SUPTECH WORKSHOP III
Giudici P., Pagnottoni P., Polinesi G.
Network models to enhance automated cryptocurrency portfolio
management
Robot advisors, intro
FinTech innovations are increasing exponentially, for the
evolving technology on the supply side and for the shifting
of consumer preferences on the demand side
The total masses managed by the automatic consultancy
are estimated around 980 billion dollars in 2019, and
2,552 billion in 2023
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 2 / 15
Robot advisors and financial automation,
Pros & Cons
Advantages
Improved financial inclusion
Lower fees
High speed of service
Customized user experience
Disadvantages:
User may not understand portfolio construction
Portfolio models may be too simple
Contagion between asset returns increases
Portfolio allocation may not be complaint with investors’
risk profile
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 3 / 15
Contribution
Build similarity network models from the available asset
return data
Models that can incorporate multiple correlations
(contagion) between asset returns in portfolio allocation
The ultimate goal is to improve portfolio allocation and
risk compliance, taking systemic risk into account
Two main original contributions
We extend the application of similarity networks from
stock returns to Exchange Traded Fund returns
We propose an extension to Markowitz’ portfolio
allocation that takes network centrality and, therefore,
contagion, explicitly into account
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 4 / 15
The Random Matrix approach
RMT separates the “systematic part” of a signal embedded
into a return correlation matrix from the “noise”
Tests the eigenvalues of the correlation matrix:
λk < λk+1; k = 1, . . . , n against the null hypothesis that
they are from a random Wishart matrix R = 1
T
AAT
Let ri for i = 1, . . . , n be a time series of Cryptocurrency returns
and C be their correlation matrix. The RMT matrix is given by:
C∗
= VLVT
where V is the eigenvector matrix and
L =
0, λi < λ+
λi, λi ≥ λ+
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 5 / 15
Similarity Network
In a similarity network nodes represent asset returns and edges the
distance between adjacent nodes.
There exist different metrics to build distances between nodes: we apply
the Euclidean distance
dij = 2(1 − cij)
There exist different algorithms to simplify a similarity network: we apply
the Minimum Spanning Tree, that reduces the number of edges from
N(N − 1)/2 to N − 1.
In the MST, at each step, two cluster nodes li and lj are merged into a single
cluster if:
d(li; lj) = min(d(li; lj)
with the distance between clusters being defined as:
d(li; lj) = min(drq)
with r ∈ li and q ∈ lj
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 6 / 15
Centrality measures
To measure the importance of each node, we can use the
eigenvector centrality.
The importance of a node depends on the importance of
the nodes to which it is connected:
xi =
1
λ
N
j=1
ˆdi,jxx
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 7 / 15
Portfolio Construction
Differently from previous works which employ centrality
measures as an alternative measure of diversification risk,
we extend Markowitz’ approach using RMT and MST in the
optimisation function itself:
min
w
wT
C∗
w + γ
n
i=1
xiwi
n
i=1
wi = 1
µp ≥ 1
n
n
i=1
µi
wi ≥ 0
A high risk propensity (represented by a high value of γ)
translates in a portfolio composed by more systemically
risky assets, that lay in the central body of the network,
avoiding isolated cryptocurrencies.
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 8 / 15
Application
The data contains 10 time series of returns referred to
cryptocurrencies traded over the period 14 September
2017 - 17 October 2019 (764 daily observations)
Cryptocurrencies were selected in terms of market
capitalization
Portfolio returns are computed using the last month of
each time window
We use eleven months of observations as a look-back
period computing asset centrality and the consequent
portfolio weights
Then we calculate the return of each portfolio over the
next month rebalancing cryptocurrencies with the
retrieved weights. Finally we connect each monthly
portfolio performances from January 2018 to October 2019
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 9 / 15
Summary statistics
Cryptocurrency summary statistics over the period
14 September 2017 – 17 October 2019
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 10 / 15
MST networks
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 11 / 15
Portfolio Results - I, Cumulative P & L
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 12 / 15
Portfolio Results - II, Value at Risk (VaR)
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 13 / 15
Portfolio Results - III, cumulative returns
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 14 / 15
Portfolio Results - IV, highlights
·SUPTECH WORKSHOP III ·Network models to enhance automated cryptocurrency portfolio management 15 / 15