Portfolio Optimization

Financial Data: Graphs

Daniel P. Palomar (2025). Portfolio Optimization: Theory and Application.

Cambridge University Press.

[Link]

Latest update: April 14, 2025

Outline

1 Graphs

2 Learning graphs

3 Learning structured graphs

4 Learning heavy-tailed graphs

5 Learning time-varying graphs

6 Summary
 Abstract
Graphs are a powerful mathematical tool for representing data and relationships between
entities in various fields, including biology, finance, machine learning, and social networks.
With the increasing availability of large datasets, graph-based analysis is crucial for
understanding the structure of networks generating the data. In practice, the underlying
graph structure is often unknown and must be inferred from the data. Over the past two
decades, numerous graph learning algorithms have been proposed, with recent advances
focusing on financial data applications (Palomar 2025, chap. 5).

Portfolio Optimization Financial Data: Graphs 3 / 58

Outline

1 Graphs

2 Learning graphs

3 Learning structured graphs

4 Learning heavy-tailed graphs

5 Learning time-varying graphs

6 Summary
Graphs

Graphs in Various Domains:

Serve as a fundamental tool for modeling data across diverse applications.
Enable understanding of large networks’ structure and data geometry visualization.
Graph Examples Across Domains:
Social Media: Models individual behaviors and influences using online activities.
Brain Activity: Correlates brain sensors, utilizing fMRI data.
Financial Stocks: Shows company interdependencies in markets, with stock prices and
volumes.
Currency & Cryptocurrency: Summarizes interdependencies in foreign and crypto
markets, using economic data.

Portfolio Optimization Financial Data: Graphs 5 / 58

Examples of graphs in different applications

Social media graph Brain activity graph

Financial stock graph Financial currency graph

Portfolio Optimization Financial Data: Graphs 6 / 58

Terminology
The basic elements of a graph are

nodes: corresponding to the entities or variables; and

edges: encoding the relationships between entities or variables.

Graph Mathematical Structure:

Nodes: V = {1, 2, 3, . . . , p}
Edges: E = {(1, 2), (1, 3), . . . , (i, j), . . .}
Weight Matrix W : strength of relationships between nodes.
Portfolio Optimization Financial Data: Graphs 7 / 58
Graph matrices
Adjacency Matrix W : Directly characterizes a graph.
[W ]ij = wij if (i, j) ∈ E, else 0; Wij = 0 if i = j.
Symmetric W implies undirected graph.
Connectivity Matrix C: Binary version of the adjacency matrix.
[C]ij = 1 if (i, j) ∈ E, else 0; Cij = 0 if i = j.
Describes adjacency matrix connectivity pattern.
Degree Matrix D: Diagonal matrix with node degrees d = (d1 , . . . , dp ).
P
Degree di = j Wij .
D = Diag(W 1).
Laplacian Matrix L: Defined as L = D − W .
Symmetric, positive semidefinite: L ⪰ 0.
Zero eigenvalue with all-one vector 1: L1 = 0.
Degree vector on diagonal: diag(L) = d .
Zero eigenvalues number indicates graph’s connected components.
Measures graph signal smoothness: xT Lx = 21 i,j Wij (xi − xj )2 .
P

Portfolio Optimization Financial Data: Graphs 8 / 58

Example of a toy undirected graph
Graph Definition:
Nodes: V = {1, 2, 3, 4}
Edges: E = {(1, 2), (2, 1), (1, 3), (3, 1), (2, 3), (3, 2), (2, 4), (4, 2)}
Weights: w12 = w21 = 2, w13 = w31 = 2, w23 = w32 = 3, w24 = w42 = 1

The connectivity, adjacency, and Laplacian graph matrices are

     
0 1 1 0 0 2 2 0 4 −2 −2 0
1 0 1 1 2 0 3 1 −2 6 −3 −1
C = , W =  , L =  .
     
1 1 0 0 2 3 0 0 −2 −3 5 0
0 1 0 0 0 1 0 0 0 −1 0 1
Portfolio Optimization Financial Data: Graphs 9 / 58
Outline

1 Graphs

2 Learning graphs

3 Learning structured graphs

4 Learning heavy-tailed graphs

5 Learning time-varying graphs

6 Summary
Learning graphs

Graph Structure in Applications:

Directly observable in social networks (nodes as users, edges as friendships).
Must be inferred in gene graphs, brain activity graphs, financial graphs.
Graph Learning Methods:
Range from heuristic techniques to statistically sound approaches.
Utilize physical interpretation or estimation theory.
Data Matrix as Starting Point:

X = [x 1 , x 2 , . . . , x p ] ∈ Rn×p ,

Columns represent signals of variables/nodes.

p: number of variables/nodes, n (T in financial time series): number of observations.
Goal in Graph Learning:
Transition from data matrix X to graph description G = (V, E, W ).

Portfolio Optimization Financial Data: Graphs 11 / 58

Learning graphs from data

Portfolio Optimization Financial Data: Graphs 12 / 58

Learning graphs from data
Illustration of graph learning for the “two-moon” dataset:

Portfolio Optimization Financial Data: Graphs 13 / 58

Development in graph learning

Pioneering Work:
(Mantegna 1999): First data-driven graphs in financial markets using correlation graphs.
Foundational Textbooks:
(Lauritzen 1996; Kolaczyk 2009): Comprehensive understanding of graph theory.
Introductory and Overview Articles:
(Mateos et al. 2019; Dong et al. 2019): For basics and overview of graph learning.
Basic Graph Learning Algorithms:
Found in (Lake and Tenenbaum 2010; Egilmez, Pavez, and Ortega 2017; Zhao et al.
2019).

Portfolio Optimization Financial Data: Graphs 14 / 58

Development in graph learning

Structured Graph Learning:

Spectral constraints approach: (Kumar et al. 2019, 2020).
Sparsity in graphs: (Ying, Cardoso, and Palomar 2020).
Convex formulation for bipartite graphs: (Cardoso, Ying, and Palomar 2022b).
Graph Learning with Financial Data:
General guidelines: (Cardoso and Palomar 2020).
Learning under heavy tails: (Cardoso, Ying, and Palomar 2021).
Bipartite-structured graphs for clustering: (Cardoso, Ying, and Palomar 2022b); overview
in (Cardoso, Ying, and Palomar 2022a).
Comprehensive Literature Overview:
On financial graphs in the past two decades: (Marti et al. 2021).

Portfolio Optimization Financial Data: Graphs 15 / 58

Learning graphs from similarity measures

Graph Inference based on Adjacency Matrix W :

Use similarity or scoring functions to measure connectivity strength.
Different functions lead to different graphs.
Illustrative Simple Methods:
Thresholded Distance Graph:
Connect nodes i and j (wij = 1) if ∥x i − x j ∥2 ≤ γ (threshold).
Otherwise, wij = 0.
Gaussian Graph:
Connect all pairs i ̸= j with weights:
 
∥x i − x j ∥2
wij = exp − ,
2σ 2

σ 2 controls neighborhood size.

Portfolio Optimization Financial Data: Graphs 16 / 58

Learning graphs from similarity measures

Illustrative Simple Methods:

k-Nearest Neighbors (k-NN) Graph:
Connect nodes i and j (wij = 1) if x i is among k closest to x j or vice-versa.
Otherwise, wij = 0.
Feature Correlation Graph:
Use pairwise feature correlation for i ̸= j:

wij = xTi x j .

If signals normalized (∥x i ∥2 = 1), Euclidean distance relates to correlation:

∥x i − x j ∥2 = 2 × (1 − xTi x j ).

Limitations of Heuristic Methods:

Measure connectivity independently for each pair.
May not perform well, especially for time series data.
Better to measure connectivity jointly for all pairs.

Portfolio Optimization Financial Data: Graphs 17 / 58

Learning graphs from signal smoothness

Variance of Graph Signals:

1
For p-dimensional signal x on p nodes: xT Lx = Wij (xi − xj )2 .
P
2 i,j
For n observations in data matrix X ∈ Rn×p :
n
X  
(x (t) )T Lx (t) = Tr XLXT .
t=1

Graph Learning Problem Formulation:

Minimize signal variance to find the generating graph.  
Choose between graphs L1 and L2 by comparing variances: Tr XL1 XT
 
vs. Tr XL2 XT .
Graph Learning Optimization:
Determine graph G that minimizes signal variance.
Include regularization for graph properties (sparsity, energy, volume).

Portfolio Optimization Financial Data: Graphs 18 / 58

Learning graphs from signal smoothness
Problem Formulation in Terms of L:
 
minimize Tr XLXT + γhL (L)
L⪰0
subject to L1 = 0, Lij = Lji ≤ 0, ∀i ̸= j,

γ: regularization level, hL (L): regularization function.

Problem Formulation in Terms of W :
1
minimize 2 Tr(W Z ) + γhW (W )
W
subject to diag(W ) = 0, W = W T ≥ 0,

hW (W ): regularization function for adjacency matrix.

Convexity and Complexity:
Formulations are convex if regularization terms are convex.
Formulation in terms of L not more complex than W despite L ⪰ 0 constraint.
Portfolio Optimization Financial Data: Graphs 19 / 58
Learning graphs from graphical model networks

Graph Data as Multivariate Distribution:

Graph signals x (t) assumed to follow a distribution, e.g., Gaussian: x (t) ∼ N (µ, Σ).
T : number of observations.
Sample Covariance Matrix S:
T
1 X 1
T
(x (t) − µ)(x (t) − µ)T =


S= X − X̄ X − X̄ ,
T t=1 T

X̄: matrix with mean vector µ along each row.

Graphical Model Estimation:
Inverse sample covariance matrix S −1 historically used to determine graph structure.

Portfolio Optimization Financial Data: Graphs 20 / 58

Graph learning methods
Correlation Networks:
Use pairwise correlations to measure node similarity.
Drawback: High correlation may result from common dependencies on other nodes.
Partial Correlation Networks:
Measure direct dependency between two nodes, conditioned on other nodes.
Information contained in precision matrix Θ = Σ−1 . p
Conditional correlation between nodes i and j: −Θij / Θii Θjj .
Conditional independence if Θij = 0.
Nonzero off-diagonal entries of Θ indicate graph edges.
Graphical LASSO (GLASSO):
Estimates sparse precision matrix Θ = Σ−1 .
Regularized maximum likelihood estimation:

maximize log det(Θ) − Tr(ΘS) − ρ∥Θ∥1,off ,

Θ≻0

ρ: controls sparsity level.

Portfolio Optimization Financial Data: Graphs 21 / 58
Graph learning methods

Laplacian-structured GLASSO:
Adapts GLASSO for Gaussian Markov random fields (GMRFs) with Laplacian constraints:

maximize log gdet(L) − Tr(LS) − ρ∥L∥1,off ,

L⪰0

Constraints ensure Laplacian matrix properties.

Sparse GMRF Graphs:
Addresses the limitation of dense graphs produced by ℓ1 -norm regularization:

maximize log gdet(L) − Tr(LS) − ρ∥L∥0,off ,

L⪰0

∥L∥0,off : promotes true sparsity.

Reweighted ℓ1 -norm regularization:
Approximates ∥L∥0,off with a smooth concave function, then a convex weighted ℓ1 -norm.
Effective for sparse graph learning.

Portfolio Optimization Financial Data: Graphs 22 / 58

Numerical experiments
Data Overview:
Three years of S&P 500 stock price data (2016-2019) from Industrials, Consumer
Staples, and Energy sectors.
Data matrix X ∈ RT ×N with log-returns of N assets.
Normalization:
Assets normalized to have volatility one.
Normalization equivalent to using correlation matrix instead of covariance matrix.
Ensures uniform dynamic ranges and unit measurements.
Financial Assets Correlation:
High correlation typically due to market or other factors.
Precision and Laplacian matrices interpret partial correlation, removing common factor
effects.
Graph Learning Methods for Time Series:
GMRF-based methods preferred for enforcing sparsity:
Laplacian-structured GLASSO with ℓ1 -norm.
Sparse GMRF graph with ℓ0 penalty, practically solved via reweighted ℓ1 -norm.
Reweighted ℓ1 -norm iterative method shows superior performance.
Portfolio Optimization Financial Data: Graphs 23 / 58
Numerical experiments
Effect of sparsity regularization term on financial graphs:

GMRF graph (l1−norm) GMRF graph (reweighted l1−norm)

Portfolio Optimization Financial Data: Graphs 24 / 58

Outline

1 Graphs

2 Learning graphs

3 Learning structured graphs

4 Learning heavy-tailed graphs

5 Learning time-varying graphs

6 Summary
Learning structured graphs

Challenges with Structured Graphs:

Learning graphs with specific structures is generally NP-hard.
Designing a universal algorithm for structured graphs is challenging.

Common Types of Structured Graphs:

Multi-component or k-component graph: Contains clusters, useful for classification.
Regular graph: Each node has the same number of neighbors, useful for balanced
graphs.
Modular graph: Satisfies shortest path distance properties among triplets of nodes,
useful for social network analysis.
Bipartite graph: Two types of nodes with inter-connections only.
Grid graph: Nodes follow a rectangular grid or two-dimensional lattice.
Tree graph: Undirected graph with exactly one path between any two vertices,
resembling a branching structure.

Portfolio Optimization Financial Data: Graphs 26 / 58

Learning structured graphs
Types of structured graphs:

Multi−component graph Regular graph Modular graph

Bipartite graph Grid graph Tree graph

Portfolio Optimization Financial Data: Graphs 27 / 58

Learning structured graphs

Controlling Structural Constraints:

Some constraints, like those in grid or regular graphs, are simpler to manage by fixing
adjacency and Laplacian matrix elements or controlling node degrees.
Constraints like the number of neighbors involve nonconvex conditions.

Spectral Properties for Structural Constraints:

More complex structural constraints can be characterized by spectral properties of
Laplacian and adjacency matrices.
Spectral properties can be enforced in graph learning formulations to achieve desired
structures.

Portfolio Optimization Financial Data: Graphs 28 / 58

k-component graphs
Characterization:
A k-component graph is defined by its Laplacian matrix having k zero eigenvalues.
This indicates the graph is divided into k distinct clusters or components.
Eigenvalue Decomposition:
Laplacian matrix L decomposed as L = UDiag(λ1 , λ2 , . . . , λp )UT .
U contains eigenvectors, λ1 ≤ λ2 ≤ · · · ≤ λp are eigenvalues in increasing order.
For k-component graph: λ1 = · · · = λk = 0.
Low-Rank Property:
rank(L) = p − k, indicating a nonconvex constraint in optimization.
Practical Handling via Ky Fan’s Theorem:
Pk
Enforce i=1 λi (L) = 0 to capture the low-rank property.
Optimization involves matrix F as a variable:
k
X
λi (L) = min Tr(FT LF ),
F ∈Rp×k :FT F =I
i=1
This approach simplifies handling the nonconvex constraint of low-rank Laplacian
matrices.
Portfolio Optimization Financial Data: Graphs 29 / 58
k-component graphs
Example of a 3-component graph (3 clusters) with corresponding Laplacian matrix
eigenvalues (3 zero eigenvalues):

Graph Laplacian matrix eigenvalues

Portfolio Optimization Financial Data: Graphs 30 / 58

Low-rank graph learning formulations
Low-Rank Constraint Handling:
Low-rank property rank(L) = p − k is nonconvex.
Pk
Enforce i=1 λi (L) = 0 using Ky Fan’s theorem.
Alternate minimization between L and F for optimization.
Regularized Optimization Formulation:
Introduce low-rank constraint as regularization in objective:

minimize f (L) + γTr(FT LF ),

L,F

F optimized to be eigenvectors corresponding to k smallest eigenvalues of L.

Low-Rank Graph Approximation Example:
Given graph Laplacian L0 , approximate with low-rank L:
 
minimize ∥L − L0 ∥2F + γTr FT LF ,
L⪰0,F

Subject to Laplacian constraints and FT F = I.

Portfolio Optimization Financial Data: Graphs 31 / 58
Low-rank graph learning formulations

Low-Rank Sparse GMRF Graphs Example:

Incorporate low-rank regularization into sparse GMRF graph learning:
 
maximize log gdet(L) − Tr(LS) − ρ∥L∥0,off − γTr FT LF ,
L⪰0,F

Subject to Laplacian constraints and FT F = I.

Avoiding Trivial Solutions:
To prevent isolated nodes, control node degrees with constraint diag(L) = 1.
R Package for Spectral Constraints:
spectralGraphTopology provides functions for graph learning with spectral constraints.

Portfolio Optimization Financial Data: Graphs 32 / 58

Bipartite graphs
Adjacency Matrix Symmetric Eigenvalues:
Bipartite graphs have adjacency matrix eigenvalues symmetric around zero.
Eigenvalue decomposition: W = V Diag(ψ1 , ψ2 , . . . , ψp )V T .
Symmetry condition: ψi = −ψp−i , ∀i.
Laplacian Matrix Structure:
Alternative characterization via Laplacian matrix:
 
Diag(B1) −B
L= ,
−BT Diag(BT 1)
B ∈ Rr+×q represents edge weights between two node types.
Constructed L inherently satisfies L1 = 0, Lij = Lji ≤ 0, ∀i ̸= j.
Optimization Challenges:
Enforcing symmetric eigenvalues is nonconvex and complex.
Laplacian matrix structure provides a more practical approach to enforce bipartite
properties.
R Package for Bipartite Graphs:
finbipartite offers methods for solving bipartite graph problems.
Portfolio Optimization Financial Data: Graphs 33 / 58
Bipartite graphs
Example of a bipartite graph with corresponding adjacency matrix eigenvalues:

Graph Adjacency matrix eigenvalues

Portfolio Optimization Financial Data: Graphs 34 / 58

Bipartite graph learning formulations

Bipartite Graph Approximation:

Goal: Find closest bipartite graph approximation to a given graph L0 .
Formulation:
minimize ∥L − L0 ∥2F
L,B  
Diag(B1) −B
subject to L =
−BT Diag(BT 1)
B ≥ 0, B1 = 1.
Objective: Minimize Frobenius norm difference between L and L0 .
Constraints: Enforce bipartite structure and non-negativity of B.

Portfolio Optimization Financial Data: Graphs 35 / 58

Bipartite graph learning formulations

Bipartite Graphs from Sparse GMRFs:

Goal: Learn a sparse bipartite graph under a GMRF framework.
Formulation:
maximize log gdet(L) − Tr(LS) − ρ∥L∥0,off
L⪰0,B  
Diag(B1) −B
subject to L =
−BT Diag(BT 1)
B ≥ 0, B1 = 1.
Objective: Maximize log-determinant of L minus trace term and sparsity penalty.
Constraints: Enforce bipartite structure, non-negativity, and uniform distribution of
connections in B.

Portfolio Optimization Financial Data: Graphs 36 / 58

Numerical experiments

Data Overview:
S&P 500 stock price data from 2016-2019 for Industrials, Consumer Staples, and Energy
sectors of the S&P 500 index.
Stock Classification into Sectors and Industries:
Stocks grouped into sectors and industries for investment diversification.
Classification criteria vary, including production-oriented and market-oriented approaches.
Major Sector Classification Systems:
GICS: Developed by MSCI and S&P for classifying companies into industry groups,
sectors, and sub-industries.
ICB: Created by Dow Jones and FTSE for categorizing companies into industries,
supersectors, sectors, and subsectors.
TRBC: Thomson Reuters’ system for classifying companies into economic sectors,
business sectors, and industries.

Portfolio Optimization Financial Data: Graphs 37 / 58

Numerical experiments
Major sector classification systems:

Level/System GICS ICB TRBC

1st 11 Sectors 10 Industries 10 Economic Sectors
2nd 24 Industry Groups 19 Supersectors 28 Business Sectors
3rd 68 Industries 41 Sectors 56 Industry Groups
4th 157 Sub-Industries 114 Subsectors 136 Industries

Relevance of Classification Systems:

Each system groups stocks differently, with some similarities.
Not always clear which classification is most relevant for portfolio investment.
Data-Oriented Approach to Stock Classification:
With data availability, traditional classification systems can be bypassed.
Learn stock graph from data to discover natural clusters and relationships.
Enforce a k-component graph for automatic clustered graph generation.
Portfolio Optimization Financial Data: Graphs 38 / 58
Numerical experiments: Two-stage versus joint design of k-component
financial graphs

Two-Stage Approach:
Step 1: Learn a connected graph using GMRF design (e.g., GLASSO or sparse GMRF).
Step 2: Perform a low-rank approximation to the learned graph.
Joint Approach:
Enforce the low-rank property directly in the GMRF formulation.
Expected to outperform the two-stage approach by integrating all constraints and
objectives from the start.
Comparison and Effectiveness:
Joint design significantly superior to two-stage design.
Demonstrated by the clear difference in outcomes when employing each approach.
Considerations for Low-Rank Graph Learning:
Importance of controlling node degrees to avoid trivial solutions or isolated nodes.
Ensures the practical utility and connectivity of the learned graph.

Portfolio Optimization Financial Data: Graphs 39 / 58

Numerical experiments: Two-stage versus joint design of k-component
financial graphs

Two−stage procedure Joint procedure

Portfolio Optimization Financial Data: Graphs 40 / 58

Numerical experiments: Isolated nodes in low-rank designs for
k-component graphs

Issue with Isolated Nodes:

Imposing a low-rank structure to learn a k-component graph can result in isolated nodes.
These nodes artificially increase the number of clusters by being disconnected.
Solution:
Control the degrees of nodes to ensure they are nonzero and preferably balanced.
Prevents the occurrence of isolated nodes, maintaining connectivity within the graph.
Impact of Degree Control:
Significant improvement in graph structure by avoiding isolated nodes.
Ensures more meaningful and connected clusters in the learned graph.

Portfolio Optimization Financial Data: Graphs 41 / 58

Numerical experiments: Isolated nodes in low-rank (clustered) designs for
k-component financial graphs

Without degree control With degree control

Portfolio Optimization Financial Data: Graphs 42 / 58

Outline

1 Graphs

2 Learning graphs

3 Learning structured graphs

4 Learning heavy-tailed graphs

5 Learning time-varying graphs

6 Summary
Transition from Gaussian to Heavy-Tailed Graphs
Gaussian Maximum Likelihood Estimation (MLE) based on:
1 1
 
f (x) = q exp − (x − µ)T Σ−1 (x − µ) ,
(2π)N |Σ| 2

leading to optimization problem for Laplacian matrix L:

maximize log gdet(L) − Tr(LS)
L⪰0
subject to L1 = 0, Lij = Lji ≤ 0, ∀i ̸= j.
Heavy-Tailed Distribution Model: Student t distribution:
− p+ν
1 1

2
f (x) ∝ p 1 + (x − µ)T Σ−1 (x − µ) ,
|Σ| ν
leading to a new optimization problem:
T
p+ν X 1
 
maximize log gdet(L) − log 1 + (x (t) )T Lx (t)
L⪰0 T t=1 ν
subject to L1 = 0, Financial
Portfolio Optimization L =Data: ≤ 0, ∀i ̸= j.
L Graphs 44 / 58
Learning heavy-tailed graphs
Majorization-Minimization (MM) Framework:
Nonconvex heavy-tailed problem solved iteratively using MM.
Logarithm upper bound used for simplification:
ν + (x (t) )T Lx (t)
   
1 1
log 1 + (x (t) )T Lx (t) ≤ log 1 + (x (t) )T L0 x (t) + − 1.
ν ν ν + (x (t) )T L0 x (t)
Iterative Gaussianized Problem Solving:
MM algorithm solves sequence of Gaussianized problems:
maximize log gdet(L) − Tr(LS k )
L⪰0
subject to L1 = 0, Lij = Lji ≤ 0, ∀i ̸= j,
where S k is a weighted sample covariance matrix.
R Package for Heavy-Tailed Graph Learning:
fingraph R package provides algorithms for learning from heavy-tailed data.
Functions include learn_regular_heavytail_graph() and
learn_kcomp_heavytail_graph() for solving heavy-tailed graph learning problems
with additional constraints.
Portfolio Optimization Financial Data: Graphs 45 / 58
Numerical experiments: From Gaussian to heavy-tailed graphs

Stock Price Data Analysis:

Data from three S&P 500 sectors: Industrials, Consumer Staples, Energy.
Time period: 2016-2019.
Comparison of Graphical Models:
Gaussian MRF with concave sparsity regularizer used for Gaussian case.
Heavy-tailed MRF formulation used for non-Gaussian case.
Results and Conclusions:
Next figure shows MRF results under Gaussian and heavy-tailed assumptions.
Heavy-tailed graphs are more suitable for financial data analysis.

Portfolio Optimization Financial Data: Graphs 46 / 58

Numerical experiments: Gaussian versus heavy-tailed graph learning with
stocks

Sparse Gaussian MRF graph Heavy−tailed MRF graph

Portfolio Optimization Financial Data: Graphs 47 / 58

Numerical experiments: k-component graphs
FX Market Data Analysis:
Data: 34 most traded currencies from Jan. 2nd, 2019 to Dec. 31st, 2020.
Data matrix: Log-returns of currency prices relative to USD.
Graph Learning Comparisons:
GMRF with ℓ1 -norm regularizer vs. concave sparsity regularizer.
Heavy-tailed MRF formulation for non-Gaussian data.
GMRF with ℓ1 -norm does not produce sparse graphs.
Concave sparsity regularizer and heavy-tailed MRF yield cleaner graphs.
Heavy-tailed MRF highlights expected correlations between geographically close
currencies (e.g., {Hong Kong, China}, {Taiwan, South Korea}, {Poland, Czech
Republic}).
k-Component Graphs for Clustering:
9-component graphs show clear clustering.
Heavy-tailed MRF graph provides clearer, more reasonable clusters (e.g., {New Zealand,
Australia}, {Poland, Czech Republic, Hungary}).
Heavy-tailed MRF with low-rank structure avoids isolated nodes, unlike GMRF
formulations.
Portfolio Optimization Financial Data: Graphs 48 / 58
Numerical experiments: Gaussian versus heavy-tailed graph learning with
FX

GMRF graph (l1−norm) GMRF graph (reweighted l1−norm) Heavy−tailed MRF graph

Portfolio Optimization Financial Data: Graphs 49 / 58

Numerical experiments: Gaussian versus heavy-tailed multi-component
graph learning with FX

GMRF graph (l1−norm) GMRF graph (reweighted l1−norm) Heavy−tailed MRF graph

Portfolio Optimization Financial Data: Graphs 50 / 58

Outline

1 Graphs

2 Learning graphs

3 Learning structured graphs

4 Learning heavy-tailed graphs

5 Learning time-varying graphs

6 Summary
Learning time-varying graphs

Dynamic vs. Static Graphs:

Static graphs do not capture time variations in dynamic systems (e.g., financial markets).
Dynamic graphs are essential for understanding time-varying behaviors.
Challenges in Learning Dynamic Graphs:
Formulation complexity and increased variable count.
Limited literature due to these challenges.
Naive Approach for Dynamic Graphs:
Divide observations into T chunks, learn graphs Lt independently.
Drawbacks: fewer observations per chunk, potential lack of time-consistency.

Portfolio Optimization Financial Data: Graphs 52 / 58

Learning time-varying graphs

Time-Consistent Dynamic Graph Learning:

Regularization term d(Lt−1 , Lt ) ensures smooth transitions.
Frobenius norm: d(Lt−1 , Lt ) = ∥Lt−1 − Lt ∥2F .
ℓ1 -norm: d(Lt−1 , Lt ) = ∥Lt−1 − Lt ∥1 .
Dynamic Graph Learning Formulation:
Objective: Minimize trace and log-det terms, plus regularization for time-consistency.
Constraints ensure positive semi-definiteness and structural conditions.
Hyper-parameter δ controls time-consistency level.
Solution and Estimation:
Dynamic graph solution L̂t|T considers all T chunks (with look-ahead bias).
Rolling window approach for causal estimate L̂t|t avoids look-ahead bias.

Portfolio Optimization Financial Data: Graphs 53 / 58

Outline

1 Graphs

2 Learning graphs

3 Learning structured graphs

4 Learning heavy-tailed graphs

5 Learning time-varying graphs

6 Summary
Summary

The first application of graphical models in financial markets goes back to 1999, using a
simple correlation graph. Since then, numerous methods have been proposed.
Among the methods herein overviewed, only a few are suitable for financial time series and
produce desirable graphs:

Sparse GMRF graphs are a good starting point.

k-component graphs with low-rank structure and degree control can be used for
data-driven asset clustering.
Due to the heavy-tailed nature of financial data, heavy-tailed graph models are more
suitable than Gaussian ones and can be solved iteratively.

Portfolio Optimization Financial Data: Graphs 55 / 58

References I
Cardoso, J. V. M., and D. P. Palomar. 2020. “Learning Undirected Graphs in Financial Markets.” In
Proceedings of the 54th Asilomar Conference on Signals, Systems and Computers. Pacific Grove, CA,
USA.
Cardoso, J. V. M., J. Ying, and D. P. Palomar. 2021. “Graphical Models for Heavy-Tailed Markets.” In
Proceedings of the Advances in Neural Information Processing Systems (NeurIPS). Virtual.
———. 2022a. “Nonconvex Graph Learning: Sparsity, Heavy-Tails, and Clustering.” In Signal Processing
and Machine Learning Theory, Digital Signal Processing Series. Elsevier.
———. 2022b. “Learning Bipartite Graphs: Heavy Tails and Multiple Components.” In Proceedings of the
Advances in Neural Information Processing Systems (NeurIPS). New Orleans, LA, USA.
Dong, X., D. Thanou, M. Rabbat, and P. Frossard. 2019. “Learning Graphs from Data: A Signal
Representation Perspective.” IEEE Signal Processing Magazine 36 (3): 44–63.
Egilmez, H. E., E. Pavez, and A. Ortega. 2017. “Graph Learning from Data Under Laplacian and Structural
Constraints.” IEEE Journal of Selected Topics in Signal Processing 11 (6): 825–41.
Kolaczyk, E. D. 2009. Statistical Analysis of Network Data: Methods and Models. New York:
Springer-Verlag.
Portfolio Optimization Financial Data: Graphs 56 / 58
References II
Kumar, S., J. Ying, J. V. M. Cardoso, and D. P. Palomar. 2019. “Structured Graph Learning via Laplacian
Spectral Constraints.” In Proceedings of the Advances in Neural Information Processing Systems
(NeurIPS). Vancouver, Canada.
———. 2020. “A Unified Framework for Structured Graph Learning via Spectral Constraints.” Journal of
Machine Learning Research (JMLR), 1–60.
Lake, B., and J. Tenenbaum. 2010. “Discovering Structure by Learning Sparse Graphs.” In Proceedings of
the 33rd Annual Cognitive Science Conference.
Lauritzen, S. 1996. Graphical Models. Oxford: Oxford University Press.
Mantegna, R. N. 1999. “Hierarchical Structure in Financial Markets.” The European Physical Journal
B-Condensed Matter and Complex Systems 11: 193–97.
Marti, G., F. Nielsen, M. Binkowski, and P. Donnat. 2021. “A Review of Two Decades of Correlations,
Hierarchies, Networks and Clustering in Financial Markets.” In Progress in Information Geometry, edited
by F. Nielsen, 245–74. Springer.
Mateos, G., S. Segarra, A. G. Marques, and A. Ribeiro. 2019. “Connecting the Dots.” IEEE Signal
Processing Magazine 36 (3): 16–43.

Portfolio Optimization Financial Data: Graphs 57 / 58

References III

Palomar, D. P. 2025. Portfolio Optimization: Theory and Application. Cambridge University Press.

Ying, J., J. V. M. Cardoso, and D. P. Palomar. 2020. “Nonconvex Sparse Graph Learning Under Laplacian
Constrained Graphical Model.” In Proceedings of the Advances in Neural Information Processing
Systems (NeurIPS). Virtual.

Zhao, L., Y. Wang, S. Kumar, and D. P. Palomar. 2019. “Optimization Algorithms for Graph Laplacian
Estimation via ADMM and MM.” IEEE Transactions on Signal Processing 67 (16): 4231–44.

Portfolio Optimization Financial Data: Graphs 58 / 58