002/for.2475

Journal of Forecasting. 2018;37:37–63. Copyright © 2017 John Wiley & Sons, Ltd.wileyonlinelibrary.com/journal/for 37

and Santa‐Clara (2011), and Pettenuzzo, Timmermann, and

Valkanov (2014) show that the application of economic con-

straints generally increases the overall forecast performance

of commonly used macroeconomic variables. However,

improvements are predominantly located in the distant past

(see Li & Tsiakas, 2016; Pettenuzzo et al., 2014). Further

forecasting advances incorporate the usage of pooling strate-

gies like principal component analysis (proposed by Stock &

Watson, 2002b) or forecast combination methods (proposed

by Rapach, Strauss, & Zhou, 2010). Although Rapach and

Zhou (2013) report evidence in favor of principal component

predictive regressions, the same set of predictors performs

considerably worse than the historical average benchmark

by considering a more recent data sample (see, e.g., Neely,

Rapach, Tu, & Zhou, 2014). This behavior also seems to be

evident for alternative pooling strategies (Baetje &

Menkhoff, 2016). Overall, reported findings confirm results

highlighted by Goyal and Welch (2008) that commonly used

predictors perform unstably over time.

This finding might be a direct consequence of model

uncertainty and/or structural instability surrounding a con-

stantly evolving data‐generating process for stock returns

(Pesaran & Timmermann, 1995). As previously mentioned,

many approaches concerning return predictability only con-

sider a small number of a priori selected indicators as poten-

tial predictor variables, which naturally limits the amount of

available information. Nowadays, numerous variables are

available for model specification, leaving the question unan-

swered of which variables are the most relevant ones for

stock return predictability.

Despite the problem of searching for the most informative

variables, structural instability might lead to changes in the

best model specification (Rapach & Zhou, 2013). Following

Timmermann (2008), Paye and Timmermann (2006), and

Rapach and Wohar (2006), among others, forecasting models

exhibit parameter instability with the consequence that the

outperformance of forecasting models is extremely restricted

to short‐lived periods (see also Goyal & Welch, 2008). Thus,

if the forecasting ability of individual variables varies over

time, it is unclear which variable should be considered in

predictive regression models (i.e., model uncertainty).

This study contributes to the existing literature on stock

return predictability in three ways. First and foremost, we

refer to studies predicting stock market risk premia with

macroeconomic information by exploiting of a large number

of potential predictor variables. Empirical studies commonly

make use of a strong prior on predictability by examining the

forecast performance of individual predictors (e.g., Goyal &

Welch, 2008) or forecasting strategies based on a small set

of variables (e.g., Rapach & Zhou, 2013; Rapach et al.,

2010). However, reported evidence shows that the relation

between the macroeconomic situation and the stock market

is difficult to grasp by relying on low‐dimensional

information sets (Timmermann, 2008). By expanding the

set of predictors, we are able to determine whether the weak

evidence of predictability can be linked to a loss of informa-

tion by using a small set of information to capture time‐vary-

ing dynamics of stock market risk premia. In general, we

follow a variety of recent studies (Stock & Watson, 2002b;

Ludvigson & Ng, 2007, 2009; among others) showing that

the consideration of a large number of predictors might

improve the forecast performance. In this context, our dataset

of potentially meaningful predictor variables consists of 124

US macroeconomic and financial time series and is related

to other studies, such as Ludvigson and Ng (2009).

Second, our research contributes to concerns regarding

model uncertainty and instability issues surrounding the

data‐generating process for stock returns. We address the

complexity of the relation between stock market risk premia

and macroeconomic fundamentals by applying various

forecasting strategies. These are able to consider many poten-

tially important predictor variables simultaneously and,

therefore, are able to limit the impact of model uncertainty

(see Rapach & Zhou, 2013). Additionally, if the degree of

potential model instability is not too dependent across series,

pooling strategies might further enhance the overall forecast

performance (Rapach et al., 2010; Stock & Watson, 2002a).

Explicitly, this study evaluates the predictive performance

of forecasting strategies combining information and strate-

gies that pool individual forecasts (Huang & Lee, 2010;

Rapach & Zhou, 2013). For this purpose, we first apply prin-

cipal component predictive regressions considering a small

number of latent common components as potential predictor

variables (Cakmakli & van Dijk, 2016; Ludvigson & Ng,

2007; Neely et al., 2014). Furthermore, we follow Kelly

and Pruitt (2013, 2015) and make use of a three‐pass regres-

sion filter estimating target relevant factors by particularly

taking into account the relationship between the estimated

factors and the target relevant variables (i.e., stock returns).

Regarding the second set of forecasting strategies, we follow

Rapach et al. (2010) by using forecast combination strategies.

Third, we provide a comprehensive overview of the pre-

dictability of various risk premia. Therefore, we extend the

set of stock market risk premia and provide further insights

into whether forecasting strategies are even beneficial to pre-

dict the size, value, and momentum risk premium. To the best

of our knowledge, other studies either focus on a single risk

premium (in particular, on the market excess return) or on a

single approach. We conduct forecast regressions concerning

short‐run (1–3 months) and long‐run (12–24 months) fore-

cast horizons. In addition to an in‐sample analysis, we evalu-

ate the forecast performance in an OOS setting, starting in

February 1975. Thus forecast accuracy is directly linked to

the end of the recession in the 1970s, which is mentioned

as the climax of the previously reported OOS predictive abil-

ity of commonly used variables (see Goyal & Welch, 2008).

38 BAETJE