The document describes a research team that studies emerging methodologies in economics and finance, including complex networks and machine learning. The team is based at Democritus University of Thrace in Greece and led by Drs. Periklis Gogas and Theophilos Papadimitriou. Their research focuses on using techniques like minimum dominating sets and support vector machines to analyze economic and financial networks and time series data. They have found these emerging methodologies can more accurately represent network structures and forecast macroeconomic and financial variables compared to traditional econometric methods.
1 of 16
Download to read offline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
More Related Content
EMEF
1. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 1 | P a g e
Department of Economics
Democritus University of Thrace
Komotini, Greece
Emerging
Methodologies
in Economics
and Finance
2. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 2 | P a g e
Emerging Methodologies
in Economics and Finance
1. SUMMARY
THE TEAM. Our research team is operates within the Department of Economics, Democritus
University of Thrace, Greece. Our research efforts are funded by a Research Grant from the
European Union (Research Funding Program THALES) under the title “Study and Forecasting of
Economic Data Using Machine Learning”, MIS 380292.
MEMBERS. The research team is led by associate professors Periklis Gogas an economist (B.A.,
M.A., Ph.D.) and Theophilos Papadimitriou a mathematician (B.A.) and electrical engineer (M.Sc.,
Ph.D). Seven Ph.D. candidates and nine Master’s students are actively working for the team.
RESEARCH INTERESTS. Our team’s interests include both classic and emerging methodologies of
Econometrics as they are applied to Economics and Finance. We currently work with: a) Machine
Learning: Support Vector Machines for Classification and Regression and Deep Learning
Architectures and b) Complex Networks: Threshold – Minimum Dominating Set, Weighted
Dominating Set, and Multivariate Networks.
COMPLEX NETWORKS. Since 2012 we explore the field of Economic and Financial Networks using
Graph Theory. Two PhD students and four graduate students work exclusively on this project. Our
main goal in these research studies is to obtain a reduced version of the initial network which: a)
contains an adequate amount of the total information for each goal (the amount of needed
information depends on the application) and b) is more efficient easier to control and analyze.
We refer to this as the representation goal. This task has been undertaken in the literature so far
mainly through the MST approach in a variety of economic systems including stock markets
(Bonanno et al, 2004; Tse et al, 2010), banking networks (Papadimitriou et al, 2013),
macroeconomic networks (Hill, 1999; Gilmore et al, 2010; Dias, 2013) and others. In our work, we
show that the Minimum Spanning Tree has inherent limitations when applied to economics and
financial networks that may lead to inefficient and/or misleading results. These limitations stem
from MST’s algorithmic calculation process. To overcome these limitations we proposed a more
efficient methodology based on a modified Minimum Dominating Set (MDS) algorithm, a
methodology mainly used in the routing of computer and communications networks. The original
MDS has its own disadvantages as in the optimization procedure the information included in the
edges is not exploited. The new method we propose overcomes this problem and is called the
Threshold-Minimum Dominating Set (T-MDS). The proposed T-MDS can: a) identify the smallest
subset of nodes which b) represents the whole network more efficiently than the two other
alternatives.
MACHINE LEARNING. Our initial attempt was to forecast financial time series using Machine
Learning techniques and specifically Support Vector Machines for Classification and Regression.
Support Vector Machines (SVM) is a binary supervised machine learning classifier. Proposed by
3. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 3 | P a g e
Cortes and Vapnik (1995) the basic notion behind the method is to find a linear separator (a
hyperplane) of the two classes in the so called feature space. With the use of a kernel function
the original data space is projected to a higher dimensional space called the feature space. Solving
a convex minimization problem the algorithm converges to a linear separator that has the largest
margin between the two classes. The margin between the classes is defined by a set of data points
called Support Vectors. We find that the Supports Vector Machines methodology when used in
macroeconomics and financial data provides superior out-of-sample forecasting accuracy as
compared to classic econometrics techniques. Recently we started employing in these data Deep
Learning Machines that is considered the current state-of-the-art.
2. Identity
Our research team operates within the Department of Economics, Democritus University of
Thrace, Komotini, Greece. The members of the team are:
Researchers
Dr. Periklis Gogas, Associate Professor
He was born in Thessaloniki, Greece in 1969. He received his Ph.D. degree from the Department
of Economics of the University of Calgary and his Master’s degree from the University of
Saskatchewan. His B.A. in Economics is from the University of Macedonia, Greece. Recently, a
Visiting Scholar in Finance at the Ross School of Business of the University of Michigan. He also
taught in the past at Plovdiv University, and the vocational center of the Athens Stock Exchange.
His research interests include macroeconomics, financial economics, chaotic and non-linear
dynamics graph theory and machine learning applied to macro and finance. He served in the past
at the position of the Financial Director of a large Greek multinational corporation. He authored
more than 40 articles in journals such as Journal of Money Credit and Banking, Journal of Banking
and Finance, Economic Modeling, Journal of Forecasting, International Finance, Computational
Economics, Open Economies Review, etc.
Dr. Theophilos Papadimitriou, Associate Professor
He was born in Thessaloniki, Greece, in 1972. He received the Diploma degree (B.Sc.) from the
Department of Mathematics, Aristotle University of Thessaloniki, Greece, and the D.E.A.
A.R.A.V.I.S (Automatique, Robotique, Algorithmique, Vision, Image, Signale) degree (M.Sc.) from
the Department of Computer Science, University of Nice-Sophia Antipolis, France, in 1996 and the
Ph.D. degree from the School of Engineering, Aristotle University of Thessaloniki, Greece in 2000.
In 2001, he joined the Department of Economics of the Democritus University of Thrace in
Komotini, Greece, where he served as a lecturer (2002-2008), assistant professor (2008-2013).
Currently he holds the position of Associate Professor in the same department. Dr. Papadimitriou
co-authored more than 80 journal papers, conference papers and book chapters combined. He
4. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 4 | P a g e
served as a reviewer for various publications and as a member to scientific committees for
Conferences and Workshops. Theophilos Papadimitriou current research interests include
complex network, machine learning, and data analysis.
Ph.D. Students
Vasileios Plakandaras
Eftymia Chrysanthidou
Efthimios Stathakis
Georgios-Antonios Sarantitis
Maria-Artemis Matthaiou
Anna Agrapetidou
Efi Parharidou
Master’s Students
Anna Amvrosiadou
Pavlina Georgiadou
Fotini Gkika
Georgios Zikidis
Dimitrios Karagkiozis
Athina Kyprianidou
Sophia Maltepioti
Gerasimos Panagiotidis
Konstantina Papathanasiou
Funding
The research team is funded by a research grant received by Prof. Papadimitriou from the
European Union and financed through the European Social Fund (ESF) and Greek national funds
through the Operational Program “Education and Lifelong Learning” of the National Strategic
Reference Framework (NSRF) Research Funding Program THALES under the title “Study and
Forecasting of Economic Data Using Machine Learning” (MIS 380292).
Research Interests
Our team is interested in both classic and emerging methodologies of Econometrics as they are
applied to Economics and Finance. Currently, we work with: a) Machine Learning (Support Vector
Machines for Classification and Regression, Deep Learning Architectures) and b) Complex
Networks (Threshold – Minimum Dominating Set, Weighted Dominating Set, and Multivariate
Networks).
Selected Publications (click on a paper to open it)
1. P. Gogas, T. Papadimitriou, E. Chrysanthidou, “Yield Curve Point Triplets in Recession
Forecasting”, accepted for publication in International Finance.
5. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 5 | P a g e
2. V. Plakandaras, T. Papadimitriou, P. Gogas, “Forecasting daily and monthly exchange rates
with machine learning techniques”, accepted for publication in Journal of Forecasting.
3. V. Plakandaras, P. Gogas, R. Gupta, T. Papadimitriou, “US inflation dynamics on long range
data”, accepted for publication in Applied Economics.
4. I. Pragidis, P. Gogas, V. Plakandaras, T. Papadimitriou, “Fiscal shocks and asymmetric effects:
A comparative analysis”, accepted for publication in Journal of Economic Asymmetries.
5. V. Plakandaras, R. Gupta, P. Gogas, T. Papadimitriou, “Forecasting the US real house index”,
accepted for publication on Economic Modelling.
6. T. Papadimitriou, P. Gogas, G. Sarantitis, “Convergence of European Business Cycles: A
Complex Networks Approach”, Computational Economics, November 2014.
7. V. Plakandaras, T. Papadimitriou, P. Gogas, K.I. Diamantaras “Market Sentiment and Exchange
Rate Directional Forecasting”, accepted for publication on Algorithmic Finance.
8. P. Gogas, T. Papadimitriou, A. Agrapetidou, “Forecasting Bank Credit Ratings”, Journal of Risk
Finance, Vol. 15, No. 2, 2014, pp. 195-209.
9. T. Papadimitriou, P. Gogas, E. Stathakis, “Forecasting Energy Markets using Support Vector
Machines”, Energy Economics, Vol. 44, pp. 135-142, July 2014.
10. P. Gogas, T. Papadimitriou, M. Matthaiou, E. Chrysanthidou, “Yield Curve and Recession
Forecasting in a Machine Learning Framework”, Computational Economics, Vol. 45, no. 4, pp
635-645, April 2015.
11. P. Gogas, V. Plakandaras, T. Papadimitriou, “Public Debt and Private Consumption in OECD
countries”, Journal of Economic Asymmetries, Vol 11, no. 1, June, 2014, pp. 1-7.
12. P. Gogas, T. Papadimitriou, G. Sarantitis, “Testing Purchasing Power Parity theory in a DFA
Framework and Rolling Hurst Exponent: the case of 23 OECD Countries”, Applied Financial
Economics, Vol. 23, no 17, pp. 1399-1406, 2013.
13. T. Papadimitriou, P. Gogas, B.M. Tabak, “Complex Networks and Banking Systems
Supervision”, Physica A, Vol. 392, pp. 4429-4434, Oct. 2013.
14. P. Gogas, T. Papadimitriou and E. Takli, (2013) ''Comparison of simple sum and Divisia
monetary aggregates in GDP forecasting: a support vector machines approach'', Economics
Bulletin, Vol. 33, No. 2, pp. 1101-1115, 2013.
3. Methodologies Employed
3.1 Complex Networks
In 2012 we started exploring the field of Economic and Financial Networks through the prism of
Graph Theory. We are still highly committed in the field with two PhD students and four graduate
students working exclusively on this project. Our main goal in these research studies is to obtain
a reduced version of the initial network which: a) contains an adequate amount of the total
information for each goal (the amount of needed information depends on the application) and b)
is more efficient easier to control and analyze. We refer to this as the representation goal. This
task has been undertaken in the literature mainly through the MST approach in a variety of
economic systems including stock markets (Bonanno et al, 2004; Tse et al, 2010), banking
6. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 6 | P a g e
networks (Papadimitriou et al, 2013), macroeconomic networks (Hill, 1999; Gilmore et al, 2010;
Dias, 2013) and others. Nonetheless, the Minimum Spanning Tree has inherent limitations that
stem from its algorithmic calculation process. We investigated these limitations and proposed a
more efficient methodology based on a modified Minimum Dominating Set (MDS) algorithm, a
methodology mainly used in the routing of computer and communications networks. The new
method is called Threshold-Minimum Dominating Set (T-MDS). By comparing these
methodologies (MST, MDS and T-MDS) we can allege that the T-MDS can: a) identify the smallest
subset of nodes which b) represents the whole network more efficiently than the two other
alternatives.
Consider the network depicted in Figure 1. In this network, the values of the edges correspond to
distances and thus nodes with more similar behavior are connected by a low distance edge and
vice-versa. The similarity can be calculated using correlation based metrics, concordance based
metrics or even a z-score multivariate distance.
Figure 1. The nodes, edges and distances of the initial network
Figure 2. The core nodes of the MST based method.
The MST-based approach identifies a set of four core nodes, namely nodes 2, 3, 5 and 6, to
describe the behavior of the entire network. They can be used as gauges for the behavior of their
neighborhood. Whenever a shift in the behavior of a core node occurs, it reflects a change in the
7. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 7 | P a g e
behavior of its neighborhood. Consequently, the state of each neighborhood (and eventually of
the entire network) can be represented by the core nodes, which is the basic concept of the
methodology in Papadimitriou et al. (2013).
However, by closely examining the topological features of the core nodes, we observe that:
The MST procedure eliminates the edge e25 to avoid the creation of a loop in the path,
although the corresponding distance d25 is one of the shortest in the network. This means
that even though a) the behavior of nodes 2 and 5 is very similar and b) 2 is a core node,
the behavior of node 5 is not represented by core node 2. In fact, this limitation of the
MST algorithm leads to the identification of node 5 as a core node as well.
On the other hand, the edge e34 is included in the MST, although the corresponding
distance d34 is the maximum in the network. This means that core node 3 describes the
behavior of node 4, even though the two nodes are very dissimilar.
Hence, the MST-based approach “misses” some essential relations (e.g. the e25 edge), and at the
same time maintains some edges that connect nodes with low similarity in the process of
identifying the core nodes (e.g. the e34 edge). These are important drawbacks that could affect
the efficiency of the MST-based solution to the representation goal.
Next, we apply the simple MDS methodology in the theoretical seven nodes’ network of Figure 1.
Figure 3. Minimum Dominating Set (gray nodes)
Τhe absence of the no-loop restriction, in the MDS algorithm, increases the number of the
considered edges in the network, which in turn is manifested through larger neighborhoods (since
more nodes remain interconnected). As the network becomes denser, fewer MDS nodes are
needed to reach the representation goal. Consequently, it is more probable that the cardinality
of the MDS nodes’ set will be smaller than the corresponding MST one. Indeed, by applying the
MDS technique, we are able to represent the behavior of the whole network using just three
nodes (one less than the MST methodology). This fact is noteworthy but not critical to our cause.
We do search for a small subset of nodes, but most importantly we search for a set that best
represents the entire network. The MDS nodes’ set, as opposed to the MST-based one, is a better
8. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 8 | P a g e
representation of the network: all the small distance edges are included in the representative
nodes’ neighborhoods. This implies that all node pairs with similar behavior are included in the
neighborhoods. On the other hand though, the same is true for all the large distance edges (see
in Figure 3 the edges e34, e35, e46 and e67). These edges connect nodes with dissimilar behavior and
their inclusion in the MDS neighborhoods may jeopardize the representation goal. Node 4 in the
example belongs to the neighborhoods of nodes 3 and 6. In both cases, the edges linking node 4
with the two representative nodes describe long distances, meaning that the node in question
does not have similar behavior with the nodes that represent it. Consequently, every decision that
we make for node 4 based on the behavior of nodes 3 and 6 is potentially unreliable.
Threshold-Minimum Dominating Set (T-MDS)
We suggest that the limitations of the MST-based approach and the classical MDS technique can
be effectively addressed with the inclusion of one extra step prior to the identification of the MDS:
the imposition of a threshold on the edges’ distances in the initial network. By doing so, we ensure
that only the edges with low distance (i.e. the edges which connect nodes with high similarity)
survive, while all the other edges are eliminated.
Definition: We call a Threshold – Minimum Dominating Set the two step methodology for the
identification of the most representative nodes of a network, defined as:
Step 1. A thresholding on the edges’ distances leading to the elimination of all edges that
correspond to large distances.
Step 2. The identification of the MDS nodes on the remaining network.
The suggested process is depicted in Figure 4. In the left part of the Figure, in Step 1 we impose a
threshold of 𝑝 = 4 to remove any edges (depicted with the dashed lines) with distance greater
than 4. Thus, we eliminate the edges that connect nodes with dissimilar behavior; this results in
the isolation of nodes 3 and 6 from the rest of the network. Then, the MDS is identified on the
right part of Figure 4..
Figure 4. Threshold-Minimum Dominating Set (colored nodes)
9. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 9 | P a g e
Apparently, the Minimum Dominating Set consists of only three nodes: one dominant node (node
2) and two isolated ones (nodes 3 and 6). The dominant node of the network describes the
behavior of the interconnected nodes 1, 4, 5, and 7. The edges connecting each of these nodes
with the dominant one represent a short distance and thus this node can successfully act as a
representative agent for the neighborhood.
The Threshold-Minimum Dominating Set (T-MDS) retains the advantages of the Minimum
Spanning Tree and the classical Minimum Dominating Set while overcoming their inherent
limitations: it is a compact and reliable subset able to describe efficiently the topology of the
entire network. Among the three presented methodologies, the T-MDS based approach yields the
best solution to the representation goal. The described methodology was successfully applied
using a simple computer to real world networks that include more than 4000 nodes.
Application Range
Banking networks
"Complex Networks and Banking System Supervision", Physica A, 392, pp. 4429-4434.
"Bank Supervision using the Threshold-Minimum Dominating Set", Physica A, under
review.
We used the T-MDS methodology to create an auxiliary supervision/monitoring mechanism that
can be used by the banking system supervising authority, usually the central bank. This system is
both efficient with respect to the required resources needed and can also promptly identify a set
of banks that are in distress so that immediate and appropriate action can be taken by the
supervising authority. The T-MDS is used to identify the smallest and most efficient subset of
banks that can be used as a) sensors of distress of a manifesting banking crisis and b) provide a
path of possible contagion. We propose the use of this method as a supplementary monitoring
tool in the arsenal of a Central Bank. One dataset we used includes the 122 largest American banks
in terms of their interbank loans. At threshold level of t = 0.8 our method identifies 47 T-MDS
nodes, with 21 of them being isolated ones. According to this, the interconnected nodes in the
thresholded network are 101, and they can be monitored by the 26 dominant nodes.
U.S. Gross State Product Synchronization
In this study we created the U.S. states’ network, based on their Gross State Product (GSP). We
use it to explore the inter-relations of the GSP growth rates and the degree of their business cycle
synchronization. Moreover, this approach allowed us to identify a subset of states that can
efficiently describe the collective behavior of the entire network. The policy implications are
obvious: the monitoring of these key states from the federal fiscal and monetary authorities may
provide an alternative and efficient tool in the designing and implementation of economic policy.
To do so, we used four alternative measures of similarity: correlation, weighted correlation, the
sign concordance index (SCI), and the weighted sign concordance index. Both types of measures
(correlations and SCI) have indicated an increased similarity between the U.S. states GSP in the
more recent years, evidenced both by the T-MDS methodological analysis and the simple network
metrics. In both cases the weighted versions (where the more recent GSP values received a higher
10. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 10 | P a g e
weight) corresponded to denser networks in which the overall correlations are higher. Thus, this
can be viewed as evidence of an increased business cycles synchronization
Figure 5. Network topology in the case of a weighted Sign Concordance Index and threshold
level t=0.75 (nodes sized according to their node degree)
Another interesting result is identification and examination of crisis dispersion paths between the
U.S. states. The imposition of a threshold on the network’s edges removes low strength edges
and allows only for the existence of the highly important and informative ones. This procedure
creates a map of GSP inter-relations that could be used as a guide for contagion in the case of a
crisis manifestation in one or more U.S. states. By studying this map, policy authorities could
intervene timely and effectively to contain the crisis and/or curtail its effects and further
contagion.
The U.K. Consumer Price Index
During the summer of 2014, our PhD student Georgios Sarantitis was employed as a research
associate at the Bank of England. His job description was exclusively to apply the T-MDS
methodology on the calculation of the United Kingdom’s Consumer Price Index (CPI). He modeled
the United Kingdom’s Consumer Price Index as a complex network and he applied clustering and
optimization techniques to study the network’s evolution through time. By doing this, he provided
11. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 11 | P a g e
a dynamic, multi-level analysis of the mechanism that drives inflation in the U.K. The analysis that
we performed on our team revealed that the CPI-classes’ network exhibits an evolving topology
through time which depends substantially on the prevailing economic conditions in the U.K. We
identified non-overlapping communities of these U.K. CPI classes and we observed that they do
not correspond to the actual categories they officially belong to; a finding suggesting that diverse
forces are driving the inter-relations of the CPI classes which are stronger for classes in different
categories rather than for the classes within them. Finally, we created and presented a reduced
version of the U.K. CPI that fulfils all the core inflation measure criteria and can possibly be used
as an appropriate measure of the underlying inflation in the U.K. Since this new measure uses
only 14 out of the 85 U.K. CPI classes, it can be employed to complement the Bank of England’s
arsenal of core inflation measures without the need for further resource allocation.
Income Inequality
"Income Inequality: A State-by-State Complex Network Analysis", Social Indicators
Research, under review.
We investigated the developing patterns of inequality in the U.S. using complex network analysis
and the Threshold-Minimum Dominating Set (T-MDS). We used two alternative measures of
income inequality, the top 1% share of income and the Gini coefficient. We performed a dynamic
analysis over four consecutive periods running from 1916 to 2012.
Table 1. T-MDS metrics for the top 1% inequality measure
1916-1929 1930-1944 1945-1979 1980-2012
T-MDS cardinality 22 28 15 3
Isolated nodes 14 22 10 0
Dominant nodes 8 6 5 3
Our findings reveal a heterogeneous pattern of income inequality and economic integration of
the U.S. states according to each focal period. Furthermore, the empirical findings differentiate
slightly with respect to the two inequality measures employed.
Business Cycle Synchronization
"International Business Cycle Synchronization since the 1870s: Evidence from a Novel Network
Approach", Physica A, Under review.
We examined the issue of business cycle synchronization from a historical perspective in 27
developed and developing countries from the 1870s to the 2010s. Based on the T-MDS, our results
reveal heterogeneous patterns of international business cycle synchronization during
fundamental globalization periods since the 1870s. In particular, the proposed methodology
reveals that worldwide business cycles decoupled during the Gold Standard, though they were
synchronized during the Great Depression. The Bretton Woods era was associated with a lower
degree of synchronization as compared to that during the Great Depression, while worldwide
business cycle synchronization increased to unprecedented levels during the latest period of
12. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 12 | P a g e
floating exchange rates and the Great Recession. This provides empirical evidence in support of
lower business cycle synchronization within periods of fixed exchange rates and lack of an
independent monetary policy.
3.2 Machine Learning
The spark, in 2010, that led to the creation of our team one year later, was the common interest
of Dr. Gogas (economist) and Dr. Papadimitriou (mathematician/electrical engineer) to fuse
methodologies from other fields to pure economic applications. Our initial attempt was to
forecast financial time series using Machine Learning techniques and specifically Support Vector
Machines for Classification and Regression.
Support Vector Machines (SVM) is a binary supervised machine learning classifier. Proposed by
Cortes and Vapnik (1995) the basic notion behind the method is to find a linear separator (a
hyperplane) of the two classes in the so called feature space. With the use of a kernel function
the original data space is projected to a higher dimensional space called the feature space. Solving
a convex minimization problem the algorithm converges to a linear separator that has the largest
margin between the two classes. The margin between the classes is defined by a set of data points
called Support Vectors. An example of an SVM classification employing the RBF kernel is depicted
in Figure 2.
Figure 6: An example of an SVM classification using the RBF kernel. The two classes are separated
with a linear separator at a higher dimensional space called the feature space. The separator in
the feature space when it is re-projected back into the original dimensions becomes a non-linear
function of the initial data space. The encircled observations are the Support Vectors defining the
decision boundary and the observations with a square rounding are misclassified observations.
13. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 13 | P a g e
Application Range
Forecasting Bank Failures
We explored the problem of forecasting bank failures using the SVM methodology with great
success. We used a dataset that includes 1443 U.S. banks and consists of 962 solvent banks and
481 banks that failed during the period 2003-2013. We start with an initial set of 144 variables
that come from the publicly available financial statements of the banks. Next, we employ the
variable selection methodology called Local Learning and we select only 5 variables out of the
original 144. These 5 variables1
are then used as the explanatory variables to train the optimum
SVM model. This produced an out-of-sample forecasting accuracy of 97.67% in a one-period-
ahead (one year) forecasting window. With a forecasting window of two and three years ahead
the model’s accuracy fell to 83.48% and 64.35% respectively. The out-of-sample forecasting
accuracy for the failed banks was 97.39% and for the solvent banks 97.81%. These results are
summarized in Table 2.
Table 2. The foresting accuracy on each class
True Solvent True Insolvent
Estimated Solvent 223 3
Estimated Insolvent 5 112
As a follow up, we investigated the three missed insolvent cases: a) In the case of The First
National bank of Davis, Davis, Oklahoma, during the 2011 examination, significant unrecognized
losses were discovered that exceeded the bank’s capital and allowance for loan and lease losses,
b) The First State Bank received an enforcement action on 21st of December 2010 and in a very
short time, on the 28th of January 2011 it was closed and c) The Community's Bank was
established in 2001 and in 2013 it failed. The same investigation for the five misclassified solvent
case revealed that the financial institutions received an enforcement action from FDIC or financial
help from the U.S. Treasury.
Stress Testing Tool
Moreover, we propose the use of the previously defined hyperplane separating the solvent from
the insolvent banks as a new form of bank stress testing. Solvent banks that are close to the
separating hyperplane are more vulnerable to class changing (become insolvent). Therefore, a
solvent bank’s distance from the separating hyperplane may be used as a measure of its
robustness. By doing this, multi-level stress tests can be performed: a) system-wide, where
alternative scenarios on various macroeconomic and/or financial indicators are reflected on the
five explanatory variables of the banks’ financial statements that are used for the classification
and b) institute specific based on their financial condition. Moreover, for a bank that that is
1 1. Tier 1 (core) risk-based capital/total assets (t-1), 2. Provision for loan and lease losses/total interest income
(t-1), 3. Loan loss allowance/total assets (t-1),4. Total interest expense/total interest income (t-1), 5. Equity
capital to assets (t-1).
14. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 14 | P a g e
forecasted to fail within the next year, the distance from the hyperplane may be exploited and
render it solvent. This distance can be used in time in order to prescribe a set of measures that
will render the bank solvent.
Our proposition is also equipped with a sensitivity analysis of each bank in distress to the five
variables that define its position in the data space.
Forecasting in Energy Markets
"Forecasting Energy Markets using Support Vector Machines ", Energy Economics, 44, pp. 135-
142, July 2014.
The efficiency of an SVM-based forecasting model for the next-day directional change of
electricity prices that are very volatile was also investigated. We first fitted the best autoregressive
SVM model and then augmented it with various relevant variables. We tested this model on the
daily Phelix index of the German and Austrian control area of the European Energy Exchange (ΕΕΧ)
wholesale electricity market. The out-of-sample forecasting accuracy we achieved is 76.12% over
a 200 day period.
We further explored the efficiency of the SVM methodology in the electricity market by
forecasting "price spikes”. In electricity markets, the comparatively high upward or downward
movements of prices within a short period are called spikes. In effect, when we are trying to
forecast price spikes in electricity, we perform outlier detection. The problem was attacked using
Multiclass SVM and Deep Learning architectures. The forecasting accuracy of the optimal model
is depicted in Table 3.
Table 3. Spikes forecasting
SVM Deep Learning
Normal Cases 76% 100%
Positive Spikes 70% 85%
Negative Spikes 56% 98%
Forecasting Exchange Rates
"Forecasting Daily and Monthly Exchange Rates with Machine Learning Techniques", Journal of
Forecasting, forthcoming.
We tested and compared the forecasting ability of various Machine Learning, Neural Networks,
and Econometrics architectures on monthly and daily spot prices of five selected exchange rates:
EUR/USD, JPY/USD, NOK/AUD, NZD/BRL and PHP/ZAR. In doing so, we combine a novel smoothing
technique (initially applied in signal processing) with a variable selection methodology and the
regressors are used in the regression estimation methodologies. After the decomposition of the
original exchange rate series using the Ensemble Empirical Mode Decomposition (EEMD) method
into a smoothed and a fluctuating component, Multivariate Adaptive Regression Splines (MARS)
are used to select the most appropriate variable set from a very large set of explanatory variables
15. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 15 | P a g e
that we collected. The selected variables are then fed into the forecasting models that produce
one-period-ahead forecasts for the two components: smoothed and fluctuating.
We implemented two versions of this hybrid methodological setup; an autoregressive and a
structural one. The autoregressive model consistently outperforms all alternative models in
forecasting out-of-sample in monthly frequency. Using daily data, the structural EEMD-MARS-SVR
model is superior for three out of the five exchange rates in our sample. A close alternative is the
structural EEMD-MARS-NN (Neural Networks) model that outperforms in forecasting the other
two exchange rates. Nevertheless, only the structural EEMD-MARS-SVR methodology consistently
outperforms the benchmark random walk model in out-of-sample forecasting. The above findings
corroborate with the microstructural aspect of the exchange rate market on the short run and the
importance of macroeconomic dynamics (such as PPP, UIP, etc.) on the long run, validating
exchange rate theory. Overall, our models seem to capture the different data generating
processes between short and long horizons. The weak form of exchange rate market efficiency is
rejected for both sampling frequencies due to the fact that the autoregressive model outperforms
the RW benchmark.
Yield Curve and Recession Forecasting
"Yield Curve Point Triplets and Recession Forecasting", International Finance, forthcoming.
Several studies have highlighted the yield curve’s ability to forecast economic activity. These
studies use the information provided by the slope of the yield curve—i.e., pairs of short- and long-
term interest rates. We constructed three models to forecast the positive and negative deviations
of real U.S. GDP from its long-run trend over the period from 1976Q3 to 2011Q4: one that uses
only pairs of interest rates and two that draw on more than two points from the yield curve. We
employ two alternative forecasting methodologies: the probit model, which is commonly used in
this line of literature, and the support vector machines (SVM) approach from the area of machine
learning. Our empirical results show that the SVM model with the RBF kernel and three interest
rates as input variables achieves a 100% out-of-sample forecasting accuracy for recessions
(unemployment gaps) and the best overall accuracy is 80%. Thus, it appears that correct
identification of upcoming unemployment gaps may be achieved at the cost of reduced accuracy
of forecasting inflationary gaps or in other words at the cost of some extra inflation. The interest
rates that produce these results are the 3-month T-bill rate and the 2- and 3-year government
bond rates. Long-term rates do not appear to complement our models’ forecasting ability. Our
interpretation of this finding is that, since the U.S. is a developed industrialized country, the
United States is considered to have a stable long-term economic outlook that is not affected by
short-term dynamics and fluctuations and instead adheres to its long-run potential output. Thus,
agents’ views of future economic activity are not significantly affected by short-term events or
fluctuations, thus rendering the long-term rates uninformative in terms of short-term recession
forecasting.
4. References
16. Department of Economics, Democritus University of Thrace
Emerging Methodologies in Economics and Finance 16 | P a g e
G. Bonanno, G. Caldarelli, F. Lillo, S. Micciche, N. Vandewalle, and R. N. Mantegna (2004),
“Networks of equities in financial markets”, European Physical Journal B, Vol. 38, pp. 363-371.
J. Dias, (2013). Spanning trees and the eurozone crisis. Physica A: Statistical Mechanics and its
Applications, 392, 5974-5984.
C.G. Gilmore, Lucey, B. M. & Boscia, M. W. (2010). Comovements in government bond markets:
A minimum spanning tree analysis. Physica A: Statistical Mechanics and its Applications,
389(21), 4875-4886.
R.J. Hill, (1999). Comparing price levels across countries using minimum-spanning trees. Review
of Economics and Statistics, 81(1), 135-142.
T. Papadimitriou, P. Gogas, B.M. Tabak (2013), "Complex Networks and Banking Systems
Supervision", Physica A: Statistical Mechanics and its Applications,
Vol. 392, no. 19, pp. 4429 – 4434
C.K. Tse, Liu, J. & Lau, F. (2010). A network perspective of the stock market. Journal of Empirical
Finance, vol. 17(4), pp. 659-667.