Journal of Econometrics 212 (2019) 203–220

Contents lists available at ScienceDirect

Journal of Econometrics
journal homepage: www.elsevier.com/locate/jeconom

Monitoring banking system connectedness with big data✩

article info a b s t r a c t

Article history: In this paper, we propose a procedure that generates measures of connectedness
Available online 16 April 2019 between individual firms and for the system as a whole based on information observed
only at the firm level; i.e., no explicit linkages are observed. We apply our procedure
JEL classification:
to large U.S. bank holding companies. We show how bank outcome variables of interest
C32
G21 can be decomposed, including with mixed-frequency models, for network analysis to
G28 measure connectedness across firms. Network analysis of these decompositions produces
measures that could be of use in financial stability monitoring as well as the analysis of
Keywords: individual firms’ linkages.
Financial stability Published by Elsevier B.V.
Bank supervision
Network centrality
Systemic connectedness

1. Introduction

In recent years, the economic literature in several fields has shown increased interest in measuring and understanding
how various economic agents are connected with each other. The literature on social, trade, industrial, and financial
networks has blossomed. Yet, in some fields, and especially in finance, measuring connectedness between financial firms
remains challenging since the data needed to gauge these linkages are not observed readily or are not available at the
appropriate level of granularity. A key insight into this challenge was provided by Diebold and Yilmaz (2014) in that they
chose to gauge financial firm connectedness using their stock returns as the input data. They analyze the network implied
by their stock return volatility, which neatly avoids data availability challenges while still capturing the aggregate and
idiosyncratic effects of the firms’ interconnectedness.1
The goal of this paper is to develop a procedure that can be used to construct firm-level connectedness measures
that rely on as much of the available data for U.S. bank holding companies (BHCs), have a transparent interpretation,
and exhibit desired statistical properties. While we frame our discussion in terms of measuring the connectedness of
these financial firms for various purposes (with an emphasis on bank supervision and financial stability monitoring),
the proposed approach can be applied readily to other networks of interest, such as industrial sectors. Our procedure

✩ The views in this paper are solely the responsibility of the authors and should not be interpreted as reflecting the views of the Federal Reserve
Bank of San Francisco or the Board of Governors of the Federal Reserve System. We thank Tesia Chuderewicz and Shannon Sledz for their excellent
research assistance. We benefited from comments from Fred Furlong, Eric Ghysels, Linda Goldberg, Eric Heitfield, Anna Kovner, James Vickery, two
anonymous referees as well as seminar participants at the Federal Reserve Banks of Dallas and New York, Banque de France, and colleagues within
the Federal Reserve System.
∗ Corresponding author.
E-mail address: [email protected] (G. Hale).
1 The recent network literature has proposed techniques that transform various forms of data into connectedness measures. For example, Demirer
et al. (2017) also use stock market returns to characterize the high-dimensional network linking the largest global banks over the period from 2003
to 2014. In contrast, Hale et al. (2016) use interbank lending exposures to measure the transmission of financial shocks across borders through
international bank connections.

https://doi.org/10.1016/j.jeconom.2019.04.027
0304-4076/Published by Elsevier B.V.
204 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220

generates distinct measures of overall system connectedness and indicators of how individual entities are both connected
to each other and to the system as a whole. The procedure relies on data that is available at the entity level, and no direct
information on actual linkages between entities is available.
From the perspective of bank supervision, the global financial crisis of 2008–2009 and the subsequent regulatory
response have brought forth the importance of measuring and monitoring financial stability and connectedness within
(and across) national banking systems.2 A key design element in such monitoring efforts is in selecting explanatory
variables useful for the identification of both macroprudential (or systemic) and microprudential (or firm-specific)
concerns. Theoretically, all data sources on financial firms are of interest for these activities, but in reality, they all have
particular strengths and weaknesses. For example, focusing on data frequency, we have extensive quarterly balance sheet
data, which provide the most comprehensive and up-to-date summaries of firm conditions (i.e., asset quality and income).
However, these data are available with time lags and likely too infrequently for use in responding to fast-moving events. At
the other extreme, daily market data is timely and aggregate much current information, but their volatility and tendency
to initially overreact to incoming news limits their usefulness for supervisory monitoring purposes.
Bank supervisors have faced related challenges regarding these data sources for some time (see the long span of
literature from Meyer and Pifer (1970) to Balasubramnian and Palvia (2018)). To date, however, these efforts have been
targeted to microprudential supervisory purposes. Our goal is to develop empirical measures that would be useful for
macroprudential monitoring as well. While many measures of financial stability have been proposed, no one source of
banking data provides a clear measure of interconnectedness for both purposes. In our work, we pay specific attention
to identifying connectedness that arises from common factors affecting all network firms separately from connectedness
that is driven more by the spillover of idiosyncratic shocks.
Our procedure consists of two steps: the decomposition of an observed outcome variable at BHC level into components
of interest and the network analysis of these components. The decomposition allows us to extract common factors
that affect all firms within the network and, more importantly, identify idiosyncratic shocks that include any remaining
connections between the firms. We propose a number of approaches on how this decomposition can be implemented,
focusing on dealing with mixed frequency data. The network analysis can then be applied to any component resulting
from the decomposition, with a particular emphasis on the idiosyncratic terms.
For our proposed measures, we use BHC data to study how the largest U.S. banks were interlinked during the period
from 2005 through 2015. The two outcome measures we analyze are weekly stock returns and quarterly returns on assets
(ROA). We decompose these measures into fitted values and residuals using various regression techniques. Specifically,
we decompose stock returns using OLS into what we refer to as common and idiosyncratic components using a simple
CAPM regression, while we decompose ROA into what we refer to as observed and unobserved components, including both
market and firms-specific factors, using a ‘‘kitchen-sink’’ approach with mixed-frequency data.
We proceed in our network analysis by constructing connectedness measures for idiosyncratic stock return components
using the vector autoregression (VAR) approach proposed by Diebold and Yilmaz (2009). However, we use a simple
correlation approach to construct connectedness measures for the observed and unobserved components of ROA since
there are insufficient quarterly observations for VAR estimation. Thus, the measures we construct span quite well the
modalities we describe in our procedure. We provide proof of concept for the resulting measures by comparing them
to other measures of interconnectedness used in the literature, interpreting their dynamics around crisis times, and
comparing them to each other.
To be clear, we are not proposing a new econometric methodology. Our contribution is to demonstrate how recently
developed methodologies can be used to produce useful measures of connectedness from large amount of available data
available at mixed frequencies. For the two steps of our procedure, we discuss a variety of options available depending on
the question of interest, outcome measures developed, and data constraints. The main innovation of our procedure is that
it allows us to separately analyze connectedness resulting from observable vs. unobservable factors or from common vs.
idiosyncratic components of the outcome variables of interest. As our example of large banks’ connectedness demonstrate,
our procedure produces measures that have intuitive dynamics and contain additional information relative to other
existing measures. Moreover, they have desired statistical properties of infrequent changes in bank rankings and smooth
dynamics. Interestingly, both stock-market-based and ROA-based measures exhibit substantial growth over the last two
years of our sample, while none of the relevant raw series and other available measures of banking system risk show
such an increase.
We begin by describing our procedure in Part 2. In part 3 we present implementation of the procedure including
data description, as well as construction, description, and analysis of the measures of systemic connectedness of the U.S.
banking system we construct. Summary and conclusions are in Part 4.

2. Methodology for variable decomposition and network analysis

Our implementation procedure consists of two steps. The first step is designed to decompose the outcome measure into
the desired components, and the second step is to construct network connectedness measures based on these components.
In this section, we discuss our procedure in general terms, and we discuss our specific implementation exercises in
Section 3.

2 Researchers have proposed a wide variety of financial stability measures, such as CoVaR (Adrian and Brunnermeir, 2016); see Bisias et al. (2012)
for a survey as well as Brave and Lopez (2017) for analysis of specific measures. Connectedness has also been studied via the interbank lending
market (Afonso and Lagos, 2015).
 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220 205

2.1. Step 1: Variable decomposition

Our analysis uses individual firm performance measures, such as weekly BHC stock returns or quarterly BHC returns
on assets, as initial data inputs, similar to Diebold and Yilmaz (2009) and related studies. However, to better examine both
the macroprudential and microprudential aspects of financial system connectedness, we propose standard decompositions
that highlight specific elements of the data.
A common criticism of connectedness measures based on firm stock returns is that they are dominated by a common
component. Within the classic CAPM framework, this caveat is clearly expressed as

yit = βi rmt + εit , (1)

where yit is a firm-specific return and rmt is the return on a market index. This equation can be applied to each firm i
to obtain firm-specific βi coefficient estimates. The term βi rmt represents the effect of the common market return, and it
can be removed to examine the idiosyncratic firm components (or the εit residuals). Thus, the first type of decomposition
we propose is into common and idiosyncratic components.
The second decomposition we consider is designed to separate out all explainable (or observable) variation in yit
relative to what cannot be explained with available data (i.e., the ‘‘unobservable’’ component). In notational form, this
decomposition is

yit = Zt′ βi + Xit′ γi + εit , (2)

where Zt is a set of common variables (such as rmt above) and Xit represents a vector of all explanatory variables available
to the researcher. The fitted values, which are a linear combination of the first two terms, can be labeled as the observable
component, and the residual εit term is the unobservable component. This approach allows us to introduce all available
information in a ‘‘kitchen-sink’’ modeling approach.
To conduct these decompositions for our analysis of banking system connectedness, we estimate individual firm
regressions, using OLS techniques if all of the variables are available at the same frequency.3 This, however, is not
commonly the case.
For this reason, Ghysels et al. (2006) propose a mixed-data sampling model based on distributed lags, commonly known
as a MIDAS regression.4 In that framework, higher-frequency explanatory variables can be included as distributed lags,
with a decay parameter estimated either non-parametrically as an unrestricted MIDAS model (i.e., U-MIDAS) as proposed
by Foroni et al. (2013), or assumed to have a decay of a specific functional form for which parameters are estimated.
For example, if yit is observed quarterly with t indicating quarters since the beginning of the sample and a subset of
Xit ∈ [Xit , Xitw ] where the latter is observed weekly, the MIDAS regression can be specified as
q

( S
)′
q′ w′
∑
yit = Xit βi + Xi(13t −s) δis γi + εit , (3)
s=1

where s is the week counter from the beginning of the sample and S = 13. Again, this equation is intended to be estimated
for each i separately.5
The discussion above does not dwell on how the filtering equation for variable decomposition is selected. The decision
might be informed by theory, a specific question that needs to be addressed, and the available data. It is important to
emphasize that this decision is key for the interpretation of the results. For example, it is unlikely that all common
components can be cleanly removed, especially since there are likely to be factors that are common to only a subset
of the observations. If desired, principle component analysis or dynamic factor models might be used to remove common
factors that are not directly observed.6 One has to be cautious using such an approach in the small-scale systems in which
idiosyncratic shocks affect system dynamics, because removing common shocks might remove a portion of idiosyncratic
ones. In terms of ‘‘kitchen-sink’’ approaches, a variety of model-selection methods are available for determining the best
fit.

3 Note that we also assume all variables are stationary in our analysis for simplicity.
4 Asymptotic and other statistical properties of the estimator are reported in Andreou et al. (2010).
5 An alternative estimation approach is a mixed frequency factor model estimated in state-space form using the Kalman filter (KF). An advantage of
this approach is not requiring any parametric assumptions; that is, one can simply cast the mixed-frequency data into state-space form and estimate
the model coefficients. Aruoba et al. (2009) demonstrate that the estimation process is valid in the cases of missing data, including data missing due
to mixed frequencies. Bai et al. (2013) develop an alternative approach using a periodic steady-state Kalman filter, as described in Assimakis and
Adam (2009). In this approach, the measurement equation consists of two parts — one at low frequency, and one at higher frequency that repeats
periodically within each low-frequency time period. Bai et al. (2013) show theoretically and with actual data that this approach produces forecasts
similar to MIDAS, including in small samples. Additional real data comparison for macroeconomic variables is conducted by Schorfheide and Song
(2009) in the context of mixed-frequency vector autoregression. A summary of the variety of approaches to estimate mixed-frequency models is
also in Foroni and Marcellino (2013).
6 These approaches are used in relation to measurement of bank linkages in Kapinos and Mitnik (2016) and Kapinos et al. (2017).
206 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220

2.2. Step 2: Measuring connectedness

The difficulty in measuring firm connectedness arises from the fact that actual linkages between firms i and j are not
commonly observed. Several approaches have been introduced in the literature using individual firms’ stock returns. The
goal is to produce a matrix of linkages with all elements ij filled, which can then be used as the adjacency matrix in
network analysis. For example, the simplest and most straightforward approach is to compute pairwise correlations for
each pair of firms. Another approach is to use a vector autoregression (VAR) model to estimate the effects of firm i on firm
j and define connectedness via forecast error variance decomposition or through a generalized impulse response function.
The VAR, of course, needs to be identified, and Diebold and Yilmaz (2009) propose to identify it with a shrinkage estimator
using the elastic net.7
Other approaches to measure the systemic risks of financial firms include conditional value-at-risk (or CoVaR) by Adrian
and Brunnermeir (2016), Marginal Expected Shortfall (or MES) by Acharya et al. (2012) and the closely related Systemic
Risk (or SRISK) by Acharya et al. (2010), Brownlees and Engle (2017), and the Distressed Insurance Premium (or DIP) by
Huang et al. (2009, 2012). Since none of these approaches are able to produce explicit measures of linkages between firms
i and j, we describe the correlation and the VAR approaches.
Correlation. We construct a matrix of pairwise correlations for the series of interest between firms i and j at a point in
time, and we use rolling windows over the full data sample to create a dynamic network. These correlations are collected
into a correlation matrix, which can be interpreted as adjacency matrix for the network, with two modifications. First,
diagonal elements need to be replaced with zeros. Second, negative elements are not meaningful in the adjacency matrix.
Thus, depending on the application, they can be replaced either with their absolute values if negative linkages are viewed
as important as positive ones (such as, for example, in social networks), or zeros if negative linkages are not as important.
If the goal is to measure contagion, one may argue that negative linkages actually provide insurance and reduce contagion,
and definitely do not contribute to contagion spread. Thus, in a systemic risk analysis it makes sense to replace negative
linkages with zeros, which would amount to ignoring any effects of insurance or diversification and thus treating it as a
conservative adjustment.
There are two main drawbacks of the correlation approach. First, it is entirely backwards-looking in that observed
correlations between past values may not be informative of how correlated they might be going forward. Second, the
correlation matrix is symmetric, which means the network described by the resulting adjacency matrix is not directional.
Accordingly, we cannot distinguish between a firm’s impact on the system and its vulnerability to the system.
Var. The VAR approach addresses both main drawbacks of the correlation approach. First, by estimating a VAR, we can
forecast future values of the outcome measure, compute forecast errors, and decompose the error variance as a measure
of linkages. While the estimation still relies on past data, at least this approach has a forecast component that makes
it predictive and not strictly backward-looking. Second, the variance decomposition matrix is not symmetric, allowing
us to construct a directional network of linkages and thus measure firm impact and vulnerability separately. Given that
components of the variance decomposition matrix are by definition between zero and one, the matrix of the variance
decomposition can be used directly as adjacency matrix upon removal of the diagonal.
The main drawback of the VAR approach is that it is data-intensive and might require longer time series than are easily
available.
Upon constructing the adjacency matrix A using one of the methods described above, we construct measures of system
connectedness, such as network density (D), and ‘‘systemic importance’’ measures for each institution, such as the degree
centrality (dc) for directed and undirected networks. Formally, these measures are defined as follows:
∑∑
2 i j ̸ =i aij
D= , (4)
N(N − 1)
∑
dci = aij , (5)
j ̸ =i

where aij is an element of A, and N is the number of firms. The network density measure is intuitively the average value
of the off-diagonal cells of the matrix, such that a sparse matrix has a low D value and a highly-populated matrix has a
high D value. The degree centrality measure similarly adds the values across the row for firm i, measuring the size of the
firm’s connectedness within the system. For interpretation, in case of correlation approach, it might be useful to divide
the dci measure by N − 1 to obtain average systemic importance of a given firm that is between 0 and 1 by construction.
A directed (or asymmetric) network allows us to distinguish directional degree centrality: out-degree (odci ) and in-
degree (idci ), which can be interpreted, respectively, as the impact of firm i on the rest of the system and the vulnerability
of firm i to shocks coming from the rest of the system. In notation,
∑
odci = aij , (6)
j̸ =i

7 This approach is also implemented in Diebold and Yilmaz (2012), Demirer et al. (2017).
 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220 207
∑
idci = aji , (7)
j ̸ =i

As before, odci is a row sum measure, but idci is a column sum measure. Note that the odci impact measure need not
equal the idci vulnerability measure, because the contribution of i to j’s forecast error variance need not be the same as
the contribution of j to i’s forecast error variance.
If the dynamics of a network measure is of interest, these calculations can be repeated for rolling windows of the
outcome measure decomposition. Note that by construction, the network density measure D is bounded to the unit
interval, regardless of how the linkages are computed. Similarly, the idci measure remains within the unit interval since
it measures the total share of forecast error variance for firm i that is explained by all other firms. In contrast, the dci and
odci measures can sum to more than one.

3. Implementation

Having described our procedure, we now turn to the concrete example of its implementation. Our example is
measuring connectedness and systemic importance in the network of large banks. We first describe our data, then
demonstrate an application with two different outcome measures: stock returns, which we measure weekly, and returns
on assets (ROA), which are available quarterly.
Since stock returns can be measured at any desired frequency, we do not have a mixed frequency problem, and
therefore employ simple linear regression to obtain our decomposition. Here we focus on common and idiosyncratic
components interpretation of decomposition and estimate an augmented CAPM regression with five Fama–French
factors (French and Fama, 1992) to separate market factors from idiosyncratic ones. With the strong common component,
we proceed with the network analysis of the residual from this decomposition, the idiosyncratic component.
With ROA, which is available quarterly, we have explanatory variables that we would like to include in the decompo-
sition that can be measured at higher frequency. Thus, we employ MIDAS approach to decompose ROA.8 Here we employ
a kitchen-sink approach, interpreting decomposition as resulting in observed and unobserved components of the ROA.
We proceed to analyze both the observed and the unobserved components.

3.1. Data

Bank supervision is a critical component of monitoring and maintaining the safety and soundness of the U.S. banking
and financial system. Dating back to the creation of the Office of the Comptroller of the Currency (OCC) in 1864,
bank examiners have collected data on-site and through reporting forms to assess compliance with federal banking
regulations (White, 2011). Starting in 1869, balance sheet data collection became more standardized with the development
of the ‘‘call’’ report of condition, which are now more commonly known as the Call Reports. After the Bank Holding
Company Act of 1956, the Federal Reserve began collecting consolidated BHC data, with the Y9 consolidated reporting
forms starting in 1978. All of this data, in conjunction with on-site examinations and other supervisory work, assists in the
creation of supervisory bank ratings, which are known as CAMELS for commercial banks and RFI/C(D) ratings for BHCs.
Statistical models that summarize supervisory information and correlate it with supervisory outcomes, such as rating
changes and bank failures, date back to the mid-1970s and are in wide use currently (King et al., 2006).9
In addition to supervisory data, market-based data is available for analysis on the largest BHCs. Pettway and Sinkey
(1980) proposed using BHC stock returns for monitoring their financial conditions. Krainer and Lopez (2004, 2008)
provide evidence using ordered logit models that both supervisory and market-based data provide leading indicators of
changes in BHC supervisory ratings. Berger et al. (2000) report similar findings, but highlight that supervisory ratings are
generally less accurate than market-based indicators in predicting future changes in bank performance, except when these
assessments occurred just after a recent on-site bank examination. All of these efforts have focused on the microprudential
goals of monitoring bank-level safety and soundness as well as allocating resource to do so. Since the financial crisis,
macroprudential concerns have arisen and been partly addresses with stress-testing of the largest BHCs (Hirtle and
Lehneart, 2015). Our efforts here work with roughly the same set of firms to derive macroprudential summary measures
of banking system vulnerability and microprudential measures of firms’ systemic importance.
We focus on the 27 large U.S. BHCs that have been subject to the Federal Reserve’s stress-testing and have data available
from 2002 to 2017, listed in Table 1. This choice of a balanced panel is intentional to allow for the network modeling.
This sample includes largest, systemically important firms that jointly account for over 75% of U.S. banking assets. We use
various data sources in our analysis, which are divided into balance sheet variables and market-based variables. Our main
market-based data is the daily stock returns of the sample, BHCs, which we aggregate to the quarterly frequency. Note
that we do not merger-adjust these returns, even in light of the large and significant BHC mergers that occurred during
2008 and 2009. Our rationale here is to focus on the specific firms involved in the Federal Reserve’s stress-testing program

8 We decided in favor of MIDAS over state-space approach because of the short time span for our data analysis.
9 While Call Reports and Y9C data are publicly available, supervisory bank ratings are confidential. In our analysis, we decided not to use
supervisory ratings for easier replicability of our results.
208 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220

Table 1
Banks included in the sample.
Ticker Bank name Total assets In MIDAS?
AXP American Express Company 181 N
BAC Bank of America 2281 Y
BK Bank of New York Mellon 372 N
BCS* Barclays 158 N
BBT BB&T Corporation 2212 Y
BBVA* BBVA Compass Bancshares 87 Y
COF Capital One Financial 366 N
C Citigroup 1842 Y
CMA Comerica Incorporated 7 2 Y
CS* Creidt Suisse Group AG 141 N
DB* Deutsche Bank US 148 N
FITB Fifth Third Bancorp 142 N
GS Goldman Sachs Group 917 N
HBAN Huntington Bancshares Incorporated 104 Y
JPM JPMorgan Chase & Co. 2534 Y
KEY KeyCorp 138 Y
MTB M&T Bank Corporation 119 Y
MS Morgan Stanley 852 N
NTRS Northern Trust Corporation 139 Y
PNC PNC Financial Services Group 381 Y
RF Regions Financial 125 N
STT State Street Corporation 238 Y
STI Suntrust Banks 207 Y
TD* TD Group US Holdings 381 N
USB U.S. Bancorp 462 Y
WFC Wells Fargo & Company 1952 Y
ZION Zion Bancorporation 66 Y

Note: Total assets are in billion USD as of 2017Q4. * indicates only U.S. assets of foreign-owned companies are
reported. ‘‘In MIDAS?" indicates whether or not the bank is included in the MIDAS measure sample.

and their network connectedness. Finally, we use S&P 500 stock market returns at the weekly frequency, effective federal
funds rate as reported by the Federal Reserve, as well as five market-wide Fama–French factors from Kenneth French.
We collect consolidated balance sheet data for the sample BHCs from the Y9-C reporting form, which is made publicly
available on a quarterly basis.10 From this source, we use firm’s total assets, income, as well as non-performing loans
(NPLs).

3.2. Measure 1: Connectedness based on stock returns

Our first example is interconnectedness of BHC’s stock returns. These are available at any frequency, but we are not
necessarily interested in intraday and day-to-day fluctuations, for this reason we choose weekly returns, but computing
log change in stock prices from Friday close to Friday close. Existing measures of stock-return-based interconnectedness
are frequency criticized for having a larger common component. We can address this criticism using our decomposition
into common and idiosyncratic components. For each bank i, we will estimate an augmented CAPM regression as follows

rit = αi + β1i rtSP + β2i rtFF + β3i SMBt + β4i HMLt + β5i RMWt + β6i CMAt + εit , (8)

where rit is the weekly stock return for firm i; r SP is a weekly return on S&P 500, r FF is federal funds rate, while the rest
of the controls are market-wide Fama–French factors. To allow for β s to vary over time, we estimate this regression for
a 3-year rolling windows from 2002 through 2017, which we roll forward quarterly.
For almost all banks, all of the coefficients estimated are of expected sign and statistically significant. Fig. 1 shows raw
returns, fitted, and residual values for Wells Fargo, as an example, since April 2016. We can see that both common and
idiosyncratic factors play an important role in the dynamics of this bank’s stock returns. The share of variance explained
by the market factors (regression R2 ) for the last rolling window is reported for all banks in Fig. 2. We can see that it
varies from 30%–50% for foreign and processing banks to as high as 80% for large and regional banks. Coefficients for each
rolling window and each bank on all the variables included in the regression are reported as time series in Appendix
figure A1.

10 https://www.federalreserve.gov/apps/reportforms/reportdetail.aspx?sOoYJ+5BzDal8cbqnRxZRg==.
 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220 209

Fig. 1. Stock returns decomposition for Wells Fargo. Note: Reported are weekly stock returns as well as fitted values and residuals from the CAPM
regression estimated for the last 3-year rolling window. Only last 6 quarters are shown for exposition purposes.
Source: Bloomberg and authors’ calculations.

Fig. 2. R2 from rolling CAPM regression (last rolling window).

Given that the fitted values only vary across banks because of differences in β s, we focus on the residuals for the
analysis of connectedness. That is, the idiosyncratic component. It should be made clear here, that the term ‘‘idiosyncratic’’
is used vaguely in this case, given that there might be banking-sector-specific common factors that will be manifested as
connectedness of the residuals. There are two reasons we choose not to remove such factors in this specific example.
First is technical — given that we are dealing with a network of only 27 institutions, removing common factors by
extracting principle components or including sector stock index in the regression has a risk of removing truly idiosyncratic
210 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220

Fig. 3. Adjacency matrix for idiosyncratic stock returns (last rolling window). Note: Reported are coefficients from variance decomposition of the
1-quarter-ahead forecast error. Diagonal is removed. Element aij shows the share of contribution of bank j to the forecast error variance of bank i.
Source: Bloomberg and authors’ calculations.

component. Second, from the regulatory point of view, correlated response of all institutions to the same shock is as
relevant as a truly idiosyncratic contagion from one institution to another.
To construct connectedness of the idiosyncratic component of stock returns at weekly frequency, we estimate a VAR
using elastic net shrinkage estimator. A similar implementation is conducted by Demirer et al. (2017) for 100 largest global
banks. Following similar methodology, we measure connectedness as variance decomposition of forecast one-quarter-
ahead error.11 To construct dynamic networks, we estimate the VAR for the same 3-year windows rolling quarterly, as
we used for the regression decomposition.
The results for the last rolling window ending in December 2017 are reported as an adjacency matrix in Fig. 3. We
can observe that the most connected pair are Credit Suisse and Deutsche Bank, while the most connected cluster is Credit
Suisse, Deutsche Bank, BBVA, and Barclays. This makes sense, because all these are foreign banks and have the largest
idiosyncratic component that likely loads on a common European factor. Toronto-Dominion bank, on the other hand, is
the least connected to the rest of the system. Time series of individual bank impact and vulnerability measures for the
four largest banks are shown in Appendix Figure A2.
The dynamics of the network are quarterly, because we roll the estimation window quarterly. Overall density is
presented in Fig. 4, where we compare density of the network of the idiosyncratic component with that of the raw stock
market returns (constructed using the same VAR approach). We can see that the contribution of the common component
makes density of raw returns networks very high throughout the sample. For the idiosyncratic returns we can see much
more pronounced dynamics — an increase in network connectedness during the Global Financial crisis that started in the
second half of 2007 as well as a pronounced increase in the last two years. This recent increase is most likely reflecting
common response of the banking system to actual and expected regulatory changes in the last couple of years.
This connectedness is not the same for all banks — Figs. 5 and 6 show measures of impact and vulnerability,
respectively, by bank business group.12 We find that during the global financial crisis the increased connectedness
was limited to the largest four banks (Bank of America, Wells Fargo, Citigroup, and JP Morgan), and regional banks.
Connectedness of other banking groups reached the level similar to that of four largest banks in 2009. This makes sense,
because this is exactly when regulatory treatment of investment banks, processing banks, and foreign banks’ affiliates
was made similar to that of the largest four banks. In recent years, the impact of investment banks seems smaller than
that of other groups, while vulnerability is lowest for foreign banks.

3.3. Comparing to other market-based measures

To see whether our approach yields estimates that are both reasonable and not spanned by existing measures, we
compare our measure to other stock-return-based measures commonly used in the literature. Fig. 7 plots standardized

11 One alternative is to use impulse response functions, as in Grant and Yung (2017).
12 Remember that impact and vulnerability here are defined narrowly as outdegree and indegree, respectively. Big4 banks are BAC, C, JPM, WFC.
Processing banks are BK, NTRS, STT. Foreign banks are BBVA, BCS, CS, DB, TD. Regional banks are BBT, CMA, FITB, HBAN, KEY, MTB, PNC, RF, STI,
USB, ZION.
 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220 211

Fig. 4. Density of the network of idiosyncratic stock returns. Note: Network density is computed as average share of 1-quarter-ahead forecast error
variance explained by other banks across all banks.
Source: Bloomberg and authors’ calculations.

Fig. 5. Network impact by bank business model. Note: Network impact is the sum across banks of the share of the 1-quarter-ahead forecast error
variance explained by a given bank. Simple average of this measure for each bank group is reported.
Source: Bloomberg and authors’ calculations.

and normalized versions of our measure as well as averages across banks of the three measures: conditional value-at-risk
or CoVaR (Adrian and Brunnermeir, 2016) for each firm i, Systemic Risk or SRISK (Acharya et al., 2010; Brownlees and
Engle, 2017), and Distressed Insurance Premium or DIP (Huang et al., 2009, 2012). These measures are commonly available
for only 11 out of our 27 banks. Thus, we repeat our entire procedure for the same 11 banks so that we have a consistent
comparison. We can see that our measure is a bit less volatile but generally has similar dynamics to other measures until
212 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220

Fig. 6. Network vulnerability by bank business model. Note: Network vulnerability is the share of the 1-quarter-ahead forecast error variance of a
given bank explained by other banks. Simple average of this measure for each bank group is reported.
Source: Bloomberg and authors’ calculations.

Fig. 7. Comparison of idiosyncratic returns connectedness measure with other measures. Note: All measures are for the following 11 banks: BAC,
BCS, BK, C, CS, DB, GS, JPM, MS, STT, WFC . Simple averages across banks. All measures are normalized and standardized for comparison.
Source: Bloomberg and authors’ calculations.

end of 2015, when our measure started to diverge from other market-based measures. Thus, we find that our measure is
consistent with others but also provides additional information.

3.4. Learning from crisis times

The Lehman crisis presents an interesting case study. While all banks performed poorly, and some failed, most of
the dynamics of their stock returns are captured by market measures. Some of the developments, however, have been
idiosyncratic. For example, only Citigroup and Bank of America received bail-out funds from the Targeted Investment
 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220 213

Fig. 8. Network impact of selected bank groups around crisis time. Note: Network impact is the sum across banks of the share of the 1-quarter-ahead
forecast error variance explained by a given bank. Simple average of this measure for each bank group is reported.
Source: Bloomberg and authors’ calculations.

Program at the end of 2008 and early in 2009. At the same time, investment banks Goldman Sachs and Morgan Stanley
were converted to Bank Holding Companies which gave them access to the same bailout funds, but with the same
regulatory strings attached, as commercial banks.
These development are reflected in our connectedness measure, as shown in Fig. 8. We can see that Bank of America
and Citigroup had a high impact measure on the rest of the system starting the beginning of the crisis in 2007Q3 and
before Lehman collapse in 2008Q3. Thus, one may argue, these were the right banks to be subject to bailout. We can
see that following the bailout the banks became less connected, especially by 2009Q3, when they were less connected
than other banks. In fact, while connectedness of other banks increased slightly through the crisis period, following the
bailout, impact of Bank of America and Citigroup declined, indicating that markets started to view them as being apart
from the rest of network.
Investment banks (Goldman Sachs and Morgan Stanley, in our sample) were designated as BHCs in the second half of
2008. We can see that this lead to substantial increase in their impact relative to the previous period. As we have seen
in previous discussion, by 2011 these two banks’ connectedness measures were not substantially different than those of
the other large banks. In Fig. 8 we can see clearly, however, that this convergence started in mid-2009.

3.5. Measure 2: Connectedness based on ROA

Our second example is interconnectedness of observed and unobserved components of banks’ return on assets (ROA),
Rit . ROA is measured as a ratio of a bank’s quarterly income to the average of bank’s assets over last three preceding
quarters. As such, Rit is measured at quarterly frequency. We use the kitchen-sink approach to separate observed portion
of the ROA for each bank from the unobserved one, acknowledging that market movements might affect book income.
To be clear, the kitchen-sink approach does not imply causal effect of explanatory variables on the outcome variable, we
simply are trying to get the best fit regression given the data available. In addition to the same set of market factors we
used in the first example, which we continue to measure at weekly frequency, we add a measure of non-performing
loans (as a share of assets), Nij , which is available quarterly. Because of this data structure, we have to use mixed-
frequency approach, which we implement with U-MIDAS, an unrestricted form of MIDAS regression which does not
require parametric assumptions.13
After extensive pretesting and observing constraints on data availability, we settled on the following regression
equation, for 16 banks, starting 2002, estimated individually for each firm i:
S
∑
ROAit = αi + βi Nit + γi (δ1si rtKBW + δ2si rtFF ) + εit , (9)
s=1

where ROAit is the quarterly return on assets for firm i; Nit is a quarterly ratio of non-performing loans to total assets;
rtKBW is the weekly return on a large banks index KBW; rtFF is weekly effective federal funds rate as defined above. We

13 We make use of R code produced by Ghysels et al. (2016) to estimate U-MIDAS.

Fig. 9. ROA decomposition for Wells Fargo. Note: Reported are quarterly return on assets as well as fitted values and residuals from the MIDAS
regression estimated for the full sample period as shown.
Source: Bloomberg and authors’ calculations.

estimate this equation via U-MIDAS for the 16 banks for which we have sufficient balance-sheet data going back to 2012.
Due to lower frequency of the dependent variable, we estimate one equation for each bank for the entire time period
and construct fitted and residual values. Even though this specification is parsimonious, it includes bank-specific, sector-
specific, and macroeconomic factors that fit the ROA quite well as shown by R2 in Fig. 10. Using the example of Wells
Fargo, Fig. 9 shows the decomposition of the ROA into observed and unobserved components. We can see that the model
captures negative returns at the peak of the crisis in 2008Q4 quite well, leaving very little to the residual. Most likely, this
is due to a combination of market impact via KBW index and bank-specific NPLs. Overall we can see that the volatility
of observed and unobserved components of ROA are quite similar. Coefficients on all variables in the regression for all
banks are reported in Appendix Figure A3.
Given that regressions include bank-specific factors, we are interested in connectedness of both, fitted values (observed
component) and residuals (unobserved component). For this case, given quarterly data and short time frame, we resort to
simple correlations as measures of connectedness. To obtain dynamic networks, we construct correlations over 16-quarter
rolling windows.
Remember that correlation methods produce a non-directional network, which is why adjacency matrices from the
last rolling window reported in Figs. 9 and 10 for observed and unobserved components, respectively, are symmetric.
Because we are interested in the measure of banking system fragility, only positive correlations are threatening to the
system, either as response to common shocks or as contagion arising from idiosyncratic shocks. Negative correlations,
instead, would serve as mitigating factors that one can think of diversification or insurance. For this reason, we replace
negative correlations with zeros instead of replacing them with absolute values.
We find that observed components are quite highly connected for a number of banks and a negative correlation arises
very infrequently. As an example, Fig. 11 shows the adjacency matrix for the last rolling window ending 2017Q4. The
least connected banks in terms of the observed component are JP Morgan, NTRS, and BBVA. Given that they all face the
same sectoral and macroeconomic conditions, this indicates that either the dynamics of their NPLs is quite different of
that of other banks, or the response of their returns to sectoral and macroeconomic factors is different from the rest of
banks in the network. BBT, Bank of America, and Citibank are, on the other hand, quite highly connected with the rest of
the network.
The unobserved components have a number of negative correlations and are generally less connected than observed
components. Fig. 12 shows the adjacency matrix for the last rolling window. Again, JP Morgan, NTRS, and BBVA are the
least connected (although unobserved component of BBVA is highly connected with that of Citigroup). Combined with
previous finding, this shows that these banks’ ROAs have moved differently in the last four years of our data from the
rest of the banks in the network. Highly connected unobserved components are apparent for Key Bank and MTB. One can
 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220 215

Fig. 10. R2 from MIDAS regression.

Fig. 11. Adjacency matrix for observed ROA component (last rolling window). Note: Reported are correlation coefficients (with negative numbers
replaced with zeros) for the fitted values from ROA MIDAS regression. Correlations are computed for the last 16 quarters of the sample.
Source: Bloomberg and authors’ calculations.

see two highly-connected clusters in the network: Wells Fargo, PNC, US Bank, and Key Bank in one, and CMA, HSBC, and
MTB in the other.
The dynamics of connectedness for the four large banks are shown in Figs. 13 and 14, for observed and unobserved
connectedness, respectively. We can see that JP Morgan stands out, with dynamics completely different from three other
large banks, for which the dynamics are more or less the same. Interestingly, for the other three large banks we observe
an increase in unobserved connectedness in recent years similar to what we observed for the market return-based
connectedness, our first measure. This is reassuring, because we applied different methodologies to different variables
in every step of our procedure.
216 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220

Fig. 12. Adjacency matrix for unobserved ROA component (last rolling window). Note: Reported are correlation coefficients (with negative numbers
replaced with zeros) for the residuals from ROA MIDAS regression. Correlations are computed for the last 16 quarters of the sample.
Source: Bloomberg and authors’ calculations.

Fig. 13. Connectedness of the observed ROA component — four largest banks. Note: Reported are average correlation coefficients (with negative
numbers replaced with zeros) for the fitted values from ROA MIDAS regression for each of the four banks with all other banks in the network.
Source: Bloomberg and authors’ calculations.

This increase in connectedness in the last two years is also visible in the overall density, as shown in Fig. 15. We
can see that, despite declining or steady connectedness of raw ROAs in recent years, both observed and unobserved
components’ connectedness is on the rise. While observed connectedness is still about half of what we observed during
the crisis, unobserved connectedness is nearly as high as the peak in early 2010. Again, this is a similar pattern to the
one resulting from Measure 1, suggesting that some underlying connectedness in bank performance that is not easily
measured otherwise, is on the rise.
 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220 217

Fig. 14. Connectedness of the unobserved ROA component — four largest banks. Note: Reported are average correlation coefficients (with negative
numbers replaced with zeros) for the residuals from ROA MIDAS regression for each of the four banks with all other banks in the network.
Source: Bloomberg and authors’ calculations.

Fig. 15. Network density for ROA and its decomposition. Note: Reported are average correlation coefficients (with negative numbers replaced with
zeros) for ROA as well as fitted values and residuals from ROA MIDAS regression.
Source: Bloomberg and authors’ calculations.
218 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220

Fig. 16. Connectedness between observed and unobserved ROA components. Note: Reported are correlation coefficients (with negative numbers
replaced with zeros) between fitted values and residuals from ROA MIDAS regression.
Source: Bloomberg and authors’ calculations.

3.6. Connectedness between observed and unobserved components

We would like to discuss one extension to our analysis. So far, we considered each component resulting from our
decomposition as completely independent. While it is true that by construction our decomposition yields components
that are orthogonal to each other for each individual institution, observed and unobserved components, for example,
may be connected across institutions. That is, if we create a 2N ∗ 2N adjacency matrix of both observed and unobserved
components, there might be interesting linkages in the off-diagonal N ∗ N submatrices. These elements are important to
measure to assess overall systemic importance of an institution.
We conduct this exercise for observed and unobserved components of ROA that we computed for our second measure.
For these, we compute 32*32 correlation matrix for observed and unobserved components for our 16 banks, using the
same 16-quarter rolling window. Because the matrix is symmetric, we only need to analyze one off-diagonal submatrix
which we can view as a cluster in the network. Fig. 16 shows this submatrix for the last rolling window in our sample.
We can see that most correlations are either negative, zero, or very small. Only a handful of correlations stand out —
it appears that unobserved component (residuals) for JP Morgan is correlated with the observed component (fitted) for
a number of banks and observed component for JP Morgan is correlated with unobserved component for the same set
of banks. This tells us that systemic importance of JP Morgan is, in fact, larger than measured by its connectedness in
observed and unobserved component networks when viewed separately. We can also note that observed and unobserved
components of ROAs for BBT and NTRS are highly correlated with each other. Finally, the unobserved component with
NTRS is correlated with observed components of most other banks in the network.
Fig. 17 shows the dynamics of this cluster density (that is, average correlation across all elements of the off-diagonal
submatrix, for which we replaced negative numbers with zeros) and compares it with the density of observed and
unobserved clusters that we described before. We find, as we expected, that correlations between observe and unobserved
components tend to be substantially lower than within observed and across unobserved components across banks. Thus,
in our case, analyzing connectedness of observed and unobserved components separately is not a bad approximation.

4. Summary and conclusions

In this paper we propose a procedure that allows us to construct systemic connectedness measures on the basis of the
data that are only available at individual level, not individual-pair level. We pay particular attention to methods available
to extract a common component, including recently developed techniques for the analysis of mixed frequency data. We
implement our procedure to construct two measures of U.S. banking system connectedness.
Bank supervision is an important public policy concern that dates back to at least the 1860s in the United States. While
the firm-specific (or microprudential) issues remain as relevant as ever, policymakers have added macroprudential (or
systemic) concerns to their supervisory responsibilities since the global financial crisis. Our measures of network density
and of firm-specific systemic impact and vulnerability should be of broad usefulness to bank supervisors.
Our methodological proposal consists of two steps that combine several recent developments in network analysis for
supervisory use. The first step based on mixed-frequency regression techniques allows us to decompose firm performance
 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220 219

Fig. 17. Density of the subcomponents of the network that includes both observed and unobserved components of ROA for each bank. Note: Reported
are average correlation coefficients (with negative numbers replaced with zeros) for fitted values and residuals as well as between fitted values and
residuals from ROA MIDAS regression.
Source: Bloomberg and authors’ calculations.

measures into a fitted (common or observed) component and a residual (idiosyncratic or unobserved) component. While
we use a particular regression specification, other users might choose to select other variables or other specification to
achieve a decomposition that suits their purpose and data structure. For bank supervisors, such decomposition provides
information when new developments outside of their standard scope of monitoring may be arising and may require
additional resources. For this study, we examine weekly BHC stock returns and quarterly ROAs, but other measures –
such as credit risk or loan losses – could be similarly analyzed. The isolation of the unobserved component is of particular
interest as it reflects factors that are not measured otherwise.
The second step of our procedure is based on network analysis, mainly of the type proposed by Diebold and Yilmaz,
Diebold and Yilmaz (2009, 2012, 2014), as well as simple correlations. We examine several measures of aggregate and
firm-specific network connectedness to provide supervisory insights on the current state of banking system stability.
There are various measures and analytics available, and we propose the use of a representative set of them; however, we
acknowledge that other users or other policy questions might require the use of other measures. We see this flexibility
as a strength of our proposed framework.
Finally, we show that measures we construct are meaningful – in terms of comparison with other existing measures
and their dynamics around crisis times – and contain additional information relative to other measures. Specifically,
both measures we construct, using different variables and approaches at every step of the procedure, show a substantial
increase in the last two years of our sample. This increase is absent from any raw data series or other systemic risk
measures available for the banking system, suggesting that our approach reveals connections between firms that are not
captured by other measures.

Appendix A. Supplementary data

Supplementary material related to this article can be found online at https://doi.org/10.1016/j.jeconom.2019.04.027.

References

Acharya, V.V., Engle, R.F., Rchardson, M.P., 2010. Capital shortfall: A new approach to ranking and regulating systemic risk. Am. Econom. Rev.: Papers
Proc. 102, 59–64.
Acharya, V.V., Pedersen, L.H., Philippon, T., Rchardson, M.P., 2012. Measuring systemic risk. CEPR Working Paper.
Adrian, T., Brunnermeir, M.K., 2016. Covar. Amer. Econ. Rev. 106 (7), 1704–1741.
Afonso, G., Lagos, R., 2015. Trade dynamics in the market for federal funds. Econometrica 83, 263–313.
220 G. Hale and J.A. Lopez / Journal of Econometrics 212 (2019) 203–220

Andreou, E., Ghysels, E., Kourtellos, A., 2010. Regression models with mixed sampling frequencies. J. Econometrics 158 (2), 246–261.
Aruoba, S., Diebold, F., Scotti, C., 2009. Real-time measurement of business conditions. J. Bus. Econom. Statist. 27 (4), 417–427.
Assimakis, N., Adam, M., 2009. Steady state kalman filter for periodic models: A new approach. Int. J. Contemp. Math. Sci. 4, 201–218.
Bai, J., Ghysels, E., Wright, J.H., 2013. State space models and MIDAS regressions. Econometric Rev. 32 (7), 779–813.
Balasubramnian, B., Palvia, A., 2018. Can short sellers inform bank supervision?. J. Financ. Serv. Res. 53 (1), 69–98.
Berger, A., Davies, S., Flannery, M., 2000. Comparing market and supervisory assessments of bank performance: Who knows what when?. J. Money,
Credit, Bank. 32 (3), 641–667.
Bisias, D., Flood, M., Lo, A., Val, S., 2012. A survey of systemic risk analytics. Ann. Rev. Financ. Econom. 4, 255–296.
Brave, S., Lopez, J., 2017. Calibrating macroprudential policy to forecasts of financial stability. Fed. Reserve Bank San Francisc. WP 201, 7–17.
Brownlees, C., Engle, R., 2017. SRISK: A conditional capital shortfall measure of systemic risk. Rev. Financ. Stud. 30 (1), 48–79.
Demirer, M., Diebold, F., Liu, L., Yilmaz, K., 2017. Estimating global bank network connectedness. National Bureau of Economic Research WP 23140.
Diebold, F., Yilmaz, K., 2009. Measuring financial asset return and volatility spillovers with application to global equity markets. Econom. J. 119 (534),
158–171.
Diebold, F., Yilmaz, K., 2012. Better to give than to receive: Predictive measurement of volatility spillovers. Int. J. Forecast. 28 (1), 57–66.
Diebold, F.X., Yilmaz, K., 2014. On the network topology of variance decompositions: measuring the connectedness of financial firms. J. Econom. 182,
119–134.
Foroni, C., Macellion, M., Schumacher, C., 2013. Unrestriced mixed data sampling (MIDAS): MIDAS regressions with unresticted lag polynomials. J. R.
Statist. Soc.A (Statistics in Society) 178 (1), 57–82.
Foroni, C., Marcellino, M., 2013. A survey of econometric methods for mixed-frequency data. norges bank WP 2013/06.
French, K.R., Fama, E., 1992. The cross-section of expected stock returns. J. Finance 47 (2), 427–465.
Ghysels, E., Kvedaras, V., Zemlys, V., 2016. Mixed freqency data sampling regression models: the R Package midasr. J. Stat. Softw. 72 (4).
Ghysels, E., Santa-Clara, P., Valkanov, R., 2006. Predicting volatility: How to get most out of returns data sampled at different frequencies. J.
Econometrics 131 (1–2), 59–95.
Grant, E., Yung, J., 2017. The double-edged sword of global integration: Robustness, fragility & contagion in the international firm network. FRB
Dallas Globalization and Monetary Policy Institute Working Paper 313.
Hale, G., Kapan, T., Minoiu, C., 2016. Crisis transmission through the global banking network. IMF Working Paper 16/91.
Hirtle, B., Lehneart, A., 2015. Supervisory stress tests. Ann. Rev. Financ. Econom. 7 (1), 339–355.
Huang, X., Zhou, H., Zhu, H., 2009. A framework for assessing the systemic risk of major financial institutions. J. Bank. Financ. 33, 2036–2049.
Huang, X., Zhou, H., Zhu, H., 2012. Assessing the systemic risk of a heterogeneous portfolio of bbanks during the recent financial crisis. J. Financ.
Stab. 8 (3), 193–205.
Kapinos, P., Kishor, K.A., Ma, J., 2017. Dynamic Comovements Among Banks, Systemic Risk, and the Macroeconomy. mimeo.
Kapinos, P., Mitnik, O.A., 2016. A top-down approach to stress-testing banks. J. Financ. Serv. Res. 49, 229–264.
King, T., Nuxoll, D., Yeager, T., 2006. Are the causes of bank distress changing? can researchers keep up?. Fed. Reserve Bank StLouis Rev. 88, 57–80.
Krainer, J., Lopez, J., 2004. Incorporating equity market information into supervisory monitoring models. J. Money, Credit, Bank. 36, 1043–1067.
Krainer, J., Lopez, J., 2008. Using securities market information for bank supervisory monitoring. Int. J. Central Bank. 4, 125–164.
Meyer, P., Pifer, H., 1970. Prediction of bank failures. J. Finance 25 (4), 853–868.
Pettway, R.H., Sinkey, J., 1980. Establishing on-site bank examination priorities: An early-warning system using accounting and market information.
J. Finance 35 (1), 137–150.
Schorfheide, F., Song, D., 2009. Real-time forecasting with a mixed-frequency VAR. J. Bus. Econom. Statist. 33 (3), 366–380.
White, E., 2011. To establish a more effective supervision of banking: How the birth of the fed altered bank supervision. national bureau of economic
research WP 16825.