PDF Download

[Link]
30 December 2025
Total Citations: 8
Total Downloads: 227
.
.
Latest updates: hps://[Link]/doi/10.1145/3490354.3494381

.
.
Published: 03 November 2021
.
.
.
RESEARCH-ARTICLE

.
Citation in BibTeX format
Fund2Vec: mutual funds similarity using graph learning

.
.
ICAIF'21: 2nd ACM International
Conference on AI in Finance
VIPUL SATONE November 3 - 5, 2021
.
Virtual Event
DHRUV DESAI

.
.
Conference Sponsors:
.
DHAGASH MEHTA ACM
.
.
.
ICAIF '21: Proceedings of the Second ACM International Conference on AI in Finance (November 2021)
hps://[Link]/10.1145/3490354.3494381
ISBN: 9781450391481
.
 Fund2Vec: Mutual Funds Similarity using Graph Learning
Vipul Satone Dhruv Desai Dhagash Mehta ∗
vipulsatone@[Link] dhruvdesai@[Link] dhagashbmehta@[Link]
The Vanguard Group, Inc. The Vanguard Group, Inc. The Vanguard Group, Inc.

ABSTRACT lower expense ratio, brand name, etc.; recommending complement-

Identifying similar mutual funds with respect to the underlying ing items to the ones the investor is buying based on what other
portfolios has found many applications in financial services ranging investors bought together; in devising a tax loss harvesting strat-
from fund recommender systems, competitors analysis, portfolio an- egy; analysing diversification within a portfolio of funds, portfolio
alytics, marketing and sales, etc. The traditional methods are either analytics, etc.
qualitative, and hence prone to biases and often not reproducible, or, Even with its wide applicability, rigorously quantifying ’fund
are known not to capture all the nuances (non-linearities) among the similarity’ is a highly complex problem as it may involve emotional
portfolios from the raw data. We propose a radically new approach aspects and personal biases, in addition to multiple technical chal-
to identify similar funds based on the weighted bipartite network lenges such as defining an objective metric of similarity, identifying
representation of funds and their underlying assets data using a so- the variables to suite the chosen definition of similarity as well as
phisticated machine learning method called Node2Vec which learns potential nonlinear relationships between the chosen variables.
an embedded low-dimensional representation of the network. We In the financial domain, one of the most popular ways to find sim-
call the embedding Fund2Vec. Ours is the first ever study of the ilar funds is to look up their respective fund categorization provided
weighted bipartite network representation of the funds-assets net- by third-party data vendors such as Morningstar [2] and Lipper
work in its original form that identifies structural similarity among [3]. The fund categorizations are recommended by certain commit-
portfolios as opposed to merely portfolio overlaps. tees of experts who rely on various quantitative and qualitative
aspects of individual funds. However, with these categorizations,
CCS CONCEPTS an investor can only get a list of similar funds to a chosen one, but
not a ranking (i.e., which of the ones out of the list are more similar
• Applied computing → Economics.
to the chosen one than the others in the same list).
Various data-driven approaches for fund similarity have been
KEYWORDS proposed in the past [4–6]. Here, one uses raw data for each mu-
Mutual Funds, Machine Learning, Network Science, Node2Vec tual fund in the selected universe of funds, and compute the Eu-
ACM Reference Format: clidean distance, cosine similarity scores or Jaccard index in the
Vipul Satone, Dhruv Desai, and Dhagash Mehta . 2021. Fund2Vec: Mutual high-dimensional space of preselected variables. In more advanced
Funds Similarity using Graph Learning. In Proceedings of ICAIF-2021: ACM investigations, they also used different unsupervised clustering
International Conference on AI in Finance (ICAIF-2021). ACM, New York, NY, techniques [7] on the preselected variables to determine clusters of
USA, 8 pages. [Link] similar funds, and they compared these clusters with the respective
Morningstar categorizations to identify mismatches [6, 8–16] with
1 INTRODUCTION the data-driven clustering though it turned out that the reason
With the surge of popularity of the mutual funds and exchange- of mismatch was lack of important variables and improper use
traded funds (ETFs), many investment managers have launched and interpretation of the data-driven approaches [6, 17] (and in a
multiple funds with different investment strategies and philoso- published comments by Gambera, Rekenthaler and Xia in [6]).
phies yielding a plethora of funds available in the market. Such wide Here, we aim to go beyond third-party categorizations in search
variety of available products in the market bring a classic business for a data-driven approach to provide not only an objective way
problem: identifying similar (and dissimilar) products. Identifying of identifying similar funds as possible, but also to come up with
similar funds have multiple applications such as helping sales repre- a similarity ranking system. To that end, we argue that manually
sentative recommend [1] similar funds to the ones the investor has choosing any aggregate level fund composition related variables
in their portfolio while bringing in additional advantages such as (e.g., % equity, % fixed income, % allocation in specific sectors)
will invariably add certain inherent bias such as rigidly classifying
∗ Corresponding author certain companies in specific sector (e.g., [Link] Inc. being
rigidly classified as strictly in retail, or technology sector).
Permission to make digital or hard copies of all or part of this work for personal In the present work, we focus on the asset level information
or classroom use is granted without fee provided that copies are not made or
distributed for profit or commercial advantage and that copies bear this notice for each fund from a chosen universe: we view the fund and their
and the full citation on the first page. Copyrights for components of this work underlying assets data as a weighted bipartite network. Then, we
owned by others than the author(s) must be honored. Abstracting with credit is
permitted. To copy otherwise, or republish, to post on servers or to redistribute translate the fund similarity problem in the network science lan-
to lists, requires prior specific permission and/or a fee. Request permissions
from Permissions@[Link].
guage. Finally, we tailor-make a sophisticated machine learning
ICAIF'21, November 3–5, 2021, Virtual Event, USA algorithm called Node2Vec to capture underlying nonlinear rela-
© 2021 Copyright is held by the owner/author(s). Publication rights licensed to
ACM. tionship within the data to identify similar funds.
ACM ISBN 978-1-4503-9148-1/21/11…$15.00
[Link]
ICAIF-2021, 3-5 November, 2021, New York, NY Anonymous et al.

2 NETWORK OF FUNDS AND ASSETS projected onto a unipartite network of funds where the weights
We begin by viewing the mutual funds and their underlying hold- between nodes were the number of common stocks. Then, the
ings data as a weighted bipartite network where each mutual fund weighted network was clustered using the k-means and spectral
as well as each asset are represented as individual nodes. The link, clustering method for graph partitioning.
or lack of it, between each fund-asset pair of nodes yields that the
fund contains the asset at the specific snapshot in time. The weight 2.1 Bipartite Networks vs Projected Networks
on each link represents the percentage weight of the asset in the Traditionally, bipartite networks have been investigated only af-
respective fund at the snapshot. Figure 1 shows a simple exam- ter projecting it onto a unipartite network with one of the two
ple of fund-asset network with dummy data where 𝐹 1, . . . , 𝐹 15 are types of nodes: the weighted bipartite networks exhibit certain
hypothetical funds and 𝐴1, . . . , 𝐴33 are their hypothetical assets. peculiar characteristics compared to regular networks and hence
many of the network quantities defined for unipartite networks
may not be directly appropriate for weighted bipartite networks.
Hence, most of the existing literature on mutual funds and assets
networks have been limited to projected networks. However, the
original weighted bipartite network encodes the complete informa-
tion about the underlying data whereas there is invariably some
loss of information during any projection of bipartite network to a
unipartite network [25–28] though certain projections may retain
more information (e.g., Ref. [29, 30]) than others. In general, such
a projection typically induces a huge number of spurious edges
which limits on the computations of different quantities in practice.
Moreover, such large number of edges in the projected network
Figure 1: A hypothetical example of funds and assets net- may yield unique properties such as high clustering coefficient
work where green and red nodes correspond to funds and which may not be representative to corresponding characteristics
asset nodes, respectively. The thickness of links is propor- of the original network.
tional to the weight of the asset in the linked fund. In [31], the networks of mutual funds and assets for the US
based mutual funds were investigated keeping the bipartite network
structure intact for 2006Q3, 2007Q3, 2008Q3, i.e., before, during
Mutual funds and their underlying assets have been analysed
and after the 2008 financial crisis. The authors showed that the
from the network point of view in the past [18, 19]. In [20], a
degree distribution of the fund-nodes as well as asset-nodes both
network of stocks and mutual funds for a selected list of total 18
exhibited scale-free characteristics. Here, the definition of degree
stock-only mutual funds was investigated. For each of the 18 funds,
was that of an unweighted bipartite network. They also showed
the top 10 stocks according to the percentage weight in the fund
that while individual mutual funds have become more diversified
were considered. This network of 18 funds and 99 stocks was shown
portfolios after the financial crisis, the mutual funds have become
to exhibit small-world characteristic. Here, the original bipartite
more similar to each other, i.e., there is a large overlap (calculated
network was projected into an unweighted unipartite network
using asset weights within individual portfolios) among mutual
whose only nodes are stocks, and two stocks are connected if they
funds, giving rise to systematic risk.
belonged to the same mutual fund. The large average clustering
In the present work, we analyze the fund-asset network in its
coefficient (a quantity to analyze local structure of each node) [21]
original weighted bipartite network form without any approxima-
of the network was interpreted as higher probability than by chance
tion nor any projection. Then, the fund similarity problem translates
for any two stocks to be in the same mutual fund if they both are
into that of finding similar fund nodes on the network. We apply
present in another common mutual fund(s).
a recently proposed machine learning algorithm called Node2Vec
In [22] (and in [23]), a similar projection was used on the weighted
[32] to obtain a lower dimensional representation of the high-
bipartite networks of US based mutual fund and assets with weights
dimensional network data, called embedded representation of the
being total net assets, for duration from 1980Q1 to 2016Q4, first to
network. The lower dimensional representation captures most of
binary bipartite network (i.e., all non-zero weights are assumed to
the variance in the data, and more importantly the nonlinear rela-
have unit weight) and then projected to weighted unipartite net-
tionships among the raw input features. We then perform similarity
works where the only nodes are mutual funds. There, with the help
computation, such as cosine similarity, in the embedded representa-
of weighted degree centrality and weighted eigen centrality, the
tion to obtain similarity scores among different pairs of funds that
author concluded that the degree centrality has a negative and sta-
are a result of nonlinear relationship of the features.
tistically significant effect on mutual fund performance, i.e., mutual
funds with low portfolio similarity with their peers perform bet-
ter. More recently, such networks corresponding to pension funds 2.2 Fund-Asset Network and Data
(Italian market) are also analyzed and found exhibiting similar char- In order to have an entire universe of funds and assets associated
acteristics as aforementioned networks [24]. with them, we scraped data from the US Security and Exchange
In [5], a network of 551 Japanese mutual funds was constructed Commission (SEC) filings for form NPORT-P retrieved from the
using top 10 stocks of each fund. Here, the bipartite network was SEC Electronic Data Gathering, Analysis, and Retrieval (EDGAR)
Fund2Vec: Mutual Funds Similarity using Graph Learning ICAIF-2021, 3-5 November, 2021, New York, NY

Fund Nodes Asset nodes the vector length would be equal to vocabulary size and each ele-
Count 1093 16,138 ment of the vector represents a word in the vocabulary, Word2Vec
Mean number of edges 431.27447 29.20950 constructs a lower dimensional space that captures meaningful
Median number of edges 159 13 semantic and syntactic relationships between words.
Table 1: Basic network statistics for the fund-asset network. Here, first, each sentence is viewed as an directed subgraph
where each word is a node of the graph corresponding to the sen-
database, which is the primary system for submissions by compa- tence. Then, a shallow two-layer neural network is then used to
nies and others who are required by law to file information about get Word2Vec embeddings. Input to Word2Vec is large corpus of
their funds with the SEC. Access to EDGAR’s public database is free. sentences, i.e., sequence of words, and it outputs a vector space
Under the SEC regulations funds must report their portfolio and where each word is represented by a unique vector. In the vector
each of their portfolio holdings as of last business day or last calen- space, the words which share a common context in the corpus lie
dar day of the month. These reports disclose portfolio information close to each other. Word2Vec uses one of the two architectures,
as calculated by the fund for the reporting period’s ending net asset namely ’Continuous Bag-of-Words (CBOW)’ or ’Skip-gram’ model
value, reported on Form N-PORT must be filed with the commis- to find word embedding.
sion no later than 30 days after the end of each month. Information CBOW: In CBOW model, the surrounding context words are used
reported on Form N-PORT for the third month of each fund’s fiscal to predict the target word.
quarter is made publicly available 60 days after the end of the funds Skip-Gram: Here, the target word is fed as input, whereas the con-
fiscal quarter. Form N-PORT provides the following information text is generated as the output.
regarding the funds: total assets, including assets attributable to
miscellaneous securities reported for each underlying security, total 3.2 Node2Vec
liabilities, net assets, certain assets and liabilities reported in US Though Word2Vec very efficiently embeds data consisting of di-
dollars, securities lending, returns information, flow information, rected subgraphs, in other applications than NLP, graphs may arise
percentage value compared to net assets of the fund, etc. in different flavors such as (un)directed, (un)weighted, (a)cyclic, and
Based on the filing data, all funds were filtered down to a timelinse hence the methodology of Word2Vec cannot be directly applicable
falling in Quarter 1 of 2020. In order to have a defined universe we to embed these graphs.
chose to subset the data to all Equity Index Funds. SEC provides a Node2vec [32] solves this problem by using a clever trick that
list of funds, their share class as Class IDs, and the Series ID which ’generates’ directed subgraphs from the other types of graphs by
map the same fund of different share class to a single ID. starting from each node of the original graph and generating ran-
We create the bipartite network with each fund and assets as dom walks. The set of nodes in the original graph is then considered
nodes and percentage investment of a fund in an asset as weight on as the ’vocabulary’ and the directed path of each random walk is
the edge between the fund nodes. There are no edges between two considered as a ’sentence’. This new data, that now resembles the
funds or two assets as we only consider funds that do not invest in textual data, can be fed into Word2Vec to finally obtain the desired
another funds, and assets do not invest in another asset. We also do embedding. This algorithm returns a feature representation that
not consider the direction of investment from fund to asset. Hence, maximizes the likelihood of preserving network neighbourhoods
we get an undirected weighted bipartite network. of nodes in a 𝑑-dimensional feature space.
The output embeddings or representations will depend on net-
2.3 Data Cleaning work neighbourhoods, and the sampling of these network neigh-
The data collected was cleaned to keep only edges with non-negative bourhoods (i.e., the random walks) is a very important task. The
weights, and assets with a proper ISIN. Only those funds with at weight on the edge (in our case, the portfolio weights of the assets)
least 95% of the portfolio is present were retained. Certain funds, are the transition probabilities while walking from the pair of the
mostly due to the missing data, were not connected to the giant nodes connected by the edge. Below we list the most important
connected fund-asset component and were removed. The data after hyperparameters for Node2Vec:
this pre-processing contained about 1093 funds and 16,138 assets, (1) The Number of random walks (𝑟 ): the number of random
and the basic network statistics is as shown in Table 1. walks to be generated from each node in the graph;
(2) The length for each random walk (𝑙): the length (number of
3 METHODOLOGY hops) of each random walk from each node in the graph;
After its publication in 2016, Node2vec [32] has become one of the (3) 𝑝: the probability with which a random walk will return to
most used algorithms to learn lower dimensional representation for the node it already visited previously; and,
nodes in graph. Node2Vec is based on a word embedding technique (4) 𝑞: the probability with which a random walk will explore
called Word2Vec, hence, we first briefly describe Word2Vec. the unexplored part of the graph.
Note that these hyperparameters are specific to Node2Vec, in addi-
3.1 Word2Vec tion to the hyperaparameters such as context window size, number
Word2vec [33] is one of the most widely used word embedding of iterations for the Word2Vec algorithm.
techniques in the natural language processing (NLP) tasks. Instead Node2vec can be fine tuned to conform to already established
of the traditional methods to encode words in a large corpus of text equivalence in network science and can interpolate between the
in the numerical vectors form such as one-hot-encoding, where breadth-first sampling (BFS) (where nodes are sampled from the
ICAIF-2021, 3-5 November, 2021, New York, NY Anonymous et al.

immediate neighbourhood of the starting node) and depth-first V-Measure: V-measure is then defined as the harmonic mean be-
sampling (DFS) (where nodes are sequentially sampled from in- tween the homogeneity and completeness, i.e.,
creasing distance from the source), i.e., between homophily (nodes
(1+𝛽) (𝐻𝑜𝑚𝑜𝑔𝑒𝑛𝑒𝑖𝑡 𝑦) (𝐶𝑜𝑚𝑝𝑙𝑒𝑡𝑒𝑛𝑒𝑠𝑠)
that are highly connected should have their embedding vectors 𝑉 − 𝑚𝑒𝑎𝑠𝑢𝑟𝑒 = (𝛽∗𝐻𝑜𝑚𝑜𝑔𝑒𝑛𝑒𝑖𝑡 𝑦+𝐶𝑜𝑚𝑝𝑙𝑒𝑡𝑒𝑛𝑒𝑠𝑠) ,
close to each other) and structural equivalence (nodes that have
similar structural roles should be embedded closely together). where 𝛽 can be used to assign more weight to either homogene-
ity ( 𝛽 < 1) or completeness (𝛽 > 1). Since the number of ground
truth labels is only 2, clearly, the completeness will decrease as 𝐾
4 EXPERIMENTS AND RESULTS increases. We rather focus on maximizing homogeneity. To system-
We aim to find the lower (𝑑) dimensional embedding of the network atically weigh homogeneity significantly more than completeness
data using Node2Vec such that the data-points corresponding to in the computation of V-measure, we use 𝛽 = 0.01.
similar funds will be closer to each other. A systematic and objective
4.1.1 Hyperparameter Optimization. With the above metrics, we
evaluation of embeddings for an unsupervised technique is yet to
perform hyperparameter optimization for Node2Vec to obtain the
be rigorously defined in the literature [34]. Here, we device novel
desired embedding. The hyperparameters we tune are 𝑑, 𝑙, 𝑟 , 𝑝
metrics to evaluate the embeddings for the application at hand and
and 𝑞. These hyperparameters explore various strategies to sample
tune hyperparmaeters with respect to the new metrics. We call the
the network starting from interpolating between BFS and DFS, to
thus obtained best embedding Fund2Vec.
smaller and larger dimensional embeddings. We did not extensively
tune hyper-parameters of the underlying skip-gram model as our
4.1 ‘Bipartiteness’ as a Metric main aim in this investigation was to study the effects of different
random sampling strategies on embeddings.
We begin with the observation that the network at hand is inher-
In summary, below is the hyperparameter optimization process
ently bipartite, and any network learning algorithm should at least
we followed: For each hyperparameter point, we employ the K-
retain the ‘bipartiteness’ of the network in the lower-dimensional
means algorithmn (with Euclidean distance) for 𝐾 ranging from 2
representation. The embedded representation, however, is an ab-
to 10 (higher values of 𝐾 did not add further information in this
stract manifold and the data-points on this representation do not
case); we then compute the V-measure (𝛽 = 0.01) for each clustering
necessarily have network interpretation making it difficult to come
with respect to the ground truth labels (i.e., fund and asset); then, we
up with a metric to measure the retained bipartiteness.
chose the hyperparameter point at which the V-measure attained
Here, we propose a pragmatic approach: in our bipartite net-
the maximum value.
work, the only true “labels“ in the data are “fund“ and “asset“. In
Plots for a few values of hyperparameter points are shown in
the embedded representation should at least bring all the funds
Figure (2). The results for optimal values of 𝐾 for each hyperpa-
(assets) close to other funds (assets). We do not expect that in the
rameter point in our experiments and the corresponding values
embedded representation there will only be exactly 2 clusters, one
of various metrics are shown in table 2. For the data at hand, the
consisting of funds and the other consisting of assets. Then, after
best hyperparameter point in our experiments was: 𝑑 = 16, 𝑙 = 128,
training Node2Vec at each hyperparameter point, we employed the
𝑟 = 128, 𝑝 = 0.1 and 𝑞 = 5. The V-measure for these parameters
K-means [35] algorithm for a range of values for 𝐾 to cluster the
is 0.86 for 𝐾 = 5, yielding a clear separation between funds and
data and then look for the clustering which most clearly distinguish
assets in the optimal embedding. From here on, we call this 16-
funds from clusters, i.e., funds and assets are respectively clustered
dimensional representation as the [Link] of the five clusters,
together. A goal of hyperparameter optimization is then to seek for
two of them only consists of funds the remaining cluster consists
a hyperparameter point at which the mismatch between clustering
of a mixture of funds and assets with majority of data-points (>99
and the ground truth labels is minimized. In summary, we propose
%) corresponding to assets.
the bipartiteness as a metric to evaluate the embeddings learned by
Node2Vec, or any other related graph learning method, when the 1.0
original network is a bipartite network. 0.9
In practice, to measure how much bipartiteness is preserved by 0.8
V-measure (beta = 0.01)

the given embedded representation, we measure the mismatch be- 0.7

tween the clustering from K-means and the ground truth labels (i.e., 0.6
0.5
fund and asset) using a classic metric called the V-measure which
0.4
takes into account both homogeneity and completeness of the clus- 0.3 Embedding
ter. Below, we recall the definitions of homogeneity, completeness 0.2 8, 64, 64, 0.1, 5.0
and V-measure [36, 37]. 0.1 8, 128, 128, 0.1, 5.0
16, 128, 128, 0.1, 5.0
Homogeneity: When each of the clusters only contains data points 0.0 32, 128, 128, 5.0, 0.1
which are members of a single ground truth class, the clustering is 0.1
1 2 3 4 5 6 7 8 9 10
called homogeneous. If a cluster has data-points from two or more K (number of clusters)
different classes it is called a heterogeneous cluster.
Completeness: Completeness is defined as the measure of a cluster Figure 2: Plot for 𝐾 vs V-measure (𝛽 = 0.01) for a few hyper-
when a cluster contains all the elements of a class, as opposed to a parameter points.
cluster that fails to capture one or more elements of the class.
Fund2Vec: Mutual Funds Similarity using Graph Learning ICAIF-2021, 3-5 November, 2021, New York, NY

d l r p q Optimal 𝐾 V-measure (𝛽 = 0.01) Cluster number # # Funds # Assets Cluster Type

8 128 128 0.1 5 10 0.365 1 60 (0.60%) 9933 (99.40%) Asset Cluster
8 128 128 5 0.1 7 0.807 2 383 (100%) 0 (0.0%) Fund Cluster
8 64 64 0.1 5 9 0.421 3 1 (0.05%) 2041 (99.95%) Asset Cluster
8 64 64 5 0.1 11 0.782 4 643 (100%) 0 (0.0%) Fund Cluster
16 128 128 0.1 5 5 0.864 5 6 (0.14%) 4164 (99.86%) Asset Cluster
16 128 128 5 0.1 10 0.607 Total 1093 16138
16 64 64 0.1 5 10 0.829 Table 3: The table summarizes the distribution of funds
16 64 64 5 0.1 11 0.722 and assets among all the 5 clusters obtained using 𝐾-means
32 128 128 0.1 5 10 0.801 within Fund2Vec.
32 128 128 5 0.1 9 0.498
32 64 64 0.1 5 9 0.782
32 64 64 5 0.1 9 0.64
Table 2: The table shows the effect of the hyperparameters
on various metrics. The best hyperparameter point from our a novel optimization framework that accounts for both the explicit
runs is shown in bold text. and implicit relations in learning the embedded representations.
In the present work, while using Node2Vec we have devised the
above bipartiteness metric that explicitly help tuning the model
Figure 3 shows the T-SNE plot for all the data-points in this
towards an embedding that preserves the properties of the original
optimal embedding. Where most of the fund and asset nodes are
bipartite network. A comparison between our methodology, which
clearly in separate clusters. Cluster 2 and cluster 4 consist only
is more ad-hoc and pragmatic, and the one proposed in [38] which
of funds, in the T-SNE plot they lie very close to each other. The
extends the Node2Vec methodology from more theoretical point of
data-points which were missclassified indeed correspond to special
view, is beyond the scope of this paper.
types of funds or assets as discussed in the next subsection.
4.1.4 Analysis of miss-classified funds. Table 3 shows the number
Cluster 1 of funds and assets in each of the 5 clusters after K-means (𝐾 = 5)
75 Cluster 2
Cluster 3 clustering within Fund2Vec. In total there are 67 funds which were
Cluster 4 “misclassified“ as assets, i.e., these funds appeared in a cluster that
50 Cluster 5 mostly consists of assets otherwise. Interestingly, all these misclas-
25 sified funds exhibited specific and rather unique characteristic with
Dimension 2

respect to the other funds: all these funds were either currency-
0 hedged equity index funds which have assets from countries other
than the US. The assets within these funds only rarely appeared in
25 other funds than the currency-hedged funds making the currency-
50 hedged funds create almost star-graph structure around them. In a
future work, we will discuss this interesting set of outlier funds.
75

75 50 25 0 25 50 75 4.2 Baseline Benchmark

Dimension 1 Here, we devise another metric to perform sanity check to ensure
the final embedding indeed respects another objective ground truth
Figure 3: A TSNE plot of data-points in the optimal embed- information: since the funds’ filings also provide information about
ding. The blue cluster contains data points that correspond the benchmarks they track, Fund2Vec is expected to place funds
to asset nodes in the original network, whereas the red clus- which track the same benchmark closer than other groups of funds.
ter contains data points that correspond to fund nodes. We chose two of the most popular benchmarks, namely, S& P
500 (27 funds) and Russell 2000 (16 funds) and computed (1) the
mean cosine similarity, and (2) standard deviation of the cosine
4.1.2 Computational Details. The computation took, on an aver-
similarities among all the pairs of funds tracking the same bench-
age, around 6 hours to generate embeddings using Node2Vec on a
mark, using the original representation as well as Fund2Vec. We
regular single core machine with 60 GB RAM. We used Scikit-Learn
also computed both these quantities for cosine similarities between
[36] for the data pre-processing and K-means computation as well
each fund tracking the benchmark and all other funds not tracking
as to compute various evaluation metrics.
the benchmark. The results are shown in Table 4 clearly yielding
4.1.3 A few words on Node2Vec for bipartite networks. In [38], it that, on an average, Fund2Vec brought funds tracking the respective
was argued that Node2Vec (and other network embedding methods) index closer together compared to funds outside the benchmarks.
focused on homogeneous networks and overlooks special proper- Note that the relatively lower mean cosine similarity score among
ties of bipartite networks. The authors proposed a novel method funds within the Russell 2000 benchmark in the original space may
which first generates node sequences that preserve the long-tail dis- be due to the curse of dimensionality, whereas Fund2Vec being a
tribution of nodes in the original bipartite network, and then utilizes low-dimensional representation evades this problem.
ICAIF-2021, 3-5 November, 2021, New York, NY Anonymous et al.

0*Benchmark 0*# Funds

Fund2Vec Original Representation
Within Benchmark Outside Benchmark Within Benchmark Outside Benchmark
Mean Std Dev Mean Std Dev Mean Std Dev Mean Std Dev
S&P 500 27 0.99117 0.01307 0.29267 0.33861 0.96891 0.04357 0.17072 0.27734
Russell 2000 9 0.94660 0.04633 0.12425 0.27225 0.38490 0.23163 0.03346 0.09417
Table 4: The table shows the mean and standard deviation of cosine similarities between funds tracking the same benchmark
and those with funds not tracking the same benchmark.

4.3 Similarities in Different Representations 4.3.3 A Metric to Compare Different Embeddings. We compare the
Though the above defined bipartiteness metric provides a quantita- Jaccard indices using the above two scenarios for which the cosine
tive measure to evaluate embeddings with respect to the underlying similarity scores for each pair of funds are available. For each fund
network structure, it does not necessarily yield a ’goodness’ of the 𝐹𝑖 , we query top 𝑚 = 5, 10, 20, 50, most similar funds (with respect
embeddings in terms of similarity of funds. For an unsupervised to the cosine similarity computation) in the 16138-dimensional
problem such as the present one, one may not even expect a unique representation as well as in Fund2Vec. Hence, for each 𝐹𝑖 we get
and objective definition of similarity to begin with. Here, we fo- two sets of similar funds corresponding to two representations, each
cus on a few reasonable and mathematically rigorous definitions set consisting of 𝑚 funds. Then, we measure the overlap between
that suits our purpose: cosine similarity and Jaccard index. Then, these two sets using Jaccard index.
following the approach proposed in [39], instead of evaluating the For a fund, thus defined Jaccard index being close to 1 for a value
goodness of the new embedding compared to the original represen- of 𝑚 means that the embedded representation is providing the same
tation of the data, we tailor-make a few metrics to ensure that the list of similar funds as the original representation, whereas the
new embedding has indeed learned a different view of the data and, value close to 0 means that the two representations are providing
in turn, of similarity. different views of similarity of the underlying data. Clearly, the
larger the value of 𝑚 is, the greater are the chances of the same
4.3.1 Cosine Similarity. The cosine similarity is one of the most funds appearing in both the lists for a given fund with the extreme
popular scores to measure similarity for multi-dimensional data. case being 𝑚 = 𝑛 − 1, where 𝑛 is the total number of funds.
Here, each data point is considered as a point in the 𝑛-dimensional The distributions of Jaccard index over all the funds for each
space where 𝑛 is the number of variables or features. Then, the value of 𝑚 is shown in the figure 4, and further statistics for each
cosine similarity between the 𝑖-th and 𝑗-th funds as distribution is summarized in Table 5. The mean of Jaccard index
W𝐹𝑖 .W𝐹 𝑗 for different values of 𝑚 ranges from 0.47 to 0.53 yielding, on an
𝐶𝑖,𝑗 = , (1) average, a relatively low overlap between list of similar funds for
W𝐹 𝑖 W𝐹 𝑗 each fund. Figure 4 demonstrates low overlaps in lists of similar
where W𝐹𝑖 and W𝐹 𝑗 are vectors whose elements are weights of all funds between the two representations for smaller values of 𝑚.
the assets within funds 𝐹𝑖 and 𝐹 𝑗 , and ∥.∥ is the Euclidean norm.
𝐶𝑖,𝑗 ranges from −1 (i.e., 𝐹𝑖 and 𝐹 𝑗 are completely dissimilar) to 1 Number of funds querried (m) Mean Median Standard deviation
(i.e., 𝐹𝑖 and 𝐹 𝑗 are completely identical). 5 0.47 0.4 0.29
Here, we compute the cosine similarity for the following two 10 0.48 0.5 0.25
20 0.49 0.55 0.24
representations:
50 0.53 0.58 0.23
(1) We compute the cosine similarity between each pair of Table 5: Table for Jaccard index that measures the overlap,
funds in the original 16138-dimensional asset-weights space. for each fund 𝐹𝑖 , between the top 𝑚 most similar funds to 𝐹𝑖
Hence, the corresponding W𝐹𝑖 s in Eq. (1) are 16138-dimensional in the original representation and Fund2Vec.
vectors.
(2) We also compute the cosine similarity between each pair of
funds in Fund2Vec. Hence, the corresponding W𝐹𝑖 s in Eq. (1) 4.3.4 Cosine Similarities and Pearson Correlation. Another way to
are 16-dimensional vectors. find if we are able to capture a different view on similarity is to use
4.3.2 Jaccard Index. Jaccard index is used to find similarity be- the Pearson correlation between the two sets of similarity values.
tween two sample sets. It is defined as size (i.e., the number of For every pair of funds, we calculated cosine similarity in the 16138
elements) of the intersection between the two sets divided by the dimensional representation as well as within the Node2Vec embed-
size of union of two sets. Jaccard index close to being 1 indicates ding. If Node2Vec did not capture any new information than the
high similarity, and close to 0 indicates little similarity between the original representation, then these cosine similarity values should
two sets. be highly correlated. In our case, the Pearson correlation is 0.66,
Jaccard index between a pair of sets 𝐴 and 𝐵 is defined as which indicating that Fund2Vec indeed attempted to capture po-
tentially a different version of similarity (apparently, the structural
|𝐴 ∩ 𝐵| similarity) among funds. Figure 5 shows the scattered plot of the
𝐽𝐴,𝐵 = . (2)
|𝐴 ∪ 𝐵| cosine similarities for each pair of funds based on both the meth-
In other words, Jaccard index measures overlap between a pair of ods and exhibits a non-linear relationship between both the cosine
sets. similarities.
 Fund2Vec: Mutual Funds Similarity using Graph Learning ICAIF-2021, 3-5 November, 2021, New York, NY

5 Based on top 5 predictions Based on top 10 predictions

7
4 6
3 5
Density

Density
4
2 3
2
1
1
00.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2 00.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2
Based on top 20 predictions Based on top 50 predictions
5 3.0
4 2.5
2.0
Density

Density
3
1.5
2
1.0
1 0.5
00.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2 0.00.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2
Jaccard score Jaccard score
Figure 4: Statistics of Jaccard index for different values of 𝑚 = 5, 10, 20, 50.

5 DISCUSSION AND CONCLUSION

Due to the recent popularity of mutual funds and ETFs, a wide
variety of mutual funds have been available in the market. In this
paper, we pose the problem of comparing different mutual funds
as a similarity problem on a corresponding weighted bipartite net-
work where the network consists of funds and their underlying
assets as nodes and weights of individual asset in given fund as the
weights on the links between the fund and the asset. Ours is the
first study providing an investigation of the funds-assets network
while retaining the weighted bipartite characteristic of the network
completely intact.
Traditionally similarity of nodes of a network is defined with
respect to various quantities such as node centrality, degree cen-
trality, eigenvector centrality, clustering coefficients etc. Instead
of hand-crafting and computing such (potentially, infinitely many)
quantities, we employ a machine learning methodology which
yields a lower-dimensional manifold which captures the nonlinear
relationship among nodes and links. Then, the cosine similarity
between each pair of funds in the lower-dimensional representation
provides an objective and data-driven definition of fund similarity.
In particular, we employ a recently proposed network embedding
technique called Node2Vec, prompting the name of embedding of
funds-asset networks, Fund2Vec.
Evaluating performance of embeddings in representation learn-
ing paradigm is an open problem. In this paper, we also develop
Figure 5: Comparison between cosine similarities in differ- domain knowledge based methodology to evaluate the quality of
ent data representations. embeddings. Specifically, as a start, since the underlying network
is a bipartite network, we used bipartiteness as our first metric
to evaluate the quality of embedding: if the K-means clustering,
ICAIF-2021, 3-5 November, 2021, New York, NY Anonymous et al.

for 𝐾 raning from 2 to 10, performed within the given embedding [16] Daniele Lamponi. A data-driven categorization of investable assets. The Journal
is able to cluster fund nodes and asset nodes into separate clus- of Investing, 24(4):73–80, 2015.
[17] Dhagash Mehta, Dhruv Desai, and Jithin Pradeep. Machine learning fund cate-
ters, then the embedding preserves the original bipartiteness of the gorizations. ACM International Conference on AI in Finance 2020., 2020.
original network. The embeddings that mix fund nodes with asset [18] Franklin Allen and Ana Babus. Networks in finance. The network challenge:
strategy, profit, and risk in an interlinked world, 367, 2009.
nodes, and vice versa, when the K-means clustering is performed, [19] Anna Maria D’Arcangelis and Giulia Rotundo. Complex networks in finance. In
are worse than the former embedding. We used this metric for Complex networks and dynamics, pages 209–235. Springer, 2016.
hyperparameter optimization. [20] Rafael Solis. Visualizing stock-mutual fund relationships through social network
analysis. Global Journal of Finance and Banking Issues, 3(3), 2009.
As a sanity check with respect to available objective ground truth [21] Duncan J Watts and Steven H Strogatz. Collective dynamics of ‘small-
information, we also computed cosine similarities between funds world’networks. nature, 393(6684):440–442, 1998.
tracking the same index (S& P 500 and Russell 2000) to demon- [22] Shema F Mitali. Common holdings and mutual fund performance. Available at
SSRN 3448494, 2019.
strate that Fund2Vec indeed place these funds closer to each other [23] Li Lin and Xin-Yu Guo. Identifying fragility for the stock market: Perspective
compared to other group of funds. from the portfolio overlaps network. Journal of International Financial Markets,
Institutions and Money, 62:132–151, 2019.
We also used Jaccard metric and Pearson correlation to demon- [24] Anna Maria D’Arcangelis, Susanna Levantesi, and Giulia Rotundo. A complex
strate that Fund2Vec indeed provides a different representation of networks approach to pension funds. Journal of Business Research, 2019.
data and, more importantly, different similarity scores. In particular, [25] Stephen P Borgatti and Martin G Everett. Network analysis of 2-mode data.
Social networks, 19(3):243–270, 1997.
Fund2Vec not only (implicitly) captures the similarity measures such [26] Stephen P Borgatti. 2-mode concepts in social network analysis. Encyclopedia of
as the portfolio overlaps between the pairs of funds, but, more im- complexity and system science, 6:8279–8291, 2009.
portantly, it also captures structural similarities between funds, i.e., [27] Stephen P Borgatti and Daniel S Halgin. Analyzing affiliation networks. The
Sage handbook of social network analysis, 1:417–433, 2011.
two mutual funds are similar if they have similar network structure. [28] Matthieu Latapy, Clémence Magnien, and Nathalie Del Vecchio. Basic notions
To the best of our knowledge, this is the first work that considers for the analysis of large two-mode networks. Social networks, 30(1):31–48, 2008.
[29] Tao Zhou, Jie Ren, Matúš Medo, and Yi-Cheng Zhang. Bipartite network projec-
structural similarity of mutual funds and assets networks, while tion and personal recommendation. Physical review E, 76(4):046115, 2007.
also saving time to manually creating individual network quantities [30] Jaime F Lavin, Mauricio A Valle, and Nicolás S Magner. Modeling overlapped
to define structural similarity by hand. mutual funds’ portfolios: A bipartite network approach. Complexity, 2019, 2019.
[31] Danilo Delpini, Stefano Battiston, Guido Caldarelli, and Massimo Riccaboni.
Another benefit that Fund2Vec provides over the other methods Systemic risk from investment similarities. PloS one, 14(5), 2019.
is that it considers the entire network while finding out embed- [32] Aditya Grover and Jure Leskovec. Node2vec: Scalable feature learning for net-
ding for each node. As a result we find similar funds even if they works. In Proceedings of the 22nd ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, KDD 2016, page 855–864, New York, NY,
have less nodes in common. We anticipate that Fund2Vec may have USA, 2016. Association for Computing Machinery.
many more applications in addition to the ones mentioned in the [33] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Dis-
tributed representations of words and phrases and their compositionality. In
paper due to the captured structural similarity, e.g., Fund2Vec may C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger,
also be used to help a portfolio constructor identify funds similar editors, Advances in Neural Information Processing Systems 26, pages 3111–3119.
to the given one but only from within a certain theme such as Curran Associates, Inc., 2013.
[34] Alexandru Mara, Jefrey Lijffijt, and Tijl De Bie. Evalne: a framework for evaluating
Environment, Social and Governance (ESG), retail, fintech, etc. network embeddings on link prediction. arXiv preprint arXiv:1901.09691, 2019.
[35] J. MacQueen. Some methods for classification and analysis of multivariate
observations. In Proceedings of the Fifth Berkeley Symposium on Mathematical
Statistics and Probability, Volume 1: Statistics, pages 281–297, Berkeley, Calif., 1967.
REFERENCES University of California Press.
[1] Charu C Aggarwal et al. Recommender systems. Springer, 2016. [36] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blon-
[2] Morningstar categorization. del, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Courna-
[3] Lipper u.s. fund classification, 2020. peau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning
[4] Achla Marathe and Hany A Shawky. Categorizing mutual funds using clusters. in Python. Journal of Machine Learning Research, 12:2825–2830, 2011.
Advances in Quantitative analysis of Finance and Accounting, 7(1):199–204, 1999. [37] Andrew Rosenberg and Julia Hirschberg. V-measure: A conditional entropy-based
[5] Takumasa Sakakibara, Tohgoroh Matsui, Atsuko Mutoh, and Nobuhiro Inuzuka. external cluster evaluation measure. In Proceedings of the 2007 joint conference
Clustering mutual funds based on investment similarity. Procedia Computer on empirical methods in natural language processing and computational natural
Science, 60:881–890, 2015. language learning (EMNLP-CoNLL), pages 410–420, 2007.
[6] John A Haslem and Carl A Scheraga. Morningstar’s classification of large-cap [38] Ming Gao, Leihui Chen, Xiangnan He, and Aoying Zhou. Bine: Bipartite net-
mutual funds. The Journal of Investing, 10(1):79–89, 2001. work embedding. In The 41st international ACM SIGIR conference on research &
[7] Fan Cai, Nhien-An Le-Khac, and Tahar Kechadi. Clustering approaches for development in information retrieval, pages 715–724, 2018.
financial data analysis: a survey. arXiv preprint arXiv:1609.08520, 2016. [39] Xin Liu, Konstantinos Pelechrinis, and Alexandros Labrinidis. hood2vec: Identify-
[8] Athanasios Orphanides et al. Compensation incentives and risk taking behavior: ing similar urban areas using mobility networks. arXiv preprint arXiv:1907.11951,
evidence from mutual funds. Citeseer, 1996. 2019.
[9] Stephen J Brown and William N Goetzmann. Mutual fund styles. Journal of
financial Economics, 43(3):373–399, 1997.
[10] Dan DiBartolomeo and Erik Witkowski. Mutual fund misclassification: Evidence
based on style analysis. Financial Analysts Journal, 53(5):32–43, 1997.
[11] Edwin J Elton, Martin J Gruber, and Christopher R Blake. Incentive fees and
mutual funds. The Journal of Finance, 58(2):779–804, 2003.
[12] Moon Kim, Ravi Shukla, and Michael Tomas. Mutual fund objective misclassifi-
cation. Journal of Economics and Business, 52(4):309–323, 2000.
[13] Arturo Rodríguez Castellanos and Belén Vallejo Alonso. Spanish mutual fund
misclassification: Empirical evidence. The Journal of Investing, 14(1):41–51, 2005.
[14] David Moreno, Paulina Marco, and Ignacio Olmeda. Self-organizing maps could
improve the classification of spanish mutual funds. European Journal of Opera-
tional Research, 174(2):1039–1054, 2006.
[15] Debashis Acharya and Gajendra Sidana. Classifying mutual funds in india: Some
results from clustering. Indian Journal of Economics and Business, 6(1):71–79,
2007.