You are on page 1of 50

Zurich Open Repository and

Archive
University of Zurich
University Library
Strickhofstrasse 39
CH-8057 Zurich
www.zora.uzh.ch

Year: 2021

Fama–French Factor Timing: The Long-Only Integrated Approach

Leippold, Markus ; Rüegg, Roger

Abstract: There is ample evidence that factor momentum exists in the standard long–short mixed ap-
proach to factor investing. However, the excess returns are put under scrutiny due to the high implemen-
tation costs. We present a novel real-life approach that relies on the long-only integrated approach to
factor investing. Instead of exploiting the potential momentum in factor portfolios, our strategy builds
on the momentum of the optimal factor score weights in the integrated approach, which allows us to
additionally profit from the serial dependence in the factors’ interaction effects. One limitation of short-
term timing strategies is their high turnover. By including the information of the covariance matrix and
minimizing the strategy’s risk to the market portfolio, we can substantially reduce turnover. The result-
ing timing alpha remains significant even after transaction costs in a robust statistical test framework
across the major stock markets.

DOI: https://doi.org/10.1111/eufm.12285

Posted at the Zurich Open Repository and Archive, University of Zurich


ZORA URL: https://doi.org/10.5167/uzh-190472
Journal Article
Published Version

Originally published at:


Leippold, Markus; Rüegg, Roger (2021). Fama–French Factor Timing: The Long-Only Integrated Ap-
proach. European Financial Management, 27(4):666-700.
DOI: https://doi.org/10.1111/eufm.12285
Fama–French factor timing: The long-only integrated approach∗

Markus Leippold† Roger Rueegg‡

September 6, 2020

Abstract
There is ample evidence that factor momentum exists in the standard long–short mixed approach
to factor investing. However, the excess returns are put under scrutiny due to the high imple-
mentation costs. We present a novel real-life approach that relies on the long-only integrated
approach to factor investing. Instead of exploiting the potential momentum in factor portfolios,
our strategy builds on the momentum of the optimal factor score weights in the integrated ap-
proach, which allows us to additionally profit from the serial dependence in the factors’ interaction
effects. One limitation of short-term timing strategies is their high turnover. By including the
information of the covariance matrix and minimizing the strategy’s risk to the market portfolio,
we can substantially reduce turnover. The resulting timing alpha remains significant even after
transaction costs in a robust statistical test framework across the major stock markets.

JEL classification: G1, G11, G14, G17

Keywords: Factor timing, equity style timing, integrated approach, momentum


We thank John Doukas (the editor) and two anonymous referees for their thoughtful comments. We are also
indebted to Fabian Ackermann, Michael Bretscher, Martin Gillholm, Andreas Kappler, Istvan Redl, and Jue Ren for
helpful discussions on the subject.

(Correspoding author) Department of Banking and Finance, University of Zurich, Plattenstrasse 14, 8032 Zurich,
Switzerland, email: markus.leippold@bf.uzh.ch.

Department of Banking and Finance, University of Zurich and Zurich Cantonal Bank, Hardstrasse 201, 8005 Zurich,
Switzerland, email: roger.rueegg@zkb.ch.
1 Introduction

Factor timing gained much popularity recently. Arnott et al. (2018) suggests that industry momentum

may be explained by factor momentum. Other recent works that document factor momentum are

Gupta and Kelly (2019) and Ehsani and Linnainmaa (2019), who find strong autocorrelations among

a broad set of factors that lead to high returns of cross- and time-series momentum strategies. Their

results indicate that the market prices of risk only adjust slowly over time. In retrospect, this factor

persistence represents a source of alpha, which could potentially be exploited with an appropriate

factor-timing strategy.

The recent papers by Novy-Marx and Velikov (2016) and Patton and Weller (2020) show that

there is a gap between the profitability of trading strategies “on paper” and the returns exploited in

practice. An open debate goes back to the work of Shleifer and Vishny (1997). They state that even

the simplest arbitrage strategies are difficult to implement profitably. In particular, state Novy-Marx

and Velikov (2016) and Patton and Weller (2020) that accounting for implementation costs diminishes

the returns of momentum strategies. Since existing literature sees factors as returns of long–short

portfolios, the turnover of the single securities and implementation costs are often neglected. Hence

there is the open question whether recently documented factor momentum returns are robust in a

realistic setting. In particular, we consider the situation in which the investor is restricted to a

long-only portfolio.1

[Figure 1 about here.]

To answer this question, we introduce the integrated approach to long-only factor timing.2 As

shown in Figure 1, the standard approach in literature on factor timing considers a broad set of long–
1
Most institutional investors are long-only investors (Cao et al., 2017; Molk and Partnoy, 2019). Moreover, short-
selling costs are still substantial, leading to an asymmetry in arbitrage (Nagel, 2005; Stambaugh et al., 2015). Lastly,
during market turmoil, regulators tend to impose short-selling bans, e.g., during the 2007-09 crisis (Beber and Pagano,
2013) and also in the wake of the COVID-19 pandemic, as numerous European jurisdictions, including France, Italy,
Spain, Greece, and Belgium have enacted short sale bans in an attempt to stabilize financial markets and maintain
investor confidence. Therefore, we believe that the analysis of a long-only investor is a relevant one.
2
There are (at least) two diverging views on how to build multi-factor portfolios. The first approach, the mixed
approach, is to mix where a portfolio is formed by combining stand-alone factor portfolios. The second approach, the
integrated approach, a portfolio is created by selecting securities that have simultaneously strong exposure to multiple
factors at once. For our long-only factor timing, we rely on the integrated approach. See Bender and Wang (2016) for
a comparison of the two approaches.

1
short strategies and analyzes performance on portfolio-level data.3 We study factor timing using

equity-level data in a pure long-only context. Since the mix of individual long-short portfolios can

only be efficient by chance, our approach offers two advantages over the standard approach. On the

one hand, it allows us to exploit factors’ interaction effects in the factor momentum portfolios.4 On

the other hand, we can extend the analysis of the simple portfolio sorts and include the information

of the covariance matrix of equity returns in the construction of the long-only factor-timing portfolio,

which gives us additional robustness for the alpha, lower turnover, and a significant improvement in

the Sharpe ratio. Moreover, by applying realistic transaction costs on our stock-level analysis, we can

explore most directly the unresolved issue of whether auspicious factor timing is practically feasible

or not.

We study factor timing in a realistic setting, but we also minimize the common pitfalls of out-

of-sample backtesting. In particular, we apply a robust statistical testing framework that adjusts

for multiple tries and respects the time-series nature of the returns by taking into account serial and

cross dependence. We jointly test three different equity universes: for the US, developed markets, and

emerging markets. We also avoid the common pitfalls of look-ahead-biases with survivor-ship-free

universes, and a data lag of two days for international portfolios, which require trading in different

time-zones. Since this paper focuses on whether we can capture factors outside the sample over time,

we do not consider the entire “zoo of factors.” 5 Instead, for our analysis, we rely on the most well-

known factor model specifications that have the most extended out-of-sample history.6 In particular,

we use the classical Fama–French three-factor model presented in Fama and French (1992) and the

Fama–French five-factor model introduced in Fama and French (2015). As the third factor model, we

add the momentum factor of Jegadeesh and Titman (1993) to the five-factor model. These models

are widely accepted in both industry and academic literature.


3
See Arnott et al. (2018), Gupta and Kelly (2019), and Ehsani and Linnainmaa (2019).
4
Our observation that the factor returns play a crucial role in generating performance is an indication for the benefit
of assuming that factors are not only time-varying but also correlated. As Haddad et al. (2020) remark, an optimal
factor timing portfolio is equivalent to a stochastic discount factor, and the maximum conditional Sharpe ratio return
depends on the inverse of the covariance matrix of factor returns. Hence, including the off-diagonal terms of factor
returns’ covariance matrix is crucial to forming such an optimal factor timing strategy.
5
As of February 2019, Harvey and Liu (2019) document over 400 factors published in top journals. However, the
new factors must clear a much higher hurdle, and there is little evidence about their post-publication returns.
6
Also, as shown in Bauer et al. (2010), there arise regional differences as to whether a premium is positive or not.
Since we analyze different regions, we are well-advised in focusing on the established factors that have been tested in
worldwide portfolios.

2
For the US market, our approach allows us to generate significant out-of-sample alphas relative

to the multi-factor benchmark that are robust to multiple-hypothesis testing, ranging between 0.36%

and 0.42% per month, depending on the factor model. As a result, we can almost double the market’s

Sharpe ratio with our long-only factor timing strategies. Our strategy also remains significant for

developed markets for all factor models. The alphas range between 0.29% and 0.47% per month. For

the emerging market, only the three-factor model remains significant, with an alpha of 0.44%.

We remark that the alpha arises solely because of the timing ability since we adjust for the tradi-

tional long–short factor exposures in our regression analysis. We emphasize this finding because our

long–only integrated approach without timing does not survive the adjustment in our robust statis-

tical test framework. Hence, the long–only integrated approach does not dominate the traditional

long–short mixed portfolio construction as it is shown in Leippold and Rüegg (2018). Consequently,

our one-month momentum strategy’s success does not rely on the actual risk premium of the under-

lying traditional long-short factors, and it is a winning strategy when we are uncertain about whether

the underlying factor bears a positive premium.

Since standard literature builds their factor portfolios equal- or value-weighted, they are only

efficient by chance. When we include the covariance matrix’s information with the nonlinear shrinkage

of Ledoit and Wolf (2020), we can further improve the portfolio construction. We show that it even

reduces turnovers, which is the primary concern of Novy-Marx and Velikov (2016) and Patton and

Weller (2020). For the US market and the period from 1963 to 2018, we find that the significance of

the abnormal returns and the increased Sharpe ratios of these optimized portfolios survive transaction

costs of up to 0.15%. For the more recent period from 1998 to 2018, we find significant alphas for

up to 0.10% transaction costs in the US and developed markets. Lastly, in the emerging markets,

the alphas after transaction costs are still positive but lose their significance. Overall, however, our

results are at odds with the recent study by Ilmanen et al. (2019). They only find disappointing

profits from implementable factor timing strategies.

A fundamental limitation of research on market timing is that most of the promoted out-of-

sample backtests fail to generate high risk-adjusted returns in real-time.7 We robustify our results by
7
As highlighted by Bailey et al. (2014) and Bailey and Lopez de Prado (2014), the reason is that the out-of-sample
backtests are prone to severe selection biases.

3
studying different markets and over different periods. Since McLean and Pontiff (2016a) demonstrate

that the post-publication factor return is 58% lower, we pay particular attention to the most recent

period when the factors were already published. To further reduce the selection bias, we apply the

robust alpha test of Leippold and Rüegg (2019) and the robust Sharpe ratio test of Ledoit and Wolf

(2008) that rely on the block-resampling suggested by Lahiri (2003). We then show that the timing

ability also survives the multiple hypothesis adjustments of the state-of-the-art multiple hypothesis

framework of Romano and Wolf (2016).8 The significant advantage of their test framework is that it

allows for the cross-dependence of the test statistics that is crucial due to the high correlations of the

tested timing strategies.9

Our work contributes to the puzzling evidence of cross-sectional momentum as a highly popular

investment style. The first discussion of momentum strategies goes back to Levy (1967) and Jegadesh

and Titman (1993) at the individual stock level. Moskowitz and Grinblatt (1999) later show that

stock momentum arises from industry momentum, whereas a recent paper by Arnott et al. (2018)

suggests that industry momentum may be explained by factor momentum. Other recent works that

document factor momentum are Gupta and Kelly (2019) and Ehsani and Linnainmaa (2019), who

find strong autocorrelations among a broad set of factors that lead to high returns of cross- and

time-series momentum strategies.10

The literature on the timing ability of risk factors concentrates on both fundamental and technical

predictors. An early paper by Kao and Shumaker (1999) presents a multivariate macroeconomic

analysis of timing the value and growth stocks. Asness et al. (2000), Cohen et al. (2003), Asness

et al. (2013), and Arnott et al. (2016) demonstrate that valuation spreads may indicate crowding in

factors with a corresponding negative outlook for the affected factors. By applying support vector

regressions with technical and economic variables, Nalbantov et al. (2006) succeed in timing the

size and value premium in the US.11 Arshanapalli et al. (2007) use fundamental and macroeconomic

factors to show that a multi-style rotation strategy based on the four style combinations with large
8
In recent years, multiple hypothesis tests, which are vital in clinical trials, have gained popularity among the
finance community. See, e.g., Leippold and Lohre (2012a,b, 2014) on multiple hypothesis methods to test for financial
anomalies, or Harvey et al. (2016) analyzing factors strategies.
9
We find an average pairwise correlation of 0.76 among the different factor sets within the same universe.
10
A recent application of factor timing is Dichtl et al. (2019).
11
Examples of technical variables are the value–growth spread or the volatility of the S&P 500, and of the economic
variables are the oil price or the yield curve spread.

4
versus small and growth versus value outperforms the buy-and-hold portfolio. Barroso and Santa-

Clara (2015) find that ex-ante risk adjustments double the Sharpe ratio of momentum. A recent

paper by Hodges et al. (2017) demonstrates that combining different well-known predictors improves

the timing ability. In addition, Moreira and Muir (2017) find that the volatility of the plain factor

has predictive power for the market as well as for value, momentum, profitability, return on equity,

investment, and a low-risk factor.

While the recent literature summarized above does not seem to converge to a conclusive verdict as

the whether factor momentum exists and can be profitably exploited, we contribute to this literature

by revisiting the portfolio construction underlying the factor timing portfolios. In particular, we

depart from the previous literature, which considers each factor individually to construct a long-short

portfolio, we construct our portfolios based on the best-performing factor weight combination of the

most recent period, computed not on a portfolio but a security level. This subtle difference in portfolio

construction leads us to the conclusion that factor timing strategies can be successfully implemented.

In contrast to the successful timing literature, Ilmanen et al. (2019) and Asness (2016), among

others, take the view that the factor timing performance was not convincing historically. They show

that factor timing strategies do not deliver significant performance. Furthermore, with regards to

practical feasibility, substantial criticism has been expressed against different factor timing literature.

For instance, Cederburg et al. (2019) find that the volatility-managed portfolios of Moreira and

Muir (2017) cannot systematically outperform their corresponding unmanaged portfolios. Although

volatility-managed portfolios tend to exhibit significantly positive alphas in spanning regressions, the

corresponding trading strategies are not implementable in real-time. Even worse, the out-of-sample

analysis reveals lower certainty equivalent returns and Sharpe ratios than the unmanaged portfolios.12

We organize this paper as follows. In Section 2, we present our novel approach on how to time the

factors in the long-only integrated approach to style investing that includes the interaction effects.

Section 3 presents the data and preliminary analysis to explore the potential source of alpha from

factor timing. In Section 4, we provide an empirical analysis of our approach, and we present various

performance tests to robustify our results. In particular, we take transaction costs into account, and

we perform a subperiod analysis. Also, we show how the performance can be improved by taking
12
The paper by Moreira and Muir (2017) has been further criticized by Liu et al. (2019) and Guo and Liu (2019).

5
into account the covariance information in the portfolio construction. Section 5 concludes.

2 The long-only integrated approach

The integrated approach to factor investment views equities as a bundle of factors. In contrast, the

standard long–short approach views factors as a bundle of securities. The abolition of this factor-

portfolio approach has a double advantage. Firstly, it offers the possibility to take into account

the interaction effects within factor momentum. While the standard approach considers each factor

individually as a long–short portfolio and constructs a portfolio out of factor portfolios, we invest

in the best factor weight combination of the most recent period, computed on security-level and,

therefore, we include the interaction effects by construction. Instead of investing in individual factor

portfolios that may consist of the same stocks, we invest in the best factor weight combination,

treating each stock as a bundle of factors. Secondly, when we trade in individual securities, a natural

extension is to include the information from the covariance matrix to build efficient portfolios.

As shown in Figure 1, our long-only portfolios start with the normalized factor scores. To obtain

these scores, we proceed as follows. In each period t, we normalize the factor scores of each individual

factor f = 1, ..., F based on its rank and normalize it from zero (worst) to one (best) to obtain a

factor score sf,i,t for each security i = 1, ..., N in the universe. This procedure results in the factor

score matrix St ∈ RF ×N . For example, when we consider momentum as the first factor, the stock

with the highest twelve-to-one-month momentum obtains a factor score of one and the company with

the lowest momentum a factor score of zero, while in between the stocks are distributed at equal

distances. Next, we build the aggregate score at ∈ RN ×1 by weighting the factor scores with possibly

time-dependent weights wt ∈ RF ×1 with the individual factor score matrix S as

at = S⊺t wt . (1)

Finally, to obtain a long-only factor portfolio, we invest in the thirty percent best stocks based

on the aggregate score at . In the first version of our factor timing implementation, we use a value-

6
weighting scheme.13 In the second version, which we lay out in Section 4.4, we improve on value-

weighting by taking explicitly into account the information of the covariance matrix. Figure 1 gives

an overview of our portfolio construction approach and its difference to the commonly used mixed

long–short approach.

Figure 1 clearly illustrates the subtle differences of our approach to the factor timing commonly

defined in the literature, such as Arnott et al. (2018), which are based on a mixed long–short approach.

This approach builds on factor portfolios and aims to predict factor returns, which is then used to

select the factor portfolios. However, our approach is a long-only approach and aims to predict

factor score weights, which we then use to select the stocks. As we show in our empirical analysis in

Appendix A, the factor score weights show considerable persistence and hence are more predictable

than factor returns.

2.1 Naive diversification and Fama MacBeth strategy

The critical element of our long-only approach is the choice of the vector wt of factor score weights

that leads to the overall ranking of the securities. As an initial natural benchmark, we create the

naive diversification strategy (ND) that equally weights the factor scores over time and hence applies

no timing. As shown in Benartzi and Thaler (2001) and DeMiguel et al. (2009), the 1/N heuristic

has a long history in asset allocation and is the natural benchmark strategy to beat for the strategies

presented next that time the weights. In our setting, however, the 1/N strategy is not applied

to stocks, but to factor score weights. Therefore, the naive diversification strategy in this setup

corresponds to the integrated strategies presented and analyzed in Bender and Wang (2016) and

Leippold and Rüegg (2018).

As an additional standard benchmark method, we implement Fama MacBeth (FM) regressions

following Lewellen (2015). For this second benchmark strategy, we replace the aggregated factor score

with the 1-month expected return. The return forecasts are based on the firm’s beta estimates for

the current month and the FM regressions’ average factor slopes. As in Lewellen (2015), we use a

10-year rolling window to estimate FM slopes.


13
Value-weighting is the standard way to build the factor portfolios as done, e.g., in Fama and French (1993a). Also,
it is a model-free approach that still takes the size of the companies into account, and hence reduces the potential
impact of size or illiquidity biases.

7
2.2 Timing strategies

The first one-month momentum strategy (1M) sets the weights wf,t to the sign function of the long–

short factor return of the past month. Hence, when the thirty percent smallest stocks outperformed

the biggest tercile of stocks, we set the factor weight of the size factor to positive one. This procedure

builds our first timing strategy. It corresponds to the traditional approach to factor timing, studied

in other recent papers in a long–short setting. Here, we transfer it to our long-only strategy. It is

important to note that such a timing strategy, by construction, disregards the interaction effects of

the other factors. As we will show later, these interaction effects might play an important role.

To also include the potential benefits from interaction effects among different factors, we create

the one-month momentum strategy that invests in the optimal weight combination of the most recent

month (1M*). Hence, we do not try to exploit momentum in the factor returns as in the previous

literature, but momentum in the optimal factor score weights. However, to do so, we must determine

the optimal weights from the previous month, which will then determine the weights for the following

month.

To solve ex post for the optimal weights vector wt∗ , we first solve recursively and under perfect

foresight for the ex-post optimal combination. Since we cannot solve for the optimal weights in closed

form, we discretize the problem. To this end, we first define a fixed grid of possible weights from

negative one to positive one with a step size of 1/x.

We then take all the possible combinations of these factor weights for the predefined set of factors.

With two factors and x = 3, we obtain weight vectors such as {1, 1} or {0.5, 0.5} that lead to the same

results due to the linearity of the overall score. To remove these redundancies, we further normalize

all long and short exposures within a combination to ±1 and focus on the distinct combinations.

This cleaning results in an aggregate weight of the factors of positive (negative) one if there are only

factors with a positive (negative) weight, or a weight of zero if there are both negative and positive

weights in the combination.

[Figure 2 about here.]

In Figure 2, we plot the factor weights for the different factor models. For example, with F = 2

8
factors and a step size of x = 3, our procedure results in the K = 21 combinations of Table 1 and a

grid of weights that belong to the distinct values illustrated in Figure 2 in the left plot. There are

several special cases. The combination with all weights equal to zero is the market strategy. In this

case, we select all the stocks without an opinion about the future impact of the individual factors.

The combinations with a positive (negative) one and otherwise zeros are the standard long (short)

factor portfolios because they concentrate on a single metric only. We use the standard long–short

portfolios to adjust for the pure factor premias in our robust alpha test in Section 4. In contrast to

Fama and French (1993a), we rebalance the portfolio every month.14

[Table 1 about here.]

The number of combinations grows exponentially with the size of the factor set. For two factors,

such as in the Fama–French three-factor model (FF3), where we try to time the size and value factor,

we obtain 21 combinations. For the four factors in the Fama–French five-factor model (FF5), which

includes the additional factors of investment and profitability, we obtain 1,141 combinations and a

finer grid of distinct weights. For our richest Fama–French five-factor model, which includes the

momentum factor (FFC), we obtain 9,325 combinations. We remark that in each of the above factor

models, we remove the market factor from our timing model since it is not a zero-cost portfolio.

However, when we compute the alpha of the strategies, we include the market factor in the analysis

to control also for market risk.

To compute the ex-post optimal weights over time, we take the average weight of the combinations

that show a return that is equal to or above the qth percentile of the corresponding period. When we

set q = 100, we obtain the particular case where we receive the weights of the best combination only.

However, we expect higher stability in our prediction when we average over the best strategies within

a period. Thus, analogous to the portfolio construction, we concentrate in our standard setting on

the best 30% strategies and set q = 70.15


14
Fama and French (1993a) only rebalance in June of each year.
15
As a robustness check, we also tested for different choices of q. See Section 4.3.

9
2.3 Including the information of the covariance matrix

In a recent paper, Ledoit et al. (2018) argue that the standard methodology to sort stocks according

to their factor scores into quantiles and forming the corresponding long-short portfolio, ignores any

information on the covariance matrix of stock returns. Using the recent dynamic covariance estimator

of Engle et al. (2019), they find that including the information of the covariance matrix in the portfolio

sorts more than doubles the t-values compared to the standard portfolio sorts. Motivated by these

results, we also implement a minimum risk optimization. First, we fill missing returns with the one-

factor model based on the equal-weighted market and under the assumption of a zero alpha. If there

are less than six values to estimate the market beta, we assume that it is equal to one. Next, we

estimate the large dimension covariance matrix with the analytical nonlinear shrinkage estimator of

Ledoit and Wolf (2020) and include the rolling information of the past five years to achieve a true out-

of-sample portfolio.16 To construct the portfolio, say p, we minimize the risk to the value-weighted

market portfolio but restrict the portfolio holdings to hold only the best tercile of stocks as follows





 pn ≥ 0, if an ≥ Qa (q), n = 1, . . . , N,

min (p − wM KT )⊤ Σ(p − wM KT ), s.t. pn = 0, if an < Qa (q), n = 1, . . . , N, (2)
p∈RN ×1 


 p⊤ 1 = 1,

where Qa (q) is the quantile function of the aggregate score distribution evaluated at the q’th percentile

and Σ is the covariance matrix of the stock returns.17

To achieve higher stability in the risk optimization and avoid extreme weights, we resample the

returns based on the one-factor model, re-estimate the covariance matrix, and optimize in total 20

times at each rebalancing to finally invest in the average optimal weights.

2.4 Overview of the strategies

To give a comprehensive overview of the strategies, Table 2 shows the acronyms, definitions, and

characteristics of the different strategies treated in this paper. We start with the benchmark strategies
16
As a robustness check we also applied the linear shrinkage method of Ledoit and Wolf (2003). While we find highly
similar results, the nonlinear shrinkage estimator further reduced turnover, which is an essential advantage in practice.
17
For our empirical analysis we set q = 70.

10
that include the value-weighted market (MKT), the ND strategy, and FM strategies that also take

into account the factor returns. None involves any short-term factor timing, and they are out-of-

sample. The benchmark strategies have, by definition, an abnormal return of zero when we correct

for the risk factors with which they were constructed.18

[Table 2 about here.]

Next, there are the two one-month momentum strategies that apply time-varying factor score

weights wt with (1M*) and without taking the interaction effects into account (1M). While all these

strategies use value weights for their selected stocks, the optimized strategy (1M*Opt) minimizes the

risk to the market-cap weighted universe (MKT) and thereby deviates from the value weights with

the help of the covariance matrix. In the next section, we will investigate whether these three timing

strategies offer high abnormal returns when we correct for the corresponding risk factors. Since they

time the factors, their alpha can be different from zero also when we adjust for the risk factors on

which we apply the timing. A positive alpha in this setting implies that there exists timing ability.

In contrast to the ND strategy, which can only outperform the market if the risk premium of the

factors is positive, the timing strategies are model-free in the sense that they can beat the market

without relying on positive factor returns.

3 Data

For our empirical analysis, we rely on a well-established set of factors. In particular, we focus on the

Fama–French three-factor model (FF3), the Fama–French five-factor model (FF5), and the Fama–

French five-factor model including momentum, as in Carhart (1997) (FFC). Table 3 provides an

overview of their definitions and their acronyms.

[Table 3 about here.]

To guarantee that we trade on information that is known at the rebalancing date, we consider

an additional data lag of two days for the factors in the MSCI universes. This additional lag is
18
It is possible that the alpha is not equal to zero. As Bender and Wang (2016) and Lewellen (2015) argue, the
integrated and FM regression approaches may create additional alpha because of how they aggregate information.

11
of importance because we trade in Asian, European, and US countries where the market closes in

different time zones, which creates further implementation delays.

We consider three datasets. The first dataset is the standard universe with all non-financial

NYSE, AMEX, and NASDAQ stocks from the CRSP and Compustat database. The period ranges

from June 1963 to June 2018. Because of concerns about liquidity, we focus on the companies with

market capitalization above the median of the NYSE. This set of stocks corresponds to the “Big”

portfolio described in Fama and French (1993a).

The second dataset is the MSCI Developed World universe and the third the MSCI Emerging

Markets universe. We have data on its constituents for the most recent twenty years from June 1998

to June 2018. Both indexes invest in the most liquid large and mid-cap stocks of the developed

respectively emerging markets countries. They cover approximately 85% of the free-float adjusted

market capitalization within each universe. Moreover, the MSCI indexes are the most commonly

applied benchmark in delegated asset management. Usually, MSCI Inc. updates its constituents

every quarter. The developed and emerging market countries also change over time, but on a less

frequent basis. Also, for the MSCI universes, we only include non-financials.

[Table 4 about here.]

Table 4 shows the summary statistics for the three universes. We express returns as excess returns

above the one-month Treasury bill rate. Since the US universe starts earlier in 1963 with a small

universe of stocks, the DM universe shows for the market capitalization a higher average market

capitalization with USD million 14,317 compared to 7,235 in the US and 2,928 for the EM. The

US universe holds between 587 and 1,550 stocks, the MSCI developed markets between 1,093 and

3,844, and the MSCI Emerging Markets between 514 and 2,337 stocks over the analyzed period. The

value-weighted returns represented by the market (MKT) show similar Sharpe ratios of 0.31 for the

DM, 0.34 for the EM, and 0.35 for the US universe. As expected, we find the highest volatility for the

emerging markets universe. Also, the best 30% of the stocks always show higher returns compared

to the 30% worst value-weighted portfolio for each of the analyzed factors from Table 3.

12
4 Empirical analysis

For our empirical analysis, we first start with exploring the timing ability of the different strategies.

After studying the risk and return trade-off, we further provide some robustness checks and we show

how the performance can further be improved by taking into account the information of the covariance

matrix.

4.1 Pure timing ability

To quantify the pure timing ability of the strategies, we regress the returns of the strategies on the

returns from a multi-factor model. This multi-factor model includes the market excess return and

the returns from the traditional long-short tercile portfolios based on the SMB, HML, RMW, CMA,

and WML factors. By controlling for these factors, the alpha of the time series regression quantifies

the timing ability of the strategy.

[Table 5 about here.]

Table 5 shows the multi-factor alpha coefficient together with the single hypothesis and multiple

hypothesis adjusted p-value. We show in Panel A the long-term analysis in the US and Panel B the

most recent 20 years for the US, developed, and emerging markets. To compute the p-values, we

apply the block-resampling robust alpha test of Leippold and Rüegg (2019).19 Further, we adjust

the single hypothesis p-value with the multiple hypothesis framework of Romano and Wolf (2016),

when we jointly test the nine strategies that arise per universe: one-month momentum including the

interaction effects (1M*), one-month excluding the interaction effects (1M), and naive diversification

(ND).

As outlined by Bailey et al. (2014), it is crucial to control for the number of tries for our out-

of-sample tests. There are two common error rates to correct for: the false discovery rate (FDR)
19
The block-resampling method has a higher statistical power than other standard tests for financial time series (see
Lahiri (2003) or Leippold and Rüegg (2019)). For the block size, we employ a value of five since most of the optimal
block sizes lie between one and five with the method of Politis and White (2004) and Patton et al. (2009), while five
results in the highest and thus most conservative p-values.

13
and the more conservative family-wise error rate (FWER).20 The FWER approach of Romano and

Wolf (2005a,b) offers first the advantage of taking into account the cross-dependence structure of the

strategies. Since we find high correlations between different factor sets within the same universe of

0.63 to 0.90 for the 1M* strategies, it is crucial to correct for cross-dependence. Also, the FWER is

more suitable for a lower number of hypotheses. With three factor sets and three universes, we have

in total nine tries per universe, and hence obtain a low number of hypotheses. For these reasons, we

apply the multiple hypothesis adjustments of Romano and Wolf (2016) to the block-bootstrapped

t-statistics of the robust alpha test of Leippold and Rüegg (2019).

Table 5 shows monthly average alphas that control for the pure factor risks of 0.38% for the

one-month momentum strategy including the interaction effects (1M*), 0.31% for the one-month

momentum strategy excluding the interaction effects (1M), 0.15% for the naive diversification strategy,

and 0.00% for the Fama MacBeth strategy. We highlight significant alphas for the multiple-tries

adjusted p-values at the 95% confidence level in bold. In the long-term analysis in the US, the 1M*

strategy, including interaction effects, offers significant alphas also when we adjust for multiple tries.

Also in the developed markets, we observe significant alphas for each factor set over the most recent 20

years, while they are weakly significant for FF3 as well as FFC in the US and significant for FF3 in the

emerging markets and FF5 in the US. The 1M strategy shows the same return patterns, however only

in the long-term analysis, the developed markets, and the FF5 model in the US, do we find significant

alphas over the most recent 20 years. As expected, the FM and naive diversification strategies lack

significance for the period from 1998 to 2018, since the alpha is controlled by the multi-factor model.

For the long-term analysis in the US, the naive diversification strategy delivers positive alphas that

must be due to the interaction effects of the bottom–up integrated approach since we correct for the

factor returns.21 Also, the FM alpha of the FFC factor set is weakly significant in the US for the

long-term analysis, which confirms the findings of Lewellen (2015), who uses a broader set of 15 firm

characteristics. To conclude, we attribute a significant timing ability to the 1M* strategy, which is

positive in every analysis for the three different universes and three separate factor sets.
20
See, e.g., Romano et al. (2007) or Bajgrowicz and Scaillet (2012) for a discussion of the differences between the
two metrics.
21
Bender and Wang (2016) find that the interaction effects in the integrated approach can have a significant impact.
Leippold and Rüegg (2018) find that the differences are not significant from a Sharpe ratio point of view, as implied
by the results in Panel B

14
4.2 Return to risk analysis

We next analyze the differences in the return to risk between the 1M*, 1M, ND, and FM strategy.

The robust Sharpe ratio test of Ledoit and Wolf (2008) in Table 6 shows the annualized Sharpe ratio

differences from the market portfolio together with the multiple hypothesis adjusted p-values. We find

that the annual Sharpe ratio of the 1M* strategy is, on average, 0.25 higher than that of the market,

while the 1M strategy, the naive diversification strategy, and FM strategy show an improvement of

0.21, 0.20, and 0.02. As shown in Table 4, we find Sharpe ratios of 0.31, 0.35, and 0.34 for the

US, developed, respectively emerging markets. Hence, the average increase of 0.27 for the long-term

analysis in the US corresponds to a Sharpe ratio of 0.58 for the 1M* strategy.

[Table 6 about here.]

The 1M* strategy also shows, in terms of the return to risk ratio, the highest improvement: on

average, by 0.25. The improvement is significant for the long-term analysis in the US, and the FF3

model in the developed and emerging markets. The improvement in the Sharpe ratio for the 1M

strategy survives the multiple hypothesis adjustment only in the long-term analysis in the US and

the most recent 20 years in the EM FF3 factor set, while the improvement for the ND strategy

lacks significance for all but the US long-term analysis in the FF5 and FFC factor sets. The second

benchmark strategy based on the FM regressions delivers only a marginal average improvement of

0.02. We conclude that the 1M* strategy also dominates the return to risk comparison. After

robust hypothesis tests, we find significant improvements for most of the Fama–French factor sets

and universes.

4.3 Turnover and transaction costs

The average two-way turnover of the 1M* strategy is 137%. Thus, transaction costs are a crucial

factor when we implement the one-month strategy. To verify the stability of our results to the high

turnover of the timing strategies, we now apply trading costs of 5, 10, and 15 basis points (Bps) to

our analysis.22 According to the study of Frazzini et al. (2018), the choice of 10 Bps seems to be the
22
Since we only regard the highest capitalized stocks within each market and constrain short-selling, we assume that
transaction costs are similar across the universe.

15
most representative one for a large institutional money manager in developed markets. Table 7 shows

the monthly multi-factor alpha after transaction costs in Panel A, and the annual improvement in

the Sharpe ratio in Panel B for the 1M* strategy. We jointly test for the three strategies for the

long-term analysis in the US and the nine strategies for the most recent 20 years and calculate the

robust multiple hypothesis adjusted p-value in italics.

[Table 7 about here.]

We find that the 1M* strategy offers significant alphas in any of the factor sets for up to 0.10%

transaction costs in the US long-term analysis from 1963 to 2018. Also, the improvement in the

Sharpe ratio survives transaction costs of up to 0.05% for the same US universe. For the more recent

period from 1998 to 2018, we find that the alphas and return to risk improvements lack significance

when we deduct transaction costs. In our analysis, we find that the turnover increases with the

number of factors. For the FF3 model, we find an average two-sided turnover of 127%, and for the

FF5 and FFC, 144%. This explains the faster decreasing abnormal returns for the larger factor sets.

In another robustness check, we compute the multi-factor alphas and Sharpe ratio improvements

when we invest in the best strategy (q = 100) and the average 20% best strategies (q = 80) weights

instead of our standard choice of q = 70. We find that the alpha stays constant at 0.38% monthly

for both choices. Regarding the Sharpe ratio and the significance, we observe a higher stability when

we average over a part of the best strategies. Due to the high parameter uncertainty that behavior

comes as no surprise. For example, the improvement in the Sharpe ratio decreases to 0.21 for q = 100

when we only invest in the best strategy but stays at robust 0.25 for q = 80 and q = 70. As outlined

in Cattaneo et al. (2018), the standard approach to build tercile, quintile, or decile portfolios to

compute the factor return is ubiquitous. Their data-driven procedure shows that the optimal number

of portfolios may be much larger. Since we use tercile portfolios when we report our results, we also

test for decile and ventile portfolios, for which we invest in the best 10% and 5% of the universe. In

line with their results, we find larger alphas the more we concentrate on the best stocks only with

an average alpha of 0.49% and 0.56%. However, they show higher p-values and turnovers, which

decreases their attractiveness in our realistic long-only approach.

16
4.4 Including the information of the covariance matrix

So far, we have focused on a model-free portfolio construction with value-weights. We next present

the results for our optimized portfolio construction that we introduced in Section 2.3. We conduct

the same analysis with transaction costs as in the previous section, but this time, we include the

information of the covariance matrix and show the results for the 1M*Opt strategy in Table 8.

[Table 8 about here.]

We find that the average alpha stays similar to the 1M* strategy with an abnormal return of 0.36%

per month without transaction costs. However, there are two main differences when we optimize the

portfolios. First, we find an improvement of the Sharpe ratio by 0.29, which is above the 0.25 without

the information of the covariance matrix. Thus, we conclude that the minimum risk optimization

towards the market portfolio increases the stability of the strategy. Second, we see lower turnovers of

116% (FF3), 129% (FF5), and 133% (FFC), which lead to lower transaction costs. Both the higher

stability and the lower transaction costs lead to a greater robustness of the results. For example,

the FF3 timing alpha is weakly significant for each universe and analysis for up to 0.05% transaction

costs, while for the US long-term analysis the abnormal return is significant at the 95% confidence

level also for 0.15% and both the timing ability in Panel A and the Sharpe ratio improvement in Panel

B for all three factor sets. Also, the Sharpe ratio differences for the short-term analysis start to show

weakly significant improvements for up to 0.05% transaction costs. For comparison, the analysis with

value-weights in Table 7 lacks significance for most of the Sharpe ratio differences in the most recent

20 years even without transaction costs.23

Table 8 also reveals that our factor strategy does not perform equally well in emerging markets.

This finding might be due to the following reasons. First, we note that transaction costs play a

crucial role. Absent of transaction costs, the EM strategies provide some significant excess returns

but at a slightly lower level than the returns for DM. When adding transaction costs, the excess

returns become insignificant. Compared to DM, the reason for this decrease is also caused by higher
23
In tests not reported, we also compute the sector-adjusted performance of our approach. We find that the alpha is
similar with the reported alphas of Table 8 with an average alpha of 0.30% and a Sharpe ratio increase of 0.28 for the
linear shrinkage covariance estimator.

17
turnover. For instance, under the FF3 model, DM has a turnover of 104% compared to a turnover

of 117% for EM. Second, compared to DM, EM is a heterogeneous set of different countries from

different continents, i.e., it includes, e.g., Hong Kong, Brazil, South Africa, Pakistan, Saudi Arabia,

and is less concentrated than DM. In particular, almost 69% of DM is represented by the US market.

Substantially different economic and monetary policies may characterize the countries in the EM

sample. Indeed, the average return correlation for EM countries (0.52) is significantly lower than the

one for DM countries (0.64). Hence, there is room for further improvement of the underlying factor

model for these countries. We leave this as an avenue for future research.

In Figure 3, we show the cumulative logarithmic excess returns for the optimized long-only strate-

gies in the US on the top graph and the FF3 excess return to the market including transaction costs

in the bottom graph.

[Figure 3 about here.]

As we can see in the bottom graph, the excess return of the 1M*Opt strategies compared to the

market portfolio increases steadily except for the period before the burst of the dot-com bubble. Since

this was the period where we also observed high excess returns of the market portfolio, it is a period

where lower returns than the market harmed the least from a return to risk perspective. On the

other hand, the excess return to the market recovers significantly in the dot-com crash and flattens

the performance of the long-only strategy in the top graph. However, when we concentrate on other

crashes, such as Black Monday in 1987 or the financial crisis, we can not observe the same flattening

of the one-month momentum strategy. Thus, we expect different return patterns, depending on the

trigger of the crisis.

[Table 9 about here.]

As the recent literature on factor timing suggests, there is ample evidence that factor momentum

exists. However, there are concerns if the traditional implementation is profitable after real-world

implementation costs. In Table 9, we report the alphas and Sharpe ratios of the 1M strategy, which

corresponds to the traditional implementation of factor timing within the integrated approach. Fur-

thermore, as we have discussed in the previous section, the results in Table 10 of the Appendix A

18
suggest that the momentum in factor returns is weaker than in the optimal weights. Indeed, we find

that the performance of the 1M strategy is much weaker compared to the 1M* and, in particular, to

the 1M*Opt strategy. As we conjectured in our preliminary analysis, the difference for the developed

market is striking. Already at a transaction cost of 5 Bps, none of the alphas is persistent.24 For

the US long-term analysis, the alphas of the 1M*Opt are significant at a 5% level for all levels of

transaction costs. For the 1M strategy, none of the strategies is significant at 15 Bps and at 10 Bps

the significance level of the FF3 model drops to 10%.

4.5 Subperiods analysis

To assess the robustness of our factor timing strategy, we can also analyze different subperiods. We

do so by first analyzing the potential effect of a publication bias, as noted by McLean and Pontiff

(2016b). One could conjecture that the success of our factor timing strategy decreases after the

publication date of the factor models. In what follows, we only focus on the 1M*Opt for the Fama–

French three factor model. As Figure 4 indicates, the monthly alpha of this strategy is 0.33% with a

(single hypothesis) robust p-value of 0.00 over the whole period. When we split the sample into the

pre- and post-publication period, the alphas hardly differ and the p-values still indicate significance

at the 1% level. Therefore, our strategy remains robust also after the publication of the Fama and

French (1993b) paper.

[Figure 4 about here.]

As a second subperiod analysis, we use the nonparametric method of Matteson and James (2014)

to identify potential change points in the time series of the factors. This method is very general

and relies on a minimal set of assumptions for the underlying observations. It allows us to estimate

the number of change points and their location. We apply this method to the equal weighted factor

portfolio. In Panel c) of Figure 4, we add the change points which correspond basically to the oil

crisis in the 1970ies and the dot-com bubble in the 2000s. Interestingly, the algorithm does not find

a change point for the latest financial crises. Then, we ask whether these change points also have
24
As previously mentioned, a transaction cost of 10 Bps seems to be more appropriate as noted by Frazzini et al.
(2018).

19
material effects on the factor timing strategies. Clearly, before the oil crisis, the 1M*Opt strategy

was highly successful with a significant alpha of 0.52%. The oil crisis caused a negative alpha, but

with a p-value of 0.61 not significantly different from zero over this relatively short period. During

the dot-com bubble, this strategy was also not capable of generating a timing alpha different from

zero. However, after 2003, the alpha increased again to 0.27% with a p-value of 0.00, indicating a

highly significant alpha. Therefore, we find that the excess performance of our long-only factor timing

strategy is a highly robust phenomena across different time periods.

[Figure 5 about here.]

In Figure 5, we perform the same analysis for the 1M*Opt, but this time with transaction costs.

For the transaction costs, we assume 10 Bps, following Frazzini et al. (2018). Clearly, the alphas

suffer from these costs. However, as we observe in Panels a) and b), we still find significant alphas for

the whole period as well as the pre- and post-publication periods. Furthermore, the post-publication

monthly alpha is still at 0.20% and significant at the 5% level. As Panel c) indicates, the alpha in

the most recent period since the last breakpoint is still weakly significant, at 0.15% with a p-value of

0.06. Hence, even after transaction costs, we find significant alphas. Only during the oil crisis do we

find a negative alpha which, however, is insignificant.

4.6 How does traditional factor timing perform?

Lastly, we can ask how traditional factor timing, represented by the 1M strategy, performs relative

to the results from our 1M*Opt strategy in terms of their alpha. Figure 6 plots the alpha of the

1M strategy for the whole period and for different subperiods. Compared to the 1M*Opt strategy

in Panel b) of Figure 4, the FF3 alpha drops from the pre- to the post-publication period by more

than 50%. After June 1992, the monthly alpha is only 0.18% and the p-value rises to 5%. Clearly,

traditional factor timing has lost some of its sparkle in the most recent period. This impression is

confirmed when we look at Panel c) of Figure 6. While the 1M*Opt earns an alpha of 0.27% with a

p-value of 0.00 since February 2003, the 1M strategy earns an alpha of 0.20% with a p-value of 0.02.

Just before that period, the 1M strategy suffered during the period from January 2000 to January

2003, with a negative monthly alpha of −1.19%. For the same period, the 1M*Opt strategy was able

20
to generate an (insignificant) alpha of 0.16%. Hence, the 1M*Opt strategy offers a much more stable

alpha for different periods of times.

[Figure 6 about here.]

In Figure 7, we add transaction costs, again 10 Bps, to our analysis.25 While the 1M*Opt strategy

offers in both the pre- and the post-publication period a positive alpha of 0.24% and 0.20% at a 5%

significance level, the 1M strategy delivers a significant alpha only in the pre-publication period.

After June 1992, the alpha shrinks to an insignificant 0.05%. In Panel c) of Figure 7, we find that

the alpha in the most recent period after the latest breakpoint completely disappears. Although the

alpha remains positive, it only amounts to 0.07%. Its p-value of 0.38 indicates that this value is not

different from zero. In contrast, even after transaction costs, the 1M*Opt strategy offers a alpha of

0.15% with a p-value of 0.06.

[Figure 7 about here.]

The above subperiod analysis shows that traditional factor timing under reasonable transaction

costs might not be capable of delivering significant alpha, particularly in the most recent period.

However, when we use the 1M*Opt strategy, we can provide a stable and significant alpha over

different subperiods.

5 Conclusion

As highlighted by Novy-Marx and Velikov (2016) and Patton and Weller (2020), one limitation of

momentum strategies is their high turnover such that the trading strategy generates high returns “on

paper” but zero returns in practice. To analyze whether this is also the case for the recently emerged

factor momentum, we have presented a novel framework for factor timing that relies on the realistic

long-only integrated approach to style investing. We find that our one-month momentum strategy

generates robust abnormal returns of 0.38% per month, i.e. 4.61% per annum, with an improvement

of the Sharpe ratio by 0.25 compared to the market portfolio.


25
We note that the turnover for 1M and 1M*Opt is practically the same.

21
Our findings support Arnott et al. (2018), Gupta and Kelly (2019), and Ehsani and Linnainmaa

(2019) in that we provide evidence for the high returns of momentum strategies also in a realistic

long-only approach. We depart from the previous literature by exploiting the momentum in factor

score weights on security-level, and not in the factor returns on portfolio level. By doing so, we benefit

from the cross-sectional dependencies of the factor scores, which helps us to further strengthen the

significance of our results.

In addition, we take advantage of our bottom-up approach when we include the information of the

covariance matrix and minimize the risk to the market portfolio. Not only can we reduce turnover,

but we also find significant alphas even after transaction costs. At the same time, the Sharpe ratio

improvement increases to 0.29.

Our results emphasize that the Fama-French factors only adjust slowly over time. Timing ability is

present among the Fama-French three-factor, five-factor, and six-factor model including momentum,

as well as among the US, developed, and emerging markets. Moreover, the additional performance

from factor timing is statistically significant and robust to the state-of-the-art block resampling

method of Ledoit and Wolf (2008) and Leippold and Rüegg (2019), even when we adjust for multiple

tries. However, in contrast to the existing literature that concentrates on long–short strategies and a

broad set of factors, we find in our long-only setting that considers also the developed and emerging

markets the most significant results for the Fama–French three-factor model, where we time only the

two most established factors: size (SMB) and value (HML).

While the success of traditional factor strategies depends on the positive risk premium of the

factors, the success of our factor momentum strategy depends on the serial dependence of the optimal

factor weights, including their interaction. Our paper has shown that these optimal weights are

persistent and that this persistence translates into a significant alpha in a long-only setting. Moreover,

when we additionally include the information of the covariance matrix for our portfolio sorts, the

resulting alphas remain significant even after transaction costs. Hence, in contrast to the standard

literature, we present a real-live long-only approach that survives the high implementation costs of

factor momentum strategies. At the same time, traditional factor timing strategies do not provide

any significant alphas at a realistic level of transaction costs for the period from 1998 to 2018, neither

in the US nor in the developed and emerging markets.

22
References

Arnott, Robert D., Noah Beck, Vitali Kalesnik, and John West, 2016, How can "smart beta" go

horribly wrong?, Technical report, Research Affiliates.

Arnott, Robert D., Mark Clements, Vitali Kalesnik, and Juhani T. Linnainmaa, 2018, Factor mo-

mentum, Available at SSRN 3116974.

Arshanapalli, Bala G., Lorne N. Switzer, and Karim Panju, 2007, Equity-style timing: A multi-style

rotation model for the Russell large-cap and small-cap growth and value style indexes, Journal of

Asset Management 8, 9–23.

Asness, Clifford S., 2016, The Siren song of factor timing aka “smart beta timing” aka “style timing”,

Journal of Portfolio Management 42, 1–6.

Asness, Clifford S., and Andrea Frazzini, 2013, The devil in HML’s details, Journal of Portfolio

Management 39, 49–68.

Asness, Clifford S., Jacques A. Friedman, Robert J. Krail, and John M. Liew, 2000, Style timing:

Value versus growth, Journal of Portfolio Management 26, 50–60.

Asness, Clifford S., Tobias J. Moskowitz, and Lasse Heje Pedersen, 2013, Value and momentum

everywhere, Journal of Finance 68, 929–985.

Bailey, David, and Marcos Lopez de Prado, 2014, The deflated sharpe ratio: Correcting for selection

bias, backtest overfitting and non-normality, Journal of Portfolio Managment 40, 94–107.

Bailey, David H., Jonathan M. Borwein, Marcos L. de Prado, and Qiji J. Zhu, 2014, Pseudomathemat-

ics and financial charlatanism: The effects of backtest over fitting on out-of-sample performance,

Notices of the American Mathematical Society 61, 458–471.

Bajgrowicz, Pierre, and Olivier Scaillet, 2012, Technical trading revisited: False discoveries, persis-

tence tests, and transaction costs, Journal of Financial Economics 106, 473–491.

Banz, Rolf W., 1981, The relationship between return and market value of common stocks, Journal

of Financial Economics 9, 3–18.

23
Barroso, Pedro, and Pedro Santa-Clara, 2015, Momentum has its moments, Journal of Financial

Economics 116, 111–120.

Bauer, Rob, Mathijs Cosemans, and Peter C. Schotman, 2010, Conditional asset pricing and stock

market anomalies in europe, European Financial Management 16, 165–190.

Beber, Alessandro, and Marco Pagano, 2013, Short-selling bans around the world: Evidence from the

2007–09 crisis, The Journal of Finance 68, 343–381.

Benartzi, Shlomo, and Richard H. Thaler, 2001, Naive diversification strategies in defined contribution

saving plans, American Economic Review 91, 79–98.

Bender, Jennifer, and Taie Wang, 2016, Can the whole be more than the sum of the parts? Bottom–up

versus top–down multifactor portfolio construction, Journal of Portfolio Management 42, 39–50.

Cao, Jie, Bing Han, and Qinghai Wang, 2017, Institutional investment constraints and stock prices,

Journal of Financial and Quantitative Analysis 52, 465–489.

Carhart, Mark M., 1997, On persistence in mutual fund performance, Journal of Finance 52, 57–82.

Cattaneo, Matias D., Richard K. Crump, Max H. Farrell, and Ernst Schaumburg, 2018,

Characteristic-sorted portfolios: Estimation and inference, Review of Economics and Statistics

1–47.

Cederburg, Scott, Michael S O’Doherty, Feifei Wang, and Xuemin Sterling Yan, 2019, On the per-

formance of volatility-managed portfolios, Journal of Financial Economics, Forthcoming .

Cohen, Randolph B., Christopher Polk, and Tuomo Vuolteenaho, 2003, The value spread, Journal of

Finance 58, 609–641.

Cooper, Michael J., Huseyin Gulen, and Michael J. Schill, 2008, Asset growth and the cross-section

of stock returns, Journal of Finance 63, 1609–1651.

DeMiguel, Victor, Lorenzo Garlappi, and Raman Uppal, 2009, Optimal versus naive diversification:

How inefficient is the 1/N portfolio strategy?, Review of Financial Studies 22, 1915–1953.

24
Dichtl, Hubert, Wolfgang Drobetz, Harald Lohre, Carsten Rother, and Patrick Vosskamp, 2019,

Optimal timing and tilting of equity factors, Financial Analyst Journal 75, 84–102.

Ehsani, Sina, and Juhani T. Linnainmaa, 2019, Factor momentum and the momentum factor, Tech-

nical Report 6508, National Bureau of Economic Research.

Engle, Robert F., Olivier Ledoit, and Michael Wolf, 2019, Large dynamic covariance matrices, Journal

of Business & Economic Statistics 37, 363–375.

Fama, Eugene F., and Kenneth R. French, 1992, The cross-section of expected stock returns, Journal

of Finance 47, 427–465.

Fama, Eugene F., and Kenneth R. French, 1993a, Common risk factors in the returns on stocks and

bonds, Journal of Financial Economics 33, 3–56.

Fama, Eugene F., and Kenneth R. French, 1993b, Common risk factors in the returns on stocks and

bonds, Journal of Financial Economics 33, 3–56.

Fama, Eugene F., and Kenneth R. French, 2015, A five-factor asset pricing model, Journal of Finan-

cial Economics 116, 1–22.

Frazzini, Andrea, Ronen Israel, and Tobias J Moskowitz, 2018, Trading costs, Available at SSRN

3229719.

Guo, Shuxin, and Qiang Liu, 2019, Volatility-managed portfolios: True market-timing with a false

theory?, Available at SSRN 3385377 .

Gupta, Tarun, and Bryan T. Kelly, 2019, Factor momentum everywhere, Journal of Portfolio Man-

agement 1–24.

Haddad, Valentin, Serhiy Kozak, and Shrihari Santosh, 2020, Factor timing, The Review of Financial

Studies 33, 1980–2018.

Harvey, Campbell R, and Yan Liu, 2019, A census of the factor zoo, Available at SSRN 3341728 .

Harvey, Campbell R., Yan Liu, and Heqing Zhu, 2016, ...and the cross-section of expected returns,

Review of Financial Studies 29, 5–68.

25
Hodges, Philip, Ked Hogan, Justin R. Peterson, and Andrew Ang, 2017, Factor timing with cross-

sectional and time-series predictors, Journal of Portfolio Management 44, 30–43.

Ilmanen, Antti, Ronen Israel, Tobias J Moskowitz, Ashwin K Thapar, and Franklin Wang, 2019,

Factor premia and factor timing: A century of evidence, Available at SSRN 3400998 .

Jegadeesh, Narasimhan, and Sheridan Titman, 1993, Returns to buying winners and selling losers:

Implications for stock market efficiency, Journal of Finance 48, 65–91.

Jegadesh, Narasimhan, 1990, Evidence of predictable behavior of security returns, Journal of Finance

45, 881–898.

Jegadesh, Narasimhan, and Sheridan Titman, 1993, Returns to buying winners and selling losers:

Implications for stock market efficiency, Journal of Finance 48, 65–91.

Kao, Duen-Li, and Robert D. Shumaker, 1999, Equity style timing, Financial Analysts Journal 55,

37–48.

Lahiri, Soumendra N., 2003, Resampling methods for dependent data (Springer-Verlag).

Ledoit, Olivier, and Michael Wolf, 2003, Improved estimation of the covariance matrix of stock returns

with an application to portfolio selection, Journal of Empirical Finance 10, 603–621.

Ledoit, Olivier, and Michael Wolf, 2008, Robust performance hypothesis testing with the Sharpe

ratio, Journal of Empirical Finance 15, 850–859.

Ledoit, Olivier, and Michael Wolf, 2020, Analytical nonlinear shrinkage of large-dimensional covari-

ance matrices, Annals of Statistics forthcoming.

Ledoit, Olivier, Michael Wolf, and Zhao Zhao, 2018, Efficient sorting: A more powerful test for

cross-sectional anomalies, Journal of Financial Econometrics 17, 645–686.

Leippold, Markus, and Harald Lohre, 2012a, Data snooping and the global accrual anomaly, Applied

Financial Economics 22, 509–535.

Leippold, Markus, and Harald Lohre, 2012b, International price and earnings momentum, European

Journal of Finance 18, 535–573.

26
Leippold, Markus, and Harald Lohre, 2014, The dispersion effect in international stock returns,

Journal of Empirical Finance 29, 331 – 342.

Leippold, Markus, and Roger Rüegg, 2018, The mixed vs the integrated approach to style investing:

Much ado about nothing?, European Financial Management 24, 829–855.

Leippold, Markus, and Roger Rüegg, 2019, How rational and competitive is the market for mutual

funds?, Review of Finance forthcoming.

Levy, Robert A., 1967, Relative strength as a criterion for investment selection, Journal of Finance

22, 595–610.

Lewellen, Jonathan, 2015, The cross-section of expected stock returns, Critical Finance Review 4,

1–44.

Liu, Fang, Xiaoxiao Tang, and Guofu Zhou, 2019, Volatility-managed portfolio: Does it really work?,

Journal of Portfolio Management forthcoming.

Matteson, David S, and Nicholas A James, 2014, A nonparametric approach for multiple change point

analysis of multivariate data, Journal of the American Statistical Association 109, 334–345.

McLean, David R., and Jeffrey Pontiff, 2016a, Does academic research destroy stock return pre-

dictability?, Journal of Finance 71, 5–32.

McLean, R David, and Jeffrey Pontiff, 2016b, Does academic research destroy stock return pre-

dictability?, The Journal of Finance 71, 5–32.

Molk, Peter, and Frank Partnoy, 2019, Institutional investors as short sellers, Boston University Law

Review 99, 837–871.

Moreira, Alan, and Tyler Muir, 2017, Volatility-managed portfolios, The Journal of Finance 72,

1611–1644.

Moskowitz, Tobias J., and Mark Grinblatt, 1999, Do industries explain momentum?, Journal of

Finance 54, 1249–1290.

27
Nagel, Stefan, 2005, Short sales, institutional investors and the cross-section of stock returns, Journal

of Financial Economics 78, 277–309.

Nalbantov, Georgi, Rob Bauer, and Ida Sprinkhuizen-Kuyper, 2006, Equity style timing using support

vector regressions, Applied Financial Economics 16, 1095–1111.

Novy-Marx, Robert, 2013, The other side of value: The gross profitability premium, Journal of

Financial Economics 108, 1–28.

Novy-Marx, Robert, and Mihail Velikov, 2016, A taxonomy of anomalies and their trading costs, The

Review of Financial Studies 29, 104–147.

Patton, Andrew, Dimitris N. Politis, and Halbert White, 2009, Correction to “Automatic block-

length selection for the dependent bootstrap” by D. Politis and H. White, Econometric Reviews

28, 372–375.

Patton, Andrew J., and Brian M. Weller, 2020, What you see is not what you get: The costs of

trading market anomalies, Journal of Financial Economics forthcoming.

Politis, Dimitris N., and Halbert White, 2004, Automatic block-length aelection for the dependent

bootstrap, Econometric Reviews 23, 53–70.

Romano, Joseph P., and Michael Wolf, 2005a, Exact and approximate stepdown methods for multiple

hypothesis testing, Journal of the American Statistical Association 100, 94–108.

Romano, Joseph P., and Michael Wolf, 2005b, Stepwise multiple testing as formalized data snooping,

Econometrica 73, 1237–1282.

Romano, Joseph P., and Michael Wolf, 2016, Efficient computation of adjusted p-values for

resampling-based stepdown multiple testing, Statistics & Probability Letters 113, 38–40.

Romano, Joseph P., Michael Wolf, et al., 2007, Control of generalized error rates in multiple testing,

The Annals of Statistics 35, 1378–1408.

Rosenberg, Barr, Kenneth Reid, and Ronald Lanstein, 1985, Persuasive evidence of market ineffi-

ciency, Journal of Portfolio Management 11, 9–16.

28
Shleifer, Andrei, and Robert W. Vishny, 1997, The limits of arbitrage, The Journal of Finance 52,

35–55.

Stambaugh, Robert F, Jianfeng Yu, and Yu Yuan, 2015, Arbitrage asymmetry and the idiosyncratic

volatility puzzle, The Journal of Finance 70, 1903–1948.

29
A Persistence in optimal factor score weights

The novelty of our factor timing approach lies in the way we construct the timing portfolio. Instead of

exploiting the momentum in the factor returns, we exploit the momentum in the optimal factor score

weights. Indeed, this subtle difference to previous endeavors to factor timing is key in generating

the excess returns of our strategy. The reason lies in the stronger persistence of factor score weights,

compared to factor returns.

To further clarify this point and to obtain an idea of what might be the source of alpha, we

perform the following preliminary analysis for the optimal factor score weights and factor returns.

For the three universes and three factor sets, we discretize the factor score weights and calculate the

different strategies over time. Next, we compute the optimal weight combination for each month

based on the thirty percent best strategies as we presented in Section 2. To explore whether a factor

timing strategy has the potential of being successful, we analyze the persistence of the optimal factor

weights. Since the optimal weights behave like bounded random variables, we first transform the

weights to normality. Then, we run a simple vector autoregressive model with one lag on different

data sets and using different factor model specifications. In particular, we assume that the optimal

weights vector wt∗ follows a VAR(1) model of the form

wt∗ = A1 wt−1

+ ǫt .

Then, we do the same analysis for factor returns, which gives us some indication about persistence

in factor returns that eventually drives the performance of the 1M strategy.

[Table 10 about here.]

In Table 10, we summarize the results in terms of significance of the entries in the matrices A∗1

and A1 for factor scores and returns. We find that for all three markets, the FF3 model exhibits

a high degree of persistence in the diagonal elements of the matrix A∗1 , which might be used to

exploit momentum in the scores and construct profitable investment strategies. For both the US and

developed markets, both diagonal elements a∗1,ii are significant at the 1% level. Also for the emerging

30
markets, the SMB factor of the FF3 model is highly significant at the 1% level. However, the HML

factor is only significant at the 10% level, with a p-value of 7%. We remark that all off-diagonal

elements a∗1,ij,i6=j for the FF3 model are insignificant. This is good news, since with the momentum

strategy, only the persistence in the diagonal elements are important

For factor returns, the FF3 model seems to perform equally well for the US market, but does not

exhibit any significant persistence in the diagonal elements a1,ii for developed markets. For emerging

markets, one diagonal and one off-diagonal element are also significant. However, for the momentum

strategy based on factor returns, this persistence cannot be exploited. If we build our momentum

strategy on factor scores, the significance of this off-diagonal element becomes important again, since

our strategy allows us to pick up interaction effects between factors. Consequently, for FF3 we expect

the 1M* strategy to outperform the 1M strategy in emerging markets. As we will see, Table 5 confirms

our conjecture.

For the other models, we find a similar pattern. However, the persistence in the diagonal factors

does not seem to be as strong as for the FF3 model. Only a few of the off-diagonal elements are

significant. For the FF3 model, the off-diagonal elements a∗1,ij,i6=j never significant. Again, this

property will be beneficial for our strategy 1M* and 1M*Opt, since when we construct a momentum

strategy on the optimal weights, we do exploit the interaction effects of the factors, but not of the

optimal weights. For the 1M strategy, the results are less convincing. For developed markets, none

of the models provides any significant diagonal term a1,ii for the 1M strategy. Therefore, we expect

also for the developed markets a superior performance of our strategies compared to the 1M strategy

based on traditional factor momentum. Hence, we now turn to the analysis whether the above results

using a simple VAR model indeed allow us to exploit the persistence in the optimal weights.

31
Figure 1: Long-only mixed and integrated approach to factor timing

The mixed approach to factor timing that is standard in the literature compared to our long-only integrated
approach to factor timing. We illustrate the different steps that lead to the optimal long-only portfolio with
N = 10 stocks and two factors in addition to the market factor.

Standard Approach: Mixed long-short approach to factor timing

Universe of Long-short Predicting Optimal weights of the


Factor portfolio
stocks factor portfolios factor portfolio factor portfolios based
returns
𝐹1 , 𝐹2 returns on the predicted returns
𝑟𝐹1 𝑟𝐹2
- + +
-
+ ± - 𝐹2 - ෞ
𝑟𝐹1 ෞ
𝑟𝐹2 + +
- -+- + + - -+ + +
𝐹1 - -+- + -
-
𝑤𝐹1 --
𝑤𝐹2
- -+ +
+ -
Our Approach: Integrated long-only approach to factor timing

Universe of Normalized Overall score Predicting Optimal portfolio based


stocks factor scores with factor factor score on the overall score of
score weights weights the stocks
𝑠1,1,𝑡 𝑠2,1,𝑡
𝑎1
⋮ ⋮ ⋮ ⋮ ෟ
𝑤1,𝑡 ෟ
𝑤2,𝑡 1
11 10
5
4
𝑠1,𝑁,𝑡 𝑠2,𝑁,𝑡 𝑎𝑁 +
⋮ 9
6

= 10
+ +
𝑠1,1,𝑡 𝑠2,1,𝑡 7
8
𝑤1,𝑡 ⋮ + 𝑤2,𝑡 ⋮
𝑠1,𝑁,𝑡 𝑠2,𝑁,𝑡

32
Figure 2: The grid of factor weights

This figure shows the grid of factor weights for the Fama–French three-factor (FF3), five–factor (FF5), and
six–factor model including momentum (FFC). We plot the factor weight values on the y-axis and show on the
x-axis the number of distinct factor values normalized from zero to one and mention the number of resulting
distinct weight combinations.

33
Figure 3: Cumulative logarithmic excess returns

This table reports the cumulative logarithmic excess return in US dollars above the one-month Treasury bill
rate of the value-weighted market portfolio in the US (MKT US) blue dashed line, the naive diversification
strategy for the Fama–French three-factor model (ND FF3) with the red dashed line, and the optimized
one-month momentum strategy 1M*Opt for the Fama–French three-factor model (FF3) with the solid line,
Fama–French five-factor model (FF5) with the dotted line, and the Fama–French five-factor model including
momentum (FFC) with the dash-dotted line (top), together with the excess return of the 1M*Opt strategy
over the MKT strategy for transaction costs of 0, 5, 10, and 15 Bps (bottom). The analyzed period is from
July 1963 to June 2018.

34
Figure 4: Subperiod analysis for 1M*Opt

Subperiod analysis for the optimized one-month momentum strategy 1M*Opt for the Fama–French three-
factor model (FF3). In Panel a), we plot the 1M*Opt strategy’s alpha against FF3 and mark the publication
date for the FF3 factor model. In Panel b), we perform a change point analysis and mark the resulting change
points accordingly. In Panel c), we plot the evolution of the equal-weighted FF3 factor portfolio. The reported
monthly alphas are expressed in percentage points. The analyzed period is from July 1963 to June 2018.

35
Figure 5: Subperiod analysis for 1M*Opt with transaction costs

Subperiod analysis for the optimized one-month momentum strategy 1M*Opt for the Fama–French three-
factor model (FF3). In Panel a), we plot the 1M*Opt strategy’s alpha against FF3 and mark the publication
date for the FF3 factor model. In Panel b), we perform a change point analysis and mark the resulting change
points accordingly. In Panel c), we plot the evolution of the equal-weighted FF3 factor portfolio. The reported
monthly alphas are expressed in percentage points. The analyzed period is from July 1963 to June 2018. We
assume a transaction cost of 10 Bps.

36
Figure 6: Subperiod analysis for 1M

Subperiod analysis for the optimized one-month momentum strategy 1M for the Fama–French three-factor
model (FF3). In Panel a), we plot the 1M strategy’s alpha against FF3 and mark the publication date for
the FF3 factor model. In Panel b), we perform a change point analysis and mark the resulting change points
accordingly. In Panel c), we plot the evolution of the equal-weighted FF3 factor portfolio. The reported
monthly alphas are expressed in percentage points. The analyzed period is from July 1963 to June 2018.

37
Figure 7: Subperiod analysis for 1M with transaction costs

Subperiod analysis for the optimized one-month momentum strategy 1M for the Fama–French three-factor
model (FF3). In Panel a), we plot the 1M strategy’s alpha against FF3 and mark the publication date for
the FF3 factor model. In Panel b), we perform a change point analysis and mark the resulting change points
accordingly. In Panel c), we plot the evolution of the equal-weighted FF3 factor portfolio. The reported
monthly alphas are expressed in percentage points. The analyzed period is from July 1963 to June 2018. We
assume a transaction cost of 10 Bps.

38
Table 1: The grid of factor weights

This table shows the 21 possible combinations for two arbitrary factors and a step size of 1/3 for the algorithm
presented in this work. The sum of the weights and is either one, zero or minus one. The number of
combinations grows exponentially with the number of factors.

# F1 F2 Σ # F1 F2 Σ # F1 F2 Σ # F1 F2 Σ # F1 F2 Σ
1 0.00 1.00 1 6 0.60 0.40 1 10 −1.00 1.00 0 13 0.00 −1.00 −1 18 −0.60 −0.40 −1
2 0.25 0.75 1 7 0.67 0.33 1 11 0.00 0.00 0 14 −0.25 −0.75 −1 19 −0.67 −0.33 −1
3 0.33 0.67 1 8 0.75 0.25 1 12 1.00 −1.00 0 15 −0.33 −0.67 −1 20 −0.75 −0.25 −1
4 0.40 0.60 1 9 1.00 0.00 1 16 −0.40 −0.60 −1 21 −1.00 0.00 −1
5 0.50 0.50 1 17 −0.50 −0.50 −1

39
Table 2: Overview of the strategies

This table shows the acronym and definition of each strategy. The last four columns show whether the strategy
applies factor timing, market-cap weights, includes interaction effects of the factors, or includes the information
of the covariance matrix (optimized).

Acronym Definition Factor Market-cap Interaction Optimized


timing weights effects
MKT Market-cap weighted portfolio of the entire universe. × X × ×

Long-only integrated approach that invests in the 30% highest


scored stocks when the score is built with . . .
ND . . . equal weights for each factor. × X × ×
FM . . . Fama MacBeth regressions to forecast the expected return. × X × ×
1M . . . each factor score receives a weight of ±1 with the same sign X X × ×
as the most recent long-short one-month factor return.
1M* . . . with the average weights of the 30% best weight- X X X ×
combinations of the most recent month.
1M*Opt . . . with the average weights of the 30% best weight- X × X X
combinations of the most recent month. The portfolio weights
are optimized with a minimum risk optimization with respect
to the market portfolio (MKT).

40
Table 3: Factor definitions and authors

This table shows the factors’ names in the first, definition in the second, authors with the year of publication
in the third, and acronyms in the fourth column. In headers five to ten to the right of the table, we show the
acronyms of the factor sets and indicate with a X the included characteristics.

FFC
FF3
FF5
Factor Definition Author Acronym
Size Inverse market capitalization of a company. Banz (1981) SMB X X X
Value Devil definition as defined in Asness and Frazzini (2013) with Rosenberg et al. (1985) HML X X X
the lagged book equity divided by the market value of the
last month
Profitability Operating profitability as defined by sales minus cost of Novy-Marx (2013) RMW X X
goods sold divided by book equity
Investment Inverse annual total book asset growth Cooper et al. (2008) CMA X X
Momentum Total return in US dollar of the past twelve months excluding Jegadesh (1990) WML X
the most recent month

41
Table 4: Summary statistics for the three universes

This table reports the (absolute) minimum and maximum monthly excess return, as well as number of stocks
and average market cap in million US dollars of the stocks in the universe on the left hand side, and the
annualized excess return, annualized standard deviation, Sharpe ratio, (absolute) minimum and maximum
monthly excess return of the value-weighted universe (MKT) and the value-weighted best and worst 30% stocks
of the Fama–French factors SMB, HML, RMW, CMA including momentum (WML) presented in Table 3 on
the right hand side. The universes are all NYSE, AMEX, and NASDAQ stocks above the NYSE median for
the US, and the MSCI Developed Markets (DM) and MSCI Emerging Markets (EM) universe of MSCI Inc.
The market cap for the US is the CRSP market cap and for the MSCI universes the free-float adjusted market
cap by MSCI Inc. All excess returns are in percentage US dollars above the one-month Treasury bill rate.

Panel US: All NYSE, AMEX, and NASDAQ stocks above the NYSE median
July 1963 to June 2018 MKT SMB HML RMW CMA WML
Min Ret 93.60 Ret ann 5.21 6.22 4.89 6.39 4.63 6.42 3.72 6.03 4.78 8.06 3.49
Max Ret 299.74 Std ann 15.03 18.97 14.57 14.70 16.48 15.13 16.18 13.77 17.70 17.57 17.73
Min Number 587 SR 0.35 0.33 0.34 0.43 0.28 0.42 0.23 0.44 0.27 0.46 0.20
Max Number 1,550 Min Ret 22.29 27.95 21.13 18.90 24.23 22.75 23.00 19.46 24.48 26.02 20.17
Avg Mcap M USD 7,235 Max Ret 16.40 23.33 16.81 22.60 21.12 17.90 14.59 14.97 21.09 19.98 25.00
Panel DM: MSCI Developed Markets
June 1998 to June 2018 MKT SMB HML RMW CMA WML
Min Ret 100.00 Ret ann 4.63 7.51 4.26 5.85 4.69 6.01 2.69 6.32 3.96 6.58 3.02
Max Ret 386.48 Std ann 14.74 18.12 14.51 18.14 14.49 13.66 16.07 14.27 17.28 15.67 20.61
Min Number 1,093 SR 0.31 0.41 0.29 0.32 0.32 0.44 0.17 0.44 0.23 0.42 0.15
Max Number 3,844 Min Ret 17.19 24.45 15.66 21.92 15.49 14.93 19.34 15.60 21.27 15.75 23.02
Avg Mcap M USD 14,317 Max Ret 10.07 19.49 10.10 19.00 10.83 9.70 11.66 10.61 13.96 17.82 26.77
Panel EM: MSCI Emerging Markets
June 1998 to June 2018 MKT SMB HML RMW CMA WML
Min Ret 90.97 Ret ann 8.05 11.14 7.60 11.16 8.32 10.68 5.44 9.26 7.58 11.45 5.24
Max Ret 453.21 Std ann 23.33 24.06 23.54 29.00 22.52 22.78 25.85 24.40 24.69 24.30 28.71
Min Number 514 SR 0.34 0.46 0.32 0.38 0.37 0.47 0.21 0.38 0.31 0.47 0.18
Max Number 2,337 Min Ret 29.39 28.64 30.34 38.33 25.57 27.21 35.35 25.91 30.16 27.57 35.26
Avg Mcap M USD 2,928 Max Ret 16.67 23.97 16.39 28.66 16.51 19.87 19.98 25.19 20.65 16.18 28.21

42
Table 5: Out-of-sample timing ability

This table reports the monthly alphas (in percentages) of the one-month momentum strategy that includes
the interaction effects (1M*), that excludes the interaction effects (1M), the naive diversification (ND), and
the Fama MacBeth (FM) strategies together with the block-resampled robust single hypothesis p-value (p-val )
and multiple hypothesis adjusted p-value (p-adj ) for the three universes from Table 4 with the US from July
1963 to December 2018, and DM as well as EM from June 1998 to June 2018. We show the alphas relative
to the FFC factor model that includes all factors MKT, SMB, HML, RMW, CMA, and WML. We mark
significant alphas below the 1%, 5%, and 10% significance levels based on the adjusted p-value by ∗∗∗ , ∗∗ , and

.

Panel A: Multi-factor monthly alpha from 1963 to 2018


FF3 FF5 FFC
1M* 1M ND FM 1M* 1M ND FM 1M* 1M ND FM
∗∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗
US 0.36 0.26 0.09 0.04 0.42 0.32 0.20 0.08 0.40 0.33 0.17 0.08∗
p-val (0.00) (0.00) (0.01) (0.44) (0.00) (0.00) (0.00) (0.07) (0.00) (0.00) (0.00) (0.04)
p-adj (0.00) (0.00) (0.04) (0.44) (0.00) (0.00) (0.00) (0.12) (0.00) (0.00) (0.00) (0.08)
Panel B: Multi-factor monthly alpha from 1998 to 2018
FF3 FF5 FFC
1M* 1M ND FM 1M* 1M ND FM 1M* 1M ND FM
US 0.42∗ 0.18 0.05 0.13 0.40∗∗ 0.36∗∗∗ 0.12 0.15 0.40∗ 0.30∗ 0.22∗ 0.13
p-val (0.04) (0.10) (0.39) (0.15) (0.01) (0.00) (0.07) (0.09) (0.04) (0.02) (0.01) (0.06)
p-adj (0.10) (0.28) (0.39) (0.28) (0.03) (0.01) (0.28) (0.28) (0.10) (0.10) (0.10) (0.22)
DM 0.42∗∗∗ 0.34∗∗ 0.01 −0.08 0.35∗∗ 0.29∗∗ 0.15 −0.03 0.47∗∗ 0.37∗∗ 0.11 −0.02
p-val (0.00) (0.00) (0.85) (0.25) (0.00) (0.01) (0.03) (0.62) (0.00) (0.00) (0.07) (0.77)
p-adj (0.00) (0.04) (0.95) (0.55) (0.03) (0.05) (0.16) (0.91) (0.01) (0.03) (0.25) (0.95)
EM 0.44∗∗ 0.29 0.16 −0.21 0.20 0.27 0.29∗ −0.16 0.33 0.38 0.23 −0.12
p-val (0.00) (0.02) (0.12) (0.03) (0.21) (0.08) (0.00) (0.03) (0.12) (0.04) (0.04) (0.12)
p-adj (0.01) (0.13) (0.31) (0.21) (0.31) (0.28) (0.05) (0.21) (0.31) (0.21) (0.23) (0.31)

43
Table 6: Out-of-sample robust Sharpe ratio test

This table reports the difference in annual Sharpe ratio from the MKT strategy together with the block-
resampled robust p-value in italics of the one-month momentum strategy that includes the interaction effects
(1M*), that excludes the interaction effects (1M), the naive diversification (ND), and the Fama MacBeth (FM)
strategies together with the robust single hypothesis p-value (p-val ) and multiple hypothesis adjusted p-value
(p-adj ) for the three universes from Table 4 with the US from July 1963 to December 2018, and DM as well
as EM from June 1998 to June 2018. We mark significant Sharpe ratio differences below the 1%, 5%, and
10% significance levels based on the adjusted p-value by ∗∗∗ , ∗∗ , and ∗ . For the Sharpe ratios, we use the test
of Ledoit and Wolf (2008).

Panel A: Annual Sharpe ratio difference to MKT from 1963 to 2018


FF3 FF5 FFC
1M* 1M ND FM 1M* 1M ND FM 1M* 1M ND FM
∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ ∗∗∗
US 0.28 0.19 0.10 0.01 0.28 0.24 0.23 0.05 0.25 0.24 0.26 0.05
p-val (0.00) (0.00) (0.16) (0.83) (0.00) (0.00) (0.00) (0.21) (0.00) (0.00) (0.00) (0.14)
p-adj (0.00) (0.01) (0.37) (0.83) (0.00) (0.00) (0.01) (0.37) (0.00) (0.00) (0.00) (0.33)
Panel B: Annual Sharpe ratio difference to MKT from 1998 to 2018
FF3 FF5 FFC
1M* 1M ND FM 1M* 1M ND FM 1M* 1M ND FM
US 0.30 0.10 0.07 0.08 0.23 0.24 0.15 0.11 0.19 0.17 0.26 0.10
p-val (0.05) (0.38) (0.56) (0.33) (0.02) (0.03) (0.34) (0.15) (0.10) (0.11) (0.09) (0.11)
p-adj (0.17) (0.70) (0.70) (0.70) (0.17) (0.17) (0.70) (0.41) (0.41) (0.41) (0.35) (0.41)
DM 0.34∗∗ 0.29 0.11 0.01 0.32 0.21 0.25 0.00 0.27 0.23 0.29 0.03
p-val (0.01) (0.02) (0.51) (0.92) (0.02) (0.02) (0.08) (0.96) (0.02) (0.03) (0.03) (0.70)
p-adj (0.04) (0.11) (0.84) (0.99) (0.10) (0.12) (0.20) (0.99) (0.10) (0.12) (0.12) (0.91)
EM 0.26∗∗ 0.19∗ 0.15 −0.10 0.14 0.16 0.22 −0.08 0.17 0.20 0.26 −0.06
p-val (0.00) (0.02) (0.12) (0.08) (0.10) (0.06) (0.04) (0.04) (0.04) (0.03) (0.01) (0.12)
p-adj (0.03) (0.07) (0.26) (0.26) (0.26) (0.20) (0.16) (0.18) (0.17) (0.13) (0.07) (0.26)

44
Table 7: Transaction costs analysis

This table reports the monthly alphas in percentage (Panel A) and annual Sharpe ratio difference from
the market (Panel B) of the one-month momentum strategy that includes the interaction effects (1M*) for
transaction costs of 0, 5, 10, and 15 Bps together with the block-resampled multiple hypothesis adjusted
p-values in italics. We jointly analyze the three factor models FF3, FF5, and FFC from Table 3 for the US
from July 1963 to June 2018 and the same factor models for the US, developed (DM), and emerging markets
(EM) from June 1998 to June 2018. We mark significant alphas below the 1%, 5%, and 10% significance levels
based on the adjusted p-value by ∗∗∗ , ∗∗ , and ∗ .

Panel A: 1M* monthly multi-factor alpha after transaction costs


1963 to 2018 1998 to 2018
US US DM EM
FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC
∗∗∗ ∗∗∗ ∗∗∗ ∗ ∗∗ ∗ ∗∗∗ ∗∗ ∗∗ ∗∗
0 Bps 0.36 0.42 0.40 0.42 0.40 0.40 0.42 0.35 0.47 0.44 0.20 0.33
p-adj (0.00) (0.00) (0.00) (0.06) (0.03) (0.06) (0.01) (0.04) (0.03) (0.03) (0.21) (0.17)
5 Bps 0.29∗∗∗ 0.35∗∗∗ 0.33∗∗∗ 0.36 0.33∗∗ 0.32 0.36∗∗ 0.28 0.40∗ 0.37∗ 0.13 0.26
p-adj (0.00) (0.00) (0.00) (0.13) (0.08) (0.14) (0.03) (0.12) (0.08) (0.08) (0.43) (0.29)
10 Bps 0.23∗∗∗ 0.27∗∗∗ 0.25∗∗∗ 0.30 0.26 0.25 0.31∗ 0.21 0.33 0.31 0.06 0.19
p-adj (0.01) (0.00) (0.01) (0.29) (0.22) (0.30) (0.09) (0.29) (0.18) (0.18) (0.72) (0.48)
15 Bps 0.16∗∗ 0.20∗∗∗ 0.18∗∗ 0.23 0.18 0.18 0.25 0.14 0.26 0.24 −0.01 0.12
p-adj (0.05) (0.02) (0.05) (0.50) (0.50) (0.62) (0.23) (0.62) (0.39) (0.39) (0.93) (0.72)

Panel B: 1M* annual Sharpe ratio difference to MKT after transaction costs
1963 to 2018 1998 to 2018
US US DM EM
FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC
0 Bps 0.28∗∗∗ 0.28∗∗∗ 0.25∗∗∗ 0.30 0.23 0.19 0.34∗∗ 0.32 0.27 0.26∗∗ 0.14 0.17
p-adj (0.00) (0.00) (0.00) (0.11) (0.11) (0.18) (0.05) (0.11) (0.11) (0.04) (0.18) (0.11)
5 Bps 0.23∗∗∗ 0.23∗∗∗ 0.20∗∗∗ 0.25 0.18 0.14 0.30∗ 0.27 0.23∗ 0.23 0.11 0.14
p-adj (0.00) (0.00) (0.00) (0.23) (0.23) (0.40) (0.10) (0.23) (0.23) (0.08) (0.40) (0.24)
10 Bps 0.18∗∗∗ 0.17∗∗∗ 0.15∗∗∗ 0.21 0.12 0.09 0.26 0.22 0.18 0.20 0.07 0.11
p-adj (0.01) (0.01) (0.01) (0.40) (0.49) (0.65) (0.20) (0.40) (0.40) (0.17) (0.65) (0.49)
15 Bps 0.13∗ 0.12∗ 0.09∗ 0.16 0.07 0.04 0.22 0.18 0.14 0.17 0.04 0.07
p-adj (0.06) (0.06) (0.10) (0.66) (0.79) (0.90) (0.36) (0.66) (0.66) (0.32) (0.90) (0.77)

45
Table 8: Including the information of the covariance matrix

This table reports the alphas in percentage (Panel A) and annual Sharpe ratio difference from the market (Panel
B) of the one-month momentum strategy that includes the interaction effects and includes the information of
the covariance matrix (1M*Opt) for transaction costs of 0, 5, 10, and 15 Bps together with the block-resampled
multiple hypothesis adjusted p-values in italics. We jointly analyze the three-factor models FF3, FF5, and
FFC from Table 3 for the US from July 1963 to June 2018 and the same factor models for the US, developed
(DM), and emerging markets (EM) from June 1998 to June 2018. We mark significant alphas below the 1%,
5%, and 10% significance levels based on the adjusted p-value by ∗∗∗ , ∗∗ , and ∗ .

Panel A: 1M*Opt monthly multi-factor alpha after transaction costs


1963 to 2018 1998 to 2018
US US DM EM
FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC
∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗ ∗
0 Bps 0.33 0.34 0.36 0.40 0.38 0.44 0.43 0.29 0.42 0.35 0.29 0.30∗
p-adj (0.00) (0.00) (0.00) (0.00) (0.00) (0.01) (0.00) (0.01) (0.00) (0.02) (0.06) (0.09)
5 Bps 0.27∗∗∗ 0.27∗∗∗ 0.30∗∗∗ 0.34∗∗ 0.31∗∗ 0.38∗∗ 0.38∗∗ 0.23∗∗ 0.35∗∗ 0.29∗ 0.22 0.24
p-adj (0.00) (0.00) (0.00) (0.02) (0.02) (0.02) (0.02) (0.04) (0.02) (0.06) (0.15) (0.18)
10 Bps 0.21∗∗∗ 0.21∗∗∗ 0.23∗∗∗ 0.28∗ 0.24∗ 0.31∗ 0.32∗∗ 0.17 0.29∗ 0.23 0.16 0.17
p-adj (0.00) (0.00) (0.00) (0.07) (0.07) (0.08) (0.05) (0.17) (0.07) (0.17) (0.35) (0.35)
15 Bps 0.15∗∗ 0.14∗∗ 0.16∗∗ 0.22 0.18 0.24 0.27 0.10 0.22 0.17 0.09 0.10
p-adj (0.01) (0.01) (0.01) (0.23) (0.25) (0.25) (0.13) (0.46) (0.23) (0.43) (0.65) (0.65)
Panel B: 1M*Opt annual Sharpe ratio difference to MKT after transaction costs
1963 to 2018 1998 to 2018
US US DM EM
FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC
∗∗∗ ∗∗∗ ∗∗∗ ∗∗ ∗∗ ∗∗ ∗∗ ∗ ∗∗ ∗ ∗
0 Bps 0.28 0.27 0.29 0.31 0.28 0.31 0.40 0.33 0.38 0.24 0.21 0.21∗
p-adj (0.00) (0.00) (0.00) (0.04) (0.04) (0.04) (0.04) (0.09) (0.04) (0.09) (0.09) (0.09)
5 Bps 0.23∗∗∗ 0.22∗∗∗ 0.24∗∗∗ 0.27∗ 0.23∗ 0.26∗ 0.36∗ 0.28 0.34∗ 0.21 0.18 0.18
p-adj (0.00) (0.00) (0.00) (0.09) (0.09) (0.09) (0.09) (0.15) (0.09) (0.15) (0.15) (0.15)
10 Bps 0.18∗∗∗ 0.17∗∗∗ 0.19∗∗∗ 0.22 0.18 0.21 0.33 0.24 0.29 0.19 0.15 0.15
p-adj (0.00) (0.00) (0.00) (0.18) (0.22) (0.18) (0.17) (0.25) (0.16) (0.25) (0.25) (0.25)
15 Bps 0.14∗∗ 0.12∗∗ 0.14∗∗ 0.18 0.13 0.16 0.29 0.20 0.25 0.16 0.12 0.12
p-adj (0.02) (0.02) (0.01) (0.32) (0.45) (0.35) (0.26) (0.45) (0.27) (0.45) (0.45) (0.45)

46
Table 9: Traditional factor momentum portfolios

This table reports the monthly alphas in percentage (Panel A) and annual Sharpe ratio difference from the
market (Panel B) of the traditional one-month factor momentum strategy 1M for transaction costs of 0, 5,
10, and 15 Bps together with the block-resampled multiple hypothesis adjusted p-values in italics. We jointly
analyze the three-factor models FF3, FF5, and FFC from Table 3 for the US from July 1963 to June 2018
and the same factor models for the US, developed (DM), and emerging markets (EM) from June 1998 to June
2018. We mark significant alphas below the 1%, 5%, and 10% significance levels based on the adjusted p-value
by ∗∗∗ , ∗∗ , and ∗ .

Panel A: 1M monthly multi-factor alpha after transaction costs


1963 to 2018 1998 to 2018
US US DM EM
FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC
∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗ ∗∗ ∗ ∗∗ ∗
0 Bps 0.26 0.32 0.33 0.18 0.36 0.30 0.34 0.29 0.37 0.29 0.27 0.38∗
p-adj (0.00) (0.00) (0.00) (0.00) (0.01) (0.06) (0.04) (0.05) (0.03) (0.06) (0.14) (0.09)
5 Bps 0.20∗∗∗ 0.25∗∗∗ 0.26∗∗∗ 0.12 0.29∗∗ 0.23 0.28 0.23 0.30 0.22 0.20 0.32
p-adj (0.01) (0.00) (0.00) (0.32) (0.03) (0.20) (0.13) (0.20) (0.11) (0.20) (0.31) (0.20)
10 Bps 0.13∗ 0.18∗∗∗ 0.19∗∗∗ 0.06 0.22 0.15 0.22 0.16 0.23 0.16 0.13 0.25
p-adj (0.06) (0.01) (0.01) (0.62) (0.14) (0.55) (0.34) (0.54) (0.34) (0.54) (0.60) (0.54)
15 Bps 0.07 0.12 0.13 −0.01 0.15 0.08 0.16 0.09 0.16 0.10 0.06 0.18
p-adj (0.31) (0.12) (0.12) (0.94) (0.52) (0.89) (0.70) (0.89) (0.70) (0.89) (0.92) (0.83)

Panel B: 1M annual Sharpe ratio difference to MKT after transaction costs


1963 to 2018 1998 to 2018
US US DM EM
FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC FF3 FF5 FFC
0 Bps 0.19∗∗∗ 0.24∗∗∗ 0.24∗∗∗ 0.10 0.24 0.17 0.29 0.21 0.23 0.19∗ 0.16 0.20
p-adj (0.00) (0.00) (0.00) (0.38) (0.14) (0.19) (0.14) (0.14) (0.14) (0.08) (0.17) (0.14)
5 Bps 0.14∗∗ 0.19∗∗∗ 0.19∗∗∗ 0.06 0.19 0.12 0.25 0.16 0.18 0.16 0.13 0.16
p-adj (0.02) (0.00) (0.00) (0.62) (0.32) (0.42) (0.28) (0.32) 0.32 (0.18) (0.34) (0.29)
10 Bps 0.10 0.13∗∗ 0.14∗∗ 0.01 0.14 0.07 0.20 0.11 0.13 0.14 0.09 0.13
p-adj (0.11) (0.04) (0.04) (0.90) (0.63) (0.73) (0.47) (0.64) (0.64) (0.38) (0.64) (0.51)
15 Bps 0.05 0.08 0.09 −0.03 0.08 0.02 0.16 0.06 0.08 0.11 0.06 0.10
p-adj (0.38) (0.23) (0.23) (0.95) (0.92) (0.95) (0.71) (0.92) (0.92) (0.65) (0.92) (0.77)

47
Table 10: Summary of VAR estimation

This table presents the results from our VAR estimation for the dynamics of the optimal factor score weights
and factor returns. Weights and returns were calculated based on the Fama–French factor models FF3, FF5,
and FFC for the US, and the MSCI Developed Markets (DM) and MSCI Emerging Markets (EM) universe
of MSCI Inc. We report the number of factors F in addition to the market factor, the number of significant
coefficients in the matrix A∗1 , and the number of significant diagonal a∗1,ii and off-diagonal elements a∗1,ij,i6=j .
We express these numbers as a fraction of the total numbers of elements. Hence, the first entry for A∗1 means
that two of the, in total, four entries of the matrix are significant. Similarly, we denote the corresponding
matrix and elements for the 1M model by A1 and a1,ij , respectively. We calculate significance at the 1% level.

Panel A: US Market, 1962–2018


Factor score persistence Factor return persistence
F A∗1 a∗1,ii a∗1,ij,i6=j F A1 a1,ii a1,ij,i6=j
FF3 2 2/4 2/2 0/2 2 2/4 2/2 0/2
FF5 4 5/16 4/4 1/12 4 2/16 1/4 1/12
FFC 5 11/25 3/5 8/20 5 3/25 2/5 1/20
Panel B: Developed Markets, 1998–2018
Factor score persistence Factor return persistence
F A∗1 a∗1,ii a∗1,ij,i6=j F A1 a1,ii a1,ij,i6=j
FF3 2 2/4 2/2 0/2 2 1/4 0/2 1/2
FF5 4 4/16 2/4 2/12 4 1/16 0/4 1/12
FFC 5 2/25 1/5 1/20 5 0/25 0/5 0/20
Panel C: Emerging Markets, 1998–2018
Factor score persistence Factor return persistence
F A∗1 a∗1,ii a∗1,ij,i6=j F A1 a1,ii a1,ij,i6=j
FF3 2 1/4 1/2 0/2 2 2/4 1/2 1/2
FF5 4 2/16 2/4 0/12 4 2/16 1/4 1/12
FFC 5 3/25 2/5 1/20 5 1/25 1/5 0/20

48

You might also like