You are on page 1of 37

Performance of Hierarchical Equal Risk

Contribution Algorithm in China Market

WEIGE HUANG∗

ABSTRACT

This paper studies the performance of the portfolios based on the Hierarchical Equal
Risk Contribution algorithm in China stock market. Specifically, we consider a variety
of risk measures for calculating weight allocations which include equal weighting, vari-
ance, standard deviation, expected shortfall and conditional drawdown risk and four
types of linkage criteria used for agglomerative clustering, namely, single, complete,
average, and Ward linkages. We compare the performance of the portfolios based on
the HERC algorithm to the equal-weighted and inverse-variance portfolios. We find
that most HERC portfolios are not able to beat the equal-weighted and inverse-variance
portfolios in terms of several comparison measures and HERC with Ward-linkage seems
to dominate the ones with other linkages. However, the results do not show that any
risk measures can beat other measures consistently.

Keywords: Hierarchical Equal Risk Contribution, Asset Allocation, Machine Learning, Port-
folio Optimization, Inverse-Variance Portfolio, Hierarchical Risk Parity
JEL Codes: G00, G10, G11

Assistant Professor, Wenlan School of Business, Zhongnan University of Economics and Law, 182
Nanhu Ave., Wuhan 430073, P.R.China. Primary Email: darren1988@163.com; Secondary Email:
weige huang@zuel.edu.cn

Electronic copy available at: https://ssrn.com/abstract=3695598


Financial markets are complex and usually organized in a hierarchical structure. That
is, assets are correlated in a hierarchical way. For instance, stocks are grouped into different
industries and assets such as stocks and bonds are grouped into different asset classes. Thus,
the need of a quantitative description of these hierarchies arises (see Anderson 1972; Simon
1962). López de Prado (2016) points out that correlation matrices lack the notion of hierarchy
and provides a concrete example:

“Stocks could be grouped in terms of liquidity, size, industry, and region, where
stocks within a given group compete for allocations. In deciding the allocation
to a large publicly-traded U.S. financial stock like J.P. Morgan, we will consider
adding or reducing the allocation to another large publicly-traded U.S. bank like
Goldman Sachs, rather than a small community bank in Switzerland, or a real
estate holding in the Caribbean.”

To utilize this hierarchy in portfolio construction, new methods are required because tradi-
tional portfolio optimization methods do not deal with hierarchical relation between assets.
The Hierarchical Equal Risk Contribution (HERC) algorithm proposed by Raffinot (2018) is
one of the approaches aimed to the purpose. Using hierarchical clustering algorithm, HERC
firstly grows a hierarchical tree which is able to represent the hierarchical structure in the
assets. HERC then allocates weights on assets based on the tree. Literature on HERC and
related topics and details of HERC are introduced in Section I and II, respectively.
In this paper, we study the HERC performance in China stock market and compare the
performance to the ones based on the equal-weighted method and traditional risk parity,
exemplified by the Inverse-Variance Portfolio (IVP).1 We also consider two new portfolio
construction approaches by combining HERC with setting weight bars to further select as-
sets after HERC gives the allocating weights and by grouping stocks by industries before
1
We note that we are not able to apply the Critical Line Algorithm in this paper because the stock pool
is large and the covariance matrix of stock returns is ill-conditional and even singular.

Electronic copy available at: https://ssrn.com/abstract=3695598


conducting HERC algorithm. We find that most portfolios based on HERC algorithm are
not able to beat IVP and EW portfolios in China market. We also find that the portfolios
based on HERC algorithm with Ward-linkage seem to perform better than the ones based
on other linkages but there does not exist any risk measures which seem to dominate other
risk measures.

I. Literature on HERC
There is an abundance of mathematical literature dealing with portfolio optimization,
such as the classical Markowitz mean variance optimisation (Markowitz and Todd 2000),
Black-Litterman models (Black and Litterman 1992) and many more. Specifically, Harry
Markowitz developed the Critical Line Algorithm (CLA, Markowitz 1955, 1959) for this
purpose (see also Bailey and López de Prado 2013). However, the algorithm comes with its
own set of disadvantages: first, CLA requires invertibility of the covariance matrix; second,
CLA is dependent on the estimation of stock-returns which are very hard to estimate with
sufficient accuracy; third, CLA considers the correlations between the returns of all the assets
in a portfolio and this leads to a very large correlation dependency connected graph where
each asset is connected to all the other assets and thus is a computationally expensive; finally,
assets are not all correlated and there are some hierarchy in the relation between assets. This
is important point. Assets can be grouped into categories and compete with each other in a
category for allocations within a portfolio (see López de Prado 2018).
To address the caveats of CLA, López de Prado (2016) first introduce the Hierarchical
Risk Parity (HRP) method. The HRP algorithm can be broken down into three major
steps: hierarchical tree clustering, matrix seriation, and recursive bisection.2 HRP portfolios
address three major concerns of quadratic optimizers in general and Markowitz’s CLA in
2
For a detailed explanation of how hierarchical risk parity works, check an excellent blog post by Aditya
Vyas: The Hierarchical Risk Parity Algorithm: An Introduction. See also for the implementation of HRP
using the class of HierarchicalRiskParity in the mlfinlab package.

Electronic copy available at: https://ssrn.com/abstract=3695598


particular: instability, concentration, and under-performance (see López de Prado 2018,
2020c).3
Exploiting the same basic idea of correlation matrices lack the notion of hierarchy in a
different way from HRP, Raffinot (2017) proposes a hierarchical clustering based asset al-
location (HCAA). HCAA allocates capital within and across clusters of assets at multiple
hierarchical levels. The “Hierarchical 1/N” proposed in Raffinot (2017) finds a diversified
weighting by distributing capital equally among each cluster hierarchy. Then, within a clus-
ter, an equal-weighted allocation is computed. This naive approach inspired by DeMiguel,
Garlappi, and Uppal (2009) requires neither expected returns nor risk measures. Yet, it may
be too simple since risk management is not part of the weighting strategy.
The Hierarchical Equal Risk Contribution (HERC) algorithm proposed by Raffinot (2018)
takes inspiration from the HRP algorithm and HCAA and uses machine learning to allocate
weights efficiently. While both the HERC and HRP algorithms use hierarchical tree cluster-
ing to allocate their weights, there are some subtle differences between the two. HRP starts
by reorganizing the covariance matrix to place similar investments together. Then, it employs
an inverse-variance weighting allocation based on the number of assets with no further use
of the clustering. Raffinot (2018)’s HERC enhances HCAA by taking the best of both HRP
and HCAA methods by merging and enhancing the machine learning approach of HCAA
and the Top-Down recursive bisection of HRP. Furthermore, HERC is extended to downside
risk measures such as Conditional Value at Risk (CVaR) and Conditional Drawdown at Risk
(CDaR). Also, HRP algorithm uses only the variance as a risk measure for assigning weights
and the HERC algorithm allows investors to use other risk metrics like standard deviation,
expected shortfall and conditional drawdown at risk. More details will be introduced in
3
There are some recent applications of HRP. For instance, Raffinot (2017) evaluates the performance of
HRP across three datasets consisting of S&P 500 sectors, multi-assets, and individual stocks. Recently, P.
Jain and S. Jain (2019) apply the HRP strategy to individual stocks that comprise the NIFTY 50 index.
Lohre, Rother, and Schäfer (2020) employ the HRP strategy in a multi-asset multi-factor allocation context,
showing that it achieves better tail risk results.

Electronic copy available at: https://ssrn.com/abstract=3695598


next section. Raffinot (2018) evaluates the performance of HCAA and HERC and shows
that the “Hierarchical 1/N” is difficult to beat, but HERC portfolios based on downside risk
measures achieve statistically better risk-adjusted performances, especially those based on
the Conditional Drawdown at Risk.

II. HERC
Both of HRP and HERC algorithms are approaches which are developed to model the
hierarchical structure in the asset returns. Although both approaches grow the tree which
represents the hierarchies, the way of growing the hierarchical tree is different. More specifi-
cally, the HRP method uses single linkage and grows the tree to maximum depth. However,
it may not be able to obtain the optimal number of clusters identified by growing the tree
maximally. Instead, HERC calculates the optimal number of clusters and cuts the hierarchi-
cal tree to the required height and clusters before allocating the weights. This is the main
difference between HERC and HRP.

II.I. Overview of HERC

The HERC algorithm includes four main steps: hierarchical tree clustering, calculating
optimal number of clusters, top-down recursive bisection and naive risk parity.

Step 1: Hierarchical Tree Clustering


The purpose of clustering is to place assets with similarity suggested by the data into same
clusters and dissimilar assets into different clusters. Hierarchical clustering is to build recur-
sively a binary tree of the assets that successively merges similar groups of points, resulting
in a tree-based representation of the assets called a dendrogram. More specifically, hier-
archical clustering uses the famous Hierarchical Tree Clustering algorithm (Agglomerative
Clustering) to break down the assets in the portfolio into different hierarchical clusters (see

Electronic copy available at: https://ssrn.com/abstract=3695598


López de Prado 2016). Hierarchical clustering segregates assets into clusters which mimic
the real-life interactions between the assets in a portfolio - some stocks are related to each
other more than others and hence can be grouped within the same cluster. We note that
hierarchical clustering requires a suitable distance measure (similarity measure) which will
be detailed in Section II.II.

Step 2: Calculating Optimal Number of Clusters


Calculating optimal number of clusters before allocating the portfolio weights is mainly
where HERC deviates from the traditional HRP algorithm. HRP uses single linkage or other
linkages and grows the tree to maximum depth. This way of maximally growing tree may
lead to a sub-optimal number of clusters and consequently sub-optimal weights. Instead,
HERC computes the optimal number of clusters and cuts the hierarchical tree formed in last
step to the required height and clusters before calculating the weights.4

Step 3: Top-Down Recursive Bisection


After calculating the optimal number of clusters, Top-Down Recursive Bisection is used to
compute the allocating weights. We note that there is a fundamental difference between the
recursive bisections of HERC and HRP algorithms. In particular, the weights in HRP trickle
down the tree by breaking it down the middle based on the number of assets without using
the structure of the dendrogram obtained in Step 1. This is a fundamental disadvantage of
HRP. This disadvantage is improved upon by HERC by dividing the tree, at each step, based
on the structure induced by the dendrogram while calculating the cluster contributions. At
4
Currently, the Gap Index is used for calculating the required number of clusters (see Tibshirani, Walther,
and Hastie 2001). The Optimal Number of Clusters (ONC) algorithm proposed by Marcos Lopez de Prado
detects the optimal number of K-Means clusters using a correlation matrix as input (López de Prado 2020c;
López de Prado and Lewis 2019). López de Prado (2020a) briefly describes the logic behind the ONC
algorithm and López de Prado (2020b) explains why the angular distance metric is used to get distances
between elements.

Electronic copy available at: https://ssrn.com/abstract=3695598


each level of the tree, an equal risk contribution allocation is used i.e. the weights are:

RC1
α1 = 1 − ; α2 = 1 − α1
RC1 + RC2

where α1 and α2 are the weights of the left and right clusters respectively and RC1 and RC2
are the risk contributions the two clusters.5 The risk measures will be introduced in Section
II.III.

Step 4: Naive Risk Parity


This step calculates the final asset weights using the cluster weights in Step 3. To the
purpose, an initial set of weights wN RP in the same cluster is calculated using the naive risk
parity allocation. That is, in the same cluster assets are allocated weights in proportion to
the inverse of their respective risk: higher risk assets will receive lower portfolio weights and
lower risk assets will receive higher weights. Here the risk can be quantified in different ways
- variance, CVaR, CDaR, max daily loss and so on. Then, the final weights are given by the
following equation:
wfi inal = wN
i i
RP C , i ∈ Clusters

i i
where wN RP refers to naive parity weights of assets in the ith cluster and C is the weight

of ith cluster calculated in Step 3.

II.II. Linkage Criteria

Hierarchical clustering starts with every observation representing a singleton cluster and
then combines the clusters sequentially, reducing the number of clusters at each step until
−1
5 diag(Σ)
To compute RC, the weights for assets in the cluster is computed: w = tr(dig(Σ) −1 ) where Σ is the

covariance matrix of assets in the same cluster, diag denotes diagonal of the matrix and tr the trace. Then,
the risk contribution from asset i, RCwi = wi √(Σw) i
w0 Σw
. Lastly, RCj = N · RCwi where N is the total number
of assets in the cluster j = 1, 2. In this case, we compute the weights using variance as risk measure. In fact
in HERC, the way to compute w can be different according to the risk metrics chosen.

Electronic copy available at: https://ssrn.com/abstract=3695598


only one cluster is left. At each step, the similar two (least dissimilar) clusters are merged into
a single cluster, producing one less cluster at the next higher level. Therefore, hierarchical
clustering requires a suitable distance measure of dissimilarity between two clusters. This
paper used the following distance (Mantegna 1999):

q
Di,j = 2(1 − ρi,j )

where Di,j is the correlation-distance index between the ith and jth asset, and ρi,j is the
respective Pearson’s correlation coefficient.6 We note that different definitions of the distance
between clusters can produce radically different dendrograms. In hierarchical clustering
analysis, the distance is known as the “linkage criterion.” This section briefly introduces
four common linkage criteria in clustering analysis as described below (Papenbrock 2011;
Raffinot 2017): single, complete, average and Ward linkages.

Single-linkage
The idea behind single-linkage is to form groups of elements, which have the smallest
distance to each other (nearest neighboring clustering). The distance between two clusters
is the minimum of the distance between any two points in the clusters. For clusters Ci , Cj ,

dCi ,Cj = minx,y {D(x, y)|x ∈ Ci , y ∈ Cj }

This method is simple but sensitive to outliers. It oftentimes forms clusters that are chained
together and leaves large clusters. namely chaining problem. The single-link clustering
algorithm is very useful to separates assets that were very different from the rest by using
insights in the correlation structure between assets. It is a good choice if the separation of
assets is preferred and high weights should be put on outliers.
6
One can use other distance measures which base on non-linear codependency measures. See for instance
López de Prado (2020c).

Electronic copy available at: https://ssrn.com/abstract=3695598


Complete-linkage
The idea behind the complete-linkage algorithm (also called the farthest neighbor cluster-
ing) is that: assets should be grouped together in a way that they are not too different from
each other when merged in a cluster. Thus, it tries to avoid those large groups by considering
the largest distances between elements and has a much stronger definition of “similar pair
of clusters”. For this reason, this method tends to produce compact clusters of similar size
but is quite sensitive to outliers. The distance between two clusters is the maximum of the
distance between any two points in the clusters. For clusters Ci , Cj ,

dCi ,Cj = maxx,y {D(x, y)|x ∈ Ci , y ∈ Cj }

The complete-link algorithm seems suitable for investors interested in grouping stocks that
are similar in one cluster.

Average-linkage
The average-linkage algorithm is a compromise between the single-linkage and complete-
linkage algorithm and is considered to be a fairly robust method. The distance between two
clusters is the average of the distance between any two points in the clusters. For clusters
Ci , Cj ,
dCi ,Cj = meanx,y {D(x, y)|x ∈ Ci , y ∈ Cj }

Ward-linkage
Whereas single-linkage, complete-linkage and average-linkage can be classified as graph-
based clustering algorithms, Ward’s method has a prototype-based view in which the clusters
are represented by a centroid. For this reason, the proximity between clusters is usually
defined as the distance between cluster centroids. The Ward method uses the increase in the

Electronic copy available at: https://ssrn.com/abstract=3695598


sum of the squares error (SSE) to determine the clusters. For clusters Ci , Cj with sizes mi
and mj , respectively,
mi mj
dCi ,Cj = ||ci − cj ||
mi + mj
where ci and cj are the centroids for the clusters.

II.III. Risk Measures

In Step 3 and 4 of HERC in Section II.I, risk metrics are used to compute the risk con-
tribution and final weights. In the investing world, investment managers use many other
important risk metrics other than variance which is a very simple and popular representa-
tion of risk. These other risk metrics can correctly reflect the true risk of a portfolio. For this
reason, HERC tweaks the original HRP algorithm to allocate its weights based on different
risk representations of the clusters and generate better weights. This paper considers the
following risk metrics: equal weighting, variance, standard deviation, expected shortfall and
conditional drawdown risk. For equal weighting, all clusters are weighed equally in terms of
risk.

II.IV. Comparison Measures

To compare the performance of different algorithms out-of-sample, we compute the follow-


ing three comparison criteria given the time series of daily out-of-sample returns generated
by each strategy:

• The adjusted sharpe ratio (ASR) (Pézier and White 2008) explicitly adjusts for skew-
ness and kurtosis by incorporating a penalty factor for negative skewness and excess
kurtosis:
µ3  µ4 − 3
SR2
 
ASR = SR 1 + SR −
6 24

10

Electronic copy available at: https://ssrn.com/abstract=3695598


where µ3 and µ4 are the skewness and kurtosis of the returns distribution and SR
µ−rf
denotes the traditional Sharpe ratio (SR = σ
, where rf is the risk-free rate).7

• The certainty-equivalent return (CER) is the risk-free rate of return that the investor
is willing to accept instead of undertaking the risky portfolio strategy. DeMiguel,
Garlappi, and Uppal (2009) define the CER as

γ
CER = (µ − rf ) − σ 2
2

where γ is the risk aversion. In this paper, results are reported for the case of γ = 1.
More precisely, the employed definition of CER captures the level of expected utility
of a mean–variance investor, which is approximately equal to the certainty-equivalent
return for an investor with quadratic utility (DeMiguel, Garlappi, and Uppal 2009).
It is the most important number to consider for building profitable portfolios (Levy
2017).

• The third measure we considered is the max drawdown (M DD) which is an indicator
of permanent loss of capital. It measures the largest single drop from peak to bottom
in the value of a portfolio. In brief, the M DD offers investors a worstcase scenario.

III. Data and Results


The price dataset we use in this paper are collected as followings: (1) we collect daily close
prices for all stocks which are listed in Shanghai and Shenzhen stock markets; (2) we use the
dataset ranging for twenty years from Jan 1th, 2000 to Jan 1th, 2020 and keep stocks which
have at least two years’ trading records. We use the prices in last year to compute the risk
measures and the portfolio weights at the first trading day of each month. Thus, we keep
7
The daily time series for long-term bond yield in China is not available for us. Therefore, in this paper
we set the risk-free rate to be 3% which is close to the average of 10-year long-term bond yield in China.

11

Electronic copy available at: https://ssrn.com/abstract=3695598


stocks with at least two years’ trading records to allow us to compute the risk measures and
construct the portfolio. We note that we replace NaN with the last non-NaN value for each
price. Then we drop stocks with maximum length of consecutive repeated prices greater
than 126 days (half year). That is, we drop stocks which continuously are not traded in half
year. Totally, our sample includes 2917 stocks.
At the first trading day of each month, we compute the portfolio weights using daily prices
of last 12 months. Therefore, we note that we compute weights for only stocks with at least
12 months of trading as of the first trading day of the month. That is, stocks listing less than
a year in the month in which the weights are computed are not used to construct a portfolio
for the month. The number of stocks becomes larger as the time passes and the dimension
of the covariance matrix of stock returns becomes greater. This leads to an ill-degenerated
or even a singular covariance matrix which makes CLA unstable or impossible to implement.
This paper implement the HERC using the python package “mlfinlab” by Hudson &
Thames. In particular, the class HierarchicalEqualRiskContribution in the mlfinlab package
is applied to compute the portfolio weights based on different combinations of risk measures
and linkage criteria.8
In this paper, we mainly consider four variants of HERC algorithms:

• Portfolio Construction - 1: Apply HECR directly to entire sample with with ONC
being set to be 10 and without setting any weight bars to further select stocks after
HECR computes the allocating weights.

• Portfolio Construction - 2: All are the same as Portfolio Construction - 1 except with
ONC being set to be 72.

• Portfolio Construction - 3: All are the same as Portfolio Construction - 1 except


setting the weight bars to be 1%. That is, the stocks with HERC allocating weights
8
In the paper, the confidence level of 0.05 is used for calculating expected shortfall and conditional
drawdown at risk.

12

Electronic copy available at: https://ssrn.com/abstract=3695598


less than 1% are not used to construct HERC portfolios. We generally name a portfolio
constructed by HERC algorithm HERC portfolio.

• Portfolio Construction - 4: The stocks are firstly grouped by industry code. Then,
HERC algorithm as the one in Portfolio Construction 1 is applied to each industry.
After obtaining the HERC returns for each industry, we compute the average returns
by equally weighing returns for all industries.

We compute ASR, CER and MDD for all portfolios and rank the portfolios by these
comparison measures to evaluate the relative performance of HERC portfolios, EW and IVP
portfolios. For each construction method, we have totally 22 portfolios to compare to each
other. The main results are shown in Tables 1 and 2. The more detailed results could be
found in Appendix A.

III.I. Portfolio Construction - 1

This section shows the performance of HERC with ONC being set to be 10. We note that
this number is arbitrary and one can try other numbers. We do not allow HERC to choose
the optimal number of clusters because the computation takes very long time to obtain the
optimal number when the number of assets is large, e.g., in thousands in this study.9
We first introduce the notations in Table 1. EW and IVP denote equal-weight and Inverse-
Variance portfolios, respectively. They are highlighted by red to note that they are compared
to HERC portfolios. We denote HERC portfolios by the combination of risk metrics and
linkages. The last letter denotes linkage. More specifically, W denotes Ward linkage, A
average linkage, S single linkage, C complete linkage. Other letters before these letters denote
9
We quote from the document for HERC class in the mlfinlab package: “The calculation of optimal number
of clusters using the Gap Index makes the algorithm run a little slower and you will notice a significant speed
difference for larger datasets. If you know the number of clusters for your data beforehand, it would be better
for you to pass that value directly to the method using the optimal num clusters parameter. This will bypass
the Gap Index method and speed up the algorithm.”

13

Electronic copy available at: https://ssrn.com/abstract=3695598


risk metrics. ES represents expected shortfall, SD standard deviation, CDR conditional
drawdown riks and V variance. For example, VW denotes the combination of risk measure,
variance, and Ward linkage.
For the construction - 1 approach, we can see that the top 5 portfolios in terms of ASR
are VW, SDW, CDRW, IVP and EW portfolios (see first column in Table 1). The top 5
portfolios in terms of different comparison measures are boxed. VW performs best in terms
of ASR. In terms of CER, VW, SDW and CDRW portfolios still beat IVP and EW portfolios.
By MDD, however, EW portfolio beats VW, SDW and CDRW portfolios.
Table 3 in Appendix A shows the actual values of the comparison measures. Figure 1
in Appendix A visualizes them by sorting portfolios according to ASR. We note that the
correlation between ASR and CER is 0.93, ASR and MDD -0.92, CER and MDD -0.98. The
correlations between comparison measures are high. We also note that Ward-linkage seems
to work better than other linkages.

III.II. Portfolio Construction - 2

For this construction method, we set the ONC to be 72 which is equal to the number of
industries in our sample and repeat the study above. From the second panel of Table 1,
we can see that ESW portfolio performs better than IVP and EW portfolios in terms of
ASR and CER. The top 5 by all comparison measures (except CDRC) are HERC portfolios
constructed by Ward-linkage, IVP and EW portfolios. This result again shows that HERC
portfolios based on Ward-linkage seem to perform better relative to HERC portfolios based
on other linkages.
Table 4 in Appendix A shows the values of these comparison measures. Figure 2 in Ap-
pendix A visualizes them by sorting portfolios according to ASR. We note that the correlation
between ASR and CER is 0.95, ASR and MDD -0.92, CER and MDD -0.95. The correlations
between comparison measures are high. That is, we are able to come to similar conclusion

14

Electronic copy available at: https://ssrn.com/abstract=3695598


with different measures.

III.III. Portfolio Construction - 3

In this case, we do not include into a portfolio the stocks with HERC allocating weights
less than 1% to reduce the possibility of obtaining a portfolio which allocates very small
weights on some stocks. We note that one can try other weight bars. From the third panel
of Table 1, we can see that IVP and EW perform better than any others in terms of ASR.
We note again that HERC portfolios based on Ward-linkage take up all positions in top 5
(excluding IVP and EW) by any comparison measures.
Table 5 in Appendix A shows the values of these comparison measures. Figure 3 in Ap-
pendix A visualizes them by sorting portfolios according to ASR. We note that the correlation
between ASR and CER is 0.33, ASR and MDD -0.07, CER and MDD -0.80. The correlations
between comparison measures are low. That is, we are not able to come to similar conclusion
with different measures.

III.IV. Portfolio Construction - 4

This section consider another way of portfolio construction which takes into account of
the industry hierarchy. More specifically, at the first trading day of very month we manually
group stocks according to the industry code A in China market. The industries include
Finance, Utilities, Properties, Conglomerates, Industrials and Commerce. We equally weigh
each industry. Then, we grow a hierarchical tree using HERC under each industry. The
rightest panel of Table 1 shows the results. We can see that IVP and EW portfolios perform
better than any other HERC portfolios by ASR and CER measures. We note that HERC
portfolios based on Ward-linkage are not in top 5 portfolios any more. This is mainly because
the hierarchy has been taken away by grouping stocks by industry and equally weighting.
Thus, HERC method does not have an advantage any more because the HERC algorithm’s

15

Electronic copy available at: https://ssrn.com/abstract=3695598


advantage is based on being able to exploit the hierarchical structure in the assets which
includes hierarchy in industries. However, in this case the weights in each industry is forced
to be equal.
Table 6 in Appendix A shows the values of these comparison measures. Figure 4 in Ap-
pendix A visualizes them by sorting portfolios according to ASR. We note that the correlation
between ASR and CER is 0.99, ASR and MDD -0.81, CER and MDD -0.85. The correlations
between comparison measures are high, which implies that the conclusions based on different
measure are robust. The ASRs shown in Table 6 show that IVP and EW portfolios have
much higher values than HERC portfolios in top 5 portfolios. However, HERC portfolios in
top 5 portfolios have ASR values much closer to IVP and EW portfolios in other construction
approach (see Tables 3 and 4). This implies that HERC algorithm may be able to capture
the hierarchy in industries.
In practice, fund manager may focus on particular industries. For this reason, we apply
HERC to six industries which are made by industry code A in China market. Specifically,
these six categories include finance, utilities, properties, conglomerates, industries and com-
merce. We note that in this case, we set the optimal number of clusters to be 5 and we do
not set weight bars to select stocks. The results are presented in Tables 7 and 8 in Appendix
B.

III.V. Summary

This section summarizes the results based on different ways of portfolio construction ap-
proaches. More specifically, we rank all portfolios by each comparison measure for each
portfolio construction method. Then, we average the ranks across different comparison mea-
sures for each portfolio construction method. Lastly, we average the ranks across all portfolio
construction methods.
The right panel ALL of Table 2 shows mean ranks for portfolios. We can see that in

16

Electronic copy available at: https://ssrn.com/abstract=3695598


terms of mean ranks (MEAN) across all comparison measures and all portfolio construction
methods, the top 5 portfolios are EW, IVP, ESW, SDW and VW. We note that ESW, SDW
and VW are all constructed based on Ward-linkage. In terms ASR and CER, the top 5
portfolios are the same as the ones by MEAN. The results do not show there exist any risk
measures which seem to perform better than other measures. In sum, we find that most
HERC portfolios do not perform better than IVP and EW in China stock market.
We also rank portfolios using construction 1 and 2 according to the averages of the com-
parison measures. The left panel of Table 2 shows the results. We can see that by the
MEAN which averages across all measures, the top 5 portfolios are SDW, IVP, ESW, EW
and EWW. That is, IVP and EW perform better than almost all HERC portfolios. This is
not consistent with the literature on other markets (see Raffinot 2017) . Also, we note that
except IVP and EW, the other three portfolios in top 5 are all HERC portfolios based on
Ward-linkage. The results again do not show there exists any risk measures which seem to
perform better than other measures.
In sum, HERC algorithm does not necessarily construct portfolios which are better than
IVP and EW portfolios in China market. HERC with Ward-linkage seems to perform better
than other linkages.

IV. Conclusion
This paper studies the performance of HERC algorithm in China stock market. More
specifically, we consider HECR algorithm with different combinations of risk measures and
linkage criteria and compare them to IVP and EW methods. Also, four variants of portfolio
constructions are considered. We find that most HERC portfolios are not able to beat IVP
and EW portfolios in terms of ASR, CER and MDD and HERC with Ward-linkage seems
to dominate the ones with other linkages. However, the results do not show that any risk
measures can beat other measures consistently.

17

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 1: Ranks of Portfolios in terms of Different Comparison Measures

Construction - 1 Construction - 2 Construction - 3 Construction - 4


ASR CER MDD ASR CER MDD ASR CER MDD ASR CER MDD

EW 4 5 2 4 4 1 1 6 8 1 1 10
IVP 3 3 7 2 1 3 0 5 11 0 0 12
ESS 20 20 20 14 16 15 10 17 16 2 2 6
ESW 5 4 3 0 0 7 5 1 1 12 12 11
ESA 17 14 15 18 18 18 11 18 18 10 10 0
ESC 10 9 8 8 7 11 15 9 7 14 13 15
SDS 19 19 19 15 14 17 13 16 17 5 4 3
SDW 1 1 4 3 3 5 2 0 4 16 15 13
SDA 13 15 14 21 21 21 20 21 21 8 7 8
SDC 11 10 10 10 9 9 9 8 6 13 14 18
CDRS 21 21 21 7 12 12 3 12 12 3 3 4
CDRW 2 2 5 5 5 2 21 4 0 15 16 14
CDRA 18 16 17 13 13 13 8 13 13 9 9 7
CDRC 9 8 9 12 10 0 18 7 5 19 20 20
VS 16 18 18 16 15 14 12 14 15 4 5 5
VW 0 0 6 6 6 4 4 2 3 17 17 16
VA 15 17 16 20 20 20 19 20 20 7 8 9
VC 12 11 11 11 11 10 17 11 9 18 18 19
EWS 14 13 13 17 17 16 7 15 14 6 6 2
EWW 6 6 0 1 2 6 6 3 2 20 19 17
EWA 7 12 12 19 19 19 14 19 19 11 11 1
EWC 8 7 1 9 8 8 16 10 10 21 21 21

Notes: EW and IVP denote equal-weight and Inverse-Variance portfolios, respectively. They are
highlighted by red to note that they are compared to HERC portfolios. We denote HERC portfolios
by the combination of risk metrics and linkages. The last letter denotes linkage. More specifically,
W denotes Ward linkage, A average linkage, S single linkage, C complete linkage. Other letters
before these letters denote risk metrics. ES represents expected shortfall, SD standard deviation,
CDR conditional drawdown riks and V variance. For example, VW denotes the combination of
risk measure, variance, and Ward linkage. 18

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 2: Mean Ranks of Portfolios in terms of Different Comparison Measures

Construction - 1 & 2 ALL


ASR CER MDD MEAN ASR CER MDD MEAN

EW 6 6 0 3 1 1 0 0
IVP 1 0 8 1 0 0 6 1
ESS 18 20 19 20 10 17 19 16
ESW 2 1 6 2 2 2 2 2
ESA 20 15 16 17 17 18 16 19
ESC 8 8 10 9 11 7 9 7
SDS 19 17 20 18 15 16 18 17
SDW 0 2 4 0 3 3 4 3
SDA 17 19 18 19 21 20 20 20
SDC 9 10 9 10 7 8 10 8
CDRS 13 18 17 15 6 11 13 9
CDRW 5 4 2 5 8 5 1 6
CDRA 14 12 13 13 12 12 14 13
CDRC 10 9 3 8 18 9 7 10
VS 16 16 15 16 13 15 17 14
VW 3 3 7 6 4 4 5 4
VA 21 21 21 21 20 21 21 21
VC 11 11 11 11 19 13 12 15
EWS 15 13 12 14 9 14 11 11
EWW 4 5 1 4 5 6 3 5
EWA 12 14 14 12 14 19 15 18
EWC 7 7 5 7 16 10 8 12

Notes: EW and IVP denote equal-weight and Inverse-Variance portfolios, respectively. They are
highlighted by red to note that they are compared to HERC portfolios. We denote HERC portfolios
by the combination of risk metrics and linkages. The last letter denotes linkage. More specifically,
W denotes Ward linkage, A average linkage, S single linkage, C complete linkage. Other letters
before these letters denote risk metrics. ES represents expected shortfall, SD standard deviation,
CDR conditional drawdown riks and V variance. For example, VW denotes the combination of
risk measure, variance, and Ward linkage.

19

Electronic copy available at: https://ssrn.com/abstract=3695598


For further research, one can do similar studies using different values of risk aversion in
computing CER which is important to consider for building profitable portfolios. Also, one
can replace the confidence level of 0.05 to be other values which is used for calculating
expected shortfall and conditional drawdown at risk. Also, whether the top 5 portfolios
created are profitable is also worthy of further studies. More research in the results shown
in Appendix A and B may be a good start.

REFERENCES
Anderson, Philip W (1972). “More is different”. Science 177.4047, pp. 393–396.
Bailey, David H and Marcos López de Prado (2013). “An open-source implementation of the
critical-line algorithm for portfolio optimization”. Algorithms 6.1, pp. 169–196.
Black, Fischer and Robert Litterman (1992). “Global portfolio optimization”. Financial an-
alysts journal 48.5, pp. 28–43.
DeMiguel, Victor, Lorenzo Garlappi, and Raman Uppal (2009). “Optimal versus naive diver-
sification: How inefficient is the 1/N portfolio strategy?” The review of Financial studies
22.5, pp. 1915–1953.
Jain, Prayut and Shashi Jain (2019). “Can machine learning-based portfolios outperform
traditional risk-based portfolios? the need to account for covariance misspecification”.
Risks 7.3, p. 74.
Levy, Moshe (2017). “Measuring Portfolio Performance: Sharpe, Alpha, or the Geometric
Mean?” Journal of Investment Management 15.3, pp. 1–17.
Lohre, Harald, Carsten Rother, and Kilian Axel Schäfer (2020). “Hierarchical risk parity: Ac-
counting for tail dependencies in multi-asset multi-factor allocations”. Machine Learning
and Asset Management, Forthcoming.
López de Prado, Marcos (2016). “Building diversified portfolios that outperform out of sam-
ple”. The Journal of Portfolio Management 42.4, pp. 59–69.

20

Electronic copy available at: https://ssrn.com/abstract=3695598


López de Prado, Marcos (2018). Advances in financial machine learning. John Wiley & Sons.
López de Prado, Marcos (2020a). “Clustering (Presentation Slides)”. Available at SSRN:
http://dx.doi.org/10.2139/ssrn.3512998.
López de Prado, Marcos (2020b). “Codependence (Presentation Slides)”. Available at SSRN:
http://dx.doi.org/10.2139/ssrn.3512994.
López de Prado, Marcos (2020c). Machine Learning for Asset Managers. Cambridge Univer-
sity Press.
López de Prado, Marcos and Michael J Lewis (2019). “Detection of false investment strategies
using unsupervised learning methods”. Quantitative Finance 19.9, pp. 1555–1565.
Mantegna, Rosario N (1999). “Hierarchical structure in financial markets”. The European
Physical Journal B-Condensed Matter and Complex Systems 11.1, pp. 193–197.
Markowitz, Harry (1955). The optimization of a quadratic function subject to linear con-
straints. Tech. rep. RAND CORP SANTA MONICA CA.
Markowitz, Harry (1959). Portfolio selection: Efficient diversification of investments. John
Wiley.
Markowitz, Harry and G Peter Todd (2000). Mean-variance analysis in portfolio choice and
capital markets. Vol. 66. John Wiley & Sons.
Papenbrock, Jochen (2011). “Asset Clusters and Asset Networks in Financial Risk Manage-
ment and Portfolio Optimization”. PhD thesis.
Pézier, Jacques and Anthony White (2008). “The relative merits of alternative investments
in passive portfolios”. The Journal of Alternative Investments 10.4, pp. 37–49.
Raffinot, Thomas (2017). “Hierarchical clustering-based asset allocation”. The Journal of
Portfolio Management 44.2, pp. 89–99.
Raffinot, Thomas (2018). “The hierarchical equal risk contribution portfolio”. Available at
SSRN 3237540.
Simon, Herbert A. (1962). “The Architecture of Complexity”. Proceedings of the American
Philosophical Society 106.6, pp. 467–482.

21

Electronic copy available at: https://ssrn.com/abstract=3695598


Tibshirani, Robert, Guenther Walther, and Trevor Hastie (2001). “Estimating the number
of clusters in a data set via the gap statistic”. Journal of the Royal Statistical Society:
Series B (Statistical Methodology) 63.2, pp. 411–423.

A. Appendix
The Appendix details the results in Table 1 by showing the values of comparison measures
for different portfolios (Tables 3 - 6) and visualizing the values (Figures 1 to 4). We note
that these results are valuable to consider a profitable strategies because the results show
the specific values of comparison measures, e.g., ASR.

22

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 3: Portfolio Construction - 1

ASR CER MDD

VW 0.00200 -0.00011 0.84489


SDW 0.00172 -0.00012 0.83808
CDRW 0.00141 -0.00013 0.83950
IVP 0.00132 -0.00013 0.84581
EW 0.00029 -0.00017 0.83174
ESW 0.00000 -0.00015 0.83669
EWW -0.00072 -0.00017 0.82483
EWA -0.00602 -0.00096 0.93705
EWC -0.00800 -0.00028 0.82622
CDRC -0.00984 -0.00032 0.87889
ESC -0.01191 -0.00035 0.86103
SDC -0.01274 -0.00036 0.87902
VC -0.01365 -0.00037 0.88551
SDA -0.02379 -0.00112 0.99047
EWS -0.02409 -0.00101 0.97600
VA -0.02544 -0.00116 0.99212
VS -0.02600 -0.00128 0.99321
ESA -0.02729 -0.00111 0.99054
CDRA -0.02936 -0.00115 0.99296
SDS -0.03397 -0.00132 0.99496
ESS -0.03691 -0.00137 0.99604
CDRS -0.04010 -0.00144 0.99710

Notes: The table shows comparison measures of a variety of portfolios constructed based on different
combination of risk metrics and linkages. We note that we set the optimal number of clusters to be
10 and sort comparison measures according to ASR. EW and VIP denote equal-weight and Inverse-
Variance portfolios, respectively. They are highlighted by red to note that they are compared to
HERC portfolios. We denote HERC portfolios by the combination of risk metrics and linkages. The
last letter denotes linkage. More specifically, W denotes Ward linkage, A average linkage, S single
linkage, C complete linkage. Other letters before these letters denote risk metrics. ES represents
expected shortfall, SD standard deviation, CDR conditional drawdown riks and V variance. For
example, VW denotes the combination of risk measure, variance, and Ward linkage.

23

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 4: Portfolio Construction - 2

ASR CER MDD

ESW 0.00231 -0.00012 0.85676


EWW 0.00134 -0.00014 0.85647
IVP 0.00132 -0.00013 0.84581
SDW 0.00127 -0.00015 0.85639
EW 0.00029 -0.00017 0.83174
CDRW -0.00091 -0.00018 0.83714
VW -0.00099 -0.00018 0.85197
CDRS -0.00613 -0.00096 0.93894
ESC -0.00865 -0.00031 0.91172
EWC -0.00993 -0.00033 0.89566
SDC -0.00998 -0.00034 0.89710
VC -0.01032 -0.00035 0.89868
CDRC -0.01214 -0.00034 0.80789
CDRA -0.02411 -0.00102 0.97603
ESS -0.02809 -0.00125 0.99622
SDS -0.02816 -0.00123 0.99636
VS -0.02819 -0.00124 0.99609
EWS -0.02850 -0.00126 0.99626
ESA -0.03300 -0.00140 0.99656
EWA -0.03386 -0.00142 0.99689
VA -0.03387 -0.00142 0.99694
SDA -0.03427 -0.00143 0.99698

Notes: The table shows comparison measures of a variety of portfolios constructed based on different
combination of risk metrics and linkages. We note that we set the optimal number of clusters to be
72 and sort comparison measures according to ASR. EW and IVP denote equal-weight and Inverse-
Variance portfolios, respectively. They are highlighted by red to note that they are compared to
HERC portfolios. We denote HERC portfolios by the combination of risk metrics and linkages. The
last letter denotes linkage. More specifically, W denotes Ward linkage, A average linkage, S single
linkage, C complete linkage. Other letters before these letters denote risk metrics. ES represents
expected shortfall, SD standard deviation, CDR conditional drawdown riks and V variance. For
example, VW denotes the combination of risk measure, variance, and Ward linkage.

24

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 5: Portfolio Construction - 3

ASR CER MDD

IVP 0.00132 -0.00013 0.84581


EW 0.00029 -0.00017 0.83174
SDW -0.00326 -0.00002 0.42659
CDRS -0.00574 -0.00094 0.93014
VW -0.01214 -0.00005 0.41350
ESW -0.01287 -0.00005 0.36246
EWW -0.01999 -0.00008 0.36788
EWS -0.02361 -0.00111 0.99013
CDRA -0.02397 -0.00101 0.97510
SDC -0.02462 -0.00026 0.78471
ESS -0.02526 -0.00115 0.99183
ESA -0.02583 -0.00128 0.99300
VS -0.02723 -0.00111 0.99033
SDS -0.02925 -0.00114 0.99278
EWA -0.03373 -0.00131 0.99477
ESC -0.03546 -0.00035 0.82423
EWC -0.03601 -0.00037 0.84069
VC -0.03607 -0.00038 0.83436
CDRC -0.03670 -0.00026 0.67312
VA -0.03728 -0.00137 0.99615
SDA -0.04039 -0.00145 0.99717
CDRW -0.05339 -0.00011 0.16858

Notes: The table shows comparison measures of a variety of portfolios constructed based on different
combination of risk metrics and linkages. We note that we set the optimal number of clusters to be
10 and sort comparison measures according to ASR. The weight bar is 1%. EW and IVP denote
equal-weight and Inverse-Variance portfolios, respectively. They are highlighted by red to note that
they are compared to HERC portfolios. We denote HERC portfolios by the combination of risk
metrics and linkages. The last letter denotes linkage. More specifically, W denotes Ward linkage, A
average linkage, S single linkage, C complete linkage. Other letters before these letters denote risk
metrics. ES represents expected shortfall, SD standard deviation, CDR conditional drawdown riks
and V variance. For example, VW denotes the combination of risk measure, variance, and Ward
linkage.

25

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 6: Portfolio Construction - 4

ASR CER MDD

IVP 0.00210 -0.00011 0.85629


EW 0.00123 -0.00014 0.84314
ESS -0.00215 -0.00019 0.81655
CDRS -0.00437 -0.00021 0.81407
VS -0.00505 -0.00023 0.81439
SDS -0.00548 -0.00023 0.81239
EWS -0.00566 -0.00024 0.81165
VA -0.00829 -0.00027 0.82884
SDA -0.00862 -0.00027 0.82540
CDRA -0.00887 -0.00028 0.81774
ESA -0.00906 -0.00030 0.79985
EWA -0.01024 -0.00030 0.80917
ESW -0.01978 -0.00047 0.84580
SDC -0.02048 -0.00055 0.92537
ESC -0.02202 -0.00051 0.87941
CDRW -0.02424 -0.00057 0.87823
SDW -0.02456 -0.00056 0.87496
VW -0.02534 -0.00058 0.88305
VC -0.02581 -0.00060 0.92722
CDRC -0.02618 -0.00061 0.92800
EWW -0.02675 -0.00061 0.89279
EWC -0.03052 -0.00068 0.93023

Notes: The table shows comparison measures of a variety of portfolios constructed based on different
combination of risk metrics and linkages. We note that we set the optimal number of clusters to be
5 and sort comparison measures according to ASR. EW and IVP denote equal-weight and Inverse-
Variance portfolios, respectively. They are highlighted by red to note that they are compared to
HERC portfolios. We denote HERC portfolios by the combination of risk metrics and linkages. The
last letter denotes linkage. More specifically, W denotes Ward linkage, A average linkage, S single
linkage, C complete linkage. Other letters before these letters denote risk metrics. ES represents
expected shortfall, SD standard deviation, CDR conditional drawdown riks and V variance. For
example, VW denotes the combination of risk measure, variance, and Ward linkage.

26

Electronic copy available at: https://ssrn.com/abstract=3695598


Figure 1: Portfolio Construction - 1

1.0

0.8
Standardized Values of CMs

0.6 ASR
CER
0.4 MDD

0.2

0.0

VA
CDRW

EWA

CDRA
CDRC

SDA
VW
SDW

IVP
EW
ESW
EWW

SDS
EWC

ESC
SDC
VC

EWS

VS

ESS
CDRS
ESA
Notes: The figure visualizes Table 3. We note that we standardize the comparison measures
to plot them nicely in the same figure. EW and IVP denote equal-weight and Inverse-
Variance portfolios, respectively. They are highlighted by red to note that they are compared
to HERC portfolios. We denote HERC portfolios by the combination of risk metrics and
linkages. The last letter denotes linkage. More specifically, W denotes Ward linkage, A
average linkage, S single linkage, C complete linkage. Other letters before these letters denote
risk metrics. ES represents expected shortfall, SD standard deviation, CDR conditional
drawdown riks and V variance. For example, VW denotes the combination of risk measure,
variance, and Ward linkage.

27

Electronic copy available at: https://ssrn.com/abstract=3695598


Figure 2: Portfolio Construction - 2

1.0

0.8
Standardized Values of CMs

0.6 ASR
CER
0.4 MDD

0.2

0.0

VA
CDRW

CDRA

EWA
CDRC

SDA
ESW
EWW
IVP
SDW
EW

VW
CDRS
ESC
EWC
SDC
VC

ESS
SDS
VS
EWS
ESA
Notes: The figure visualizes Table 4. We note that we standardize the comparison measures
to plot them nicely in the same figure. EW and IVP denote equal-weight and Inverse-
Variance portfolios, respectively. They are highlighted by red to note that they are compared
to HERC portfolios. We denote HERC portfolios by the combination of risk metrics and
linkages. The last letter denotes linkage. More specifically, W denotes Ward linkage, A
average linkage, S single linkage, C complete linkage. Other letters before these letters denote
risk metrics. ES represents expected shortfall, SD standard deviation, CDR conditional
drawdown riks and V variance. For example, VW denotes the combination of risk measure,
variance, and Ward linkage.

28

Electronic copy available at: https://ssrn.com/abstract=3695598


Figure 3: Portfolio Construction - 3

1.0

0.8
Standardized Values of CMs

0.6

0.4

0.2 ASR
CER
0.0 MDD

VA
CDRA

EWA

CDRW
CDRC

SDA
IVP
EW
SDW

VW
CDRS

ESW
EWW
EWS

SDC
ESS

VS
SDS

ESC
EWC
VC
ESA

Notes: The figure visualizes Table 5. We note that we standardize the comparison measures
to plot them nicely in the same figure. EW and IVP denote equal-weight and Inverse-
Variance portfolios, respectively. They are highlighted by red to note that they are compared
to HERC portfolios. We denote HERC portfolios by the combination of risk metrics and
linkages. The last letter denotes linkage. More specifically, W denotes Ward linkage, A
average linkage, S single linkage, C complete linkage. Other letters before these letters denote
risk metrics. ES represents expected shortfall, SD standard deviation, CDR conditional
drawdown riks and V variance. For example, VW denotes the combination of risk measure,
variance, and Ward linkage.

29

Electronic copy available at: https://ssrn.com/abstract=3695598


Figure 4: Portfolio Construction - 4

1.0

0.8
Standardized Values of CMs

0.6 ASR
CER
0.4 MDD

0.2

0.0
VA

CDRA

EWA

CDRW

CDRC
SDA
IVP
EW
ESS
CDRS
VS
SDS
EWS

ESW

SDW
VW
SDC
ESC

VC

EWW
EWC
ESA

Notes: The figure visualizes Table 6. We note that we standardize the comparison measures
to plot them nicely in the same figure. EW and IVP denote equal-weight and Inverse-
Variance portfolios, respectively. They are highlighted by red to note that they are compared
to HERC portfolios. We denote HERC portfolios by the combination of risk metrics and
linkages. The last letter denotes linkage. More specifically, W denotes Ward linkage, A
average linkage, S single linkage, C complete linkage. Other letters before these letters denote
risk metrics. ES represents expected shortfall, SD standard deviation, CDR conditional
drawdown riks and V variance. For example, VW denotes the combination of risk measure,
variance, and Ward linkage.

30

Electronic copy available at: https://ssrn.com/abstract=3695598


B. Portfolio Construction - 5
In practice, fund manager may focus on particular industries. For this reason, we apply
HERC to six industries which are made by industry code A in China market. Specifically,
these six categories include finance, utilities, properties, conglomerates, industries and com-
merce. We note that in this case, we set the ONC to be 5 and we do not set weight bars
to further select stocks. That is, we apply directly HERC with ONC of 5 to each industry.
One can further explore by choosing other values of ONC and see if the results are robust.
Table 7 shows the relative performance of the portfolios and Table 8 details the values of all
comparison measures. We note that some portfolios have profitable ASR. For instance, in
Finance industry the daily ASR for the top 5 portfolios are about 0.02. Further investigation
could be done here.

Table 7: Portfolio Construction - 5

Finance Utilities
ASR CER MDD ASR CER MDD

EW 19 17 5 1 1 5
IVP 17 14 6 0 0 6
ESS 7 6 4 2 3 0
ESW 20 19 7 12 12 12
ESA 13 11 9 7 7 7
ESC 14 15 12 15 13 13
SDS 15 10 0 6 6 2
SDW 16 16 11 17 16 16
SDA 1 1 16 14 11 11
SDC 4 5 20 21 21 21

31

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 7: Portfolio Construction - 5

CDRS 9 8 1 3 2 4
CDRW 18 20 8 13 15 15
CDRA 2 2 13 9 8 9
CDRC 6 12 19 19 19 19
VS 8 7 2 4 5 3
VW 10 13 17 16 17 17
VA 0 0 18 10 10 10
VC 3 4 21 20 20 20
EWS 12 9 3 5 4 1
EWW 21 21 10 11 14 14
EWA 5 3 15 8 9 8
EWC 11 18 14 18 18 18

Properties Conglomerates
ASR CER MDD ASR CER MDD

EW 1 1 6 0 1 0
IVP 0 0 8 1 0 1
ESS 2 2 5 2 2 2
ESW 14 12 12 11 13 13
ESA 3 3 2 12 7 3
ESC 18 17 14 7 11 11
SDS 6 4 11 3 3 5
SDW 13 14 15 19 19 19
SDA 7 6 0 10 8 4

32

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 7: Portfolio Construction - 5

SDC 19 19 19 9 14 14
CDRS 5 5 10 4 4 6
CDRW 12 13 13 18 18 18
CDRA 4 7 1 16 10 10
CDRC 21 20 21 8 15 15
VS 10 10 9 5 5 8
VW 16 15 17 21 21 21
VA 11 11 3 15 9 9
VC 17 18 18 14 17 17
EWS 9 9 7 6 6 7
EWW 15 16 16 20 20 20
EWA 8 8 4 17 12 12
EWC 20 21 20 13 16 16

Industrials Commerce
ASR CER MDD ASR CER MDD

EW 1 1 10 1 1 8
IVP 0 0 11 0 0 9
ESS 2 2 9 3 3 3
ESW 15 13 13 15 12 12
ESA 9 9 1 11 11 0
ESC 14 12 12 17 16 13
SDS 7 5 5 6 5 5
SDW 18 15 15 12 13 16

33

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 7: Portfolio Construction - 5

SDA 8 7 2 4 7 10
SDC 6 21 17 18 18 18
CDRS 4 4 7 2 2 6
CDRW 17 17 16 14 15 15
CDRA 10 8 3 8 8 7
CDRC 13 14 14 20 20 20
VS 3 3 8 7 6 2
VW 20 18 19 13 14 14
VA 12 11 4 9 9 11
VC 16 16 18 19 19 19
EWS 5 6 6 5 4 1
EWW 21 20 21 16 17 17
EWA 11 10 0 10 10 4
EWC 19 19 20 21 21 21

Table 8: Portfolio Construction - 5

Finance Utilities
ASR CER MDD ASR CER MDD

EW 0.00665 -4e-05 0.88189 -0.00052 -0.00018 0.83458


IVP 0.00847 -0.0 0.88996 0.00032 -0.00015 0.83806
ESS 0.0145 0.0001 0.88073 -0.0054 -0.00026 0.77875
ESW 0.00593 -0.00012 0.90261 -0.02596 -0.00076 0.94442
ESA 0.01172 4e-05 0.91377 -0.01476 -0.00041 0.85188

34

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 8: Portfolio Construction - 5

ESC 0.01159 -2e-05 0.93584 -0.02987 -0.00085 0.96073


SDS 0.01102 4e-05 0.8723 -0.00754 -0.00028 0.78891
SDW 0.01098 -4e-05 0.92488 -0.03022 -0.00103 0.98551
SDA 0.02195 0.00024 0.94006 -0.02846 -0.00066 0.94035
SDC 0.01911 0.0001 0.96876 -0.03578 -0.00127 0.994
CDRS 0.01345 8e-05 0.8754 -0.00608 -0.00026 0.79767
CDRW 0.00773 -0.00013 0.90572 -0.02668 -0.00096 0.98031
CDRA 0.02039 0.00021 0.93835 -0.02435 -0.00058 0.92997
CDRC 0.0161 1e-05 0.95742 -0.03288 -0.0012 0.99186
VS 0.01359 8e-05 0.87623 -0.00643 -0.00027 0.79517
VW 0.01295 0.0 0.9417 -0.03008 -0.00105 0.98579
VA 0.02556 0.00031 0.95479 -0.02461 -0.00061 0.93251
VC 0.0193 0.0001 0.97163 -0.03383 -0.00121 0.99219
EWS 0.01195 5e-05 0.87697 -0.00655 -0.00027 0.78401
EWW 0.00485 -0.00023 0.91384 -0.02464 -0.00093 0.9772
EWA 0.01851 0.00017 0.93987 -0.02361 -0.00058 0.92025
EWC 0.01272 -0.00011 0.93866 -0.03146 -0.00117 0.99063

Properties Conglomerates
ASR CER MDD ASR CER MDD

EW -0.00034 -0.0002 0.83216 -0.00229 -0.00025 0.81281


IVP 0.00161 -0.00015 0.84653 -0.00232 -0.00024 0.81797
ESS -0.00331 -0.00025 0.83167 -0.01451 -0.00051 0.8657
ESW -0.01046 -0.00049 0.9337 -0.02212 -0.00074 0.94541

35

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 8: Portfolio Construction - 5

ESA -0.0056 -0.00029 0.79944 -0.02228 -0.00065 0.90711


ESC -0.01815 -0.00067 0.97155 -0.02114 -0.00073 0.93333
SDS -0.00688 -0.00031 0.84817 -0.01719 -0.00057 0.91395
SDW -0.01033 -0.00063 0.97383 -0.02804 -0.001 0.98227
SDA -0.00692 -0.00032 0.79695 -0.0221 -0.00066 0.9113
SDC -0.02102 -0.00092 0.99297 -0.02195 -0.00088 0.9648
CDRS -0.00678 -0.00032 0.84788 -0.01764 -0.00059 0.9142
CDRW -0.00909 -0.00061 0.97142 -0.02722 -0.001 0.98197
CDRA -0.0067 -0.00032 0.7981 -0.02435 -0.00071 0.92683
CDRC -0.02234 -0.00096 0.99397 -0.02176 -0.00088 0.96552
VS -0.00804 -0.00035 0.8476 -0.01915 -0.00063 0.92232
VW -0.01085 -0.00064 0.97726 -0.02917 -0.00103 0.98555
VA -0.00821 -0.00035 0.81125 -0.02404 -0.0007 0.92593
VC -0.01781 -0.00084 0.99138 -0.0236 -0.00091 0.96899
EWS -0.00778 -0.00034 0.84428 -0.01974 -0.00064 0.92196
EWW -0.01066 -0.00066 0.9759 -0.02829 -0.00103 0.98416
EWA -0.00713 -0.00033 0.81203 -0.02555 -0.00073 0.93411
EWC -0.02228 -0.00096 0.9938 -0.02228 -0.00089 0.96729

Industrials Commerce
ASR CER MDD ASR CER MDD

EW 2e-05 -0.00017 0.82281 0.00406 -0.0001 0.87334


IVP 0.00049 -0.00015 0.834 0.00421 -9e-05 0.88054
ESS -0.00256 -0.00022 0.81187 0.00105 -0.00014 0.85271

36

Electronic copy available at: https://ssrn.com/abstract=3695598


Table 8: Portfolio Construction - 5

ESW -0.02112 -0.00096 0.96657 -0.01197 -0.00048 0.91376


ESA -0.01117 -0.00036 0.76626 -0.00766 -0.0003 0.80176
ESC -0.02051 -0.00096 0.96551 -0.01787 -0.00065 0.95375
SDS -0.00635 -0.00026 0.79171 -0.00035 -0.00015 0.8581
SDW -0.02915 -0.0012 0.9902 -0.01015 -0.00056 0.95869
SDA -0.01083 -0.00033 0.76974 5e-05 -0.00016 0.89073
SDC -0.00611 -0.00152 0.99096 -0.02043 -0.00085 0.97488
CDRS -0.00542 -0.00025 0.79773 0.00165 -0.00012 0.86305
CDRW -0.02842 -0.00123 0.99095 -0.01178 -0.00062 0.95826
CDRA -0.01165 -0.00035 0.78103 -0.00105 -0.00018 0.87035
CDRC -0.01865 -0.00115 0.98927 -0.02385 -0.00095 0.98359
VS -0.00434 -0.00024 0.80247 -0.00035 -0.00016 0.84978
VW -0.03238 -0.00126 0.99271 -0.01118 -0.0006 0.9572
VA -0.01224 -0.00037 0.78543 -0.00227 -0.00021 0.89209
VC -0.02355 -0.0012 0.9916 -0.02286 -0.00092 0.98102
EWS -0.00577 -0.00026 0.79724 -6e-05 -0.00015 0.84952
EWW -0.03435 -0.00129 0.99352 -0.0138 -0.0007 0.96044
EWA -0.01178 -0.00036 0.76086 -0.00584 -0.00029 0.85368
EWC -0.03168 -0.00128 0.99328 -0.0269 -0.00106 0.98889

37

Electronic copy available at: https://ssrn.com/abstract=3695598

You might also like