Capital Regulation and Bank Failure Contagion

Daniel McKeever*

This version: September 2020

Abstract

I provide simulation evidence that the complex systems framework is well-suited for explaining
the historical distribution of U.S. commercial bank failure cascades. I use the complex systems
framework to test a model of the efficacy of microprudential (bank-level) ratio-based capital
adequacy regulations. I find that these requirements generally have the adverse effect of
increasing the likelihood of large failure cascades. This adverse effect is especially pronounced
under simulation conditions that mimic economic downturns. If banks use more leverage in
response to decreased capital adequacy requirements, the likelihood of failure cascades increases
only minimally. These results suggest that existing microprudential capital adequacy
requirements might be counterproductive to the goal of mitigating bank failure cascades, and
provide theoretical support for a countercyclical capital buffer.

__________

*Binghamton University School of Management. The author thanks seminar participants at

Binghamton University, the Northeast Regional Conference on Complex Systems (NERCCS),
and the Center for Collective Dynamics of Complex Systems (CoCo) at Binghamton University
for helpful comments, and Andrew Abbott, Venkata Sai Abhijit Duggirala, and Eric Lee for
excellent research assistance. The author bears sole responsibility for all errors and omissions.
Contact the author at mckeever@binghamton.edu.

Electronic copy available at: https://ssrn.com/abstract=3564267

 I. Introduction

Are microprudential (bank-level) capital regulations effective in reducing the overall rate

of insolvencies among U.S. commercial banks?

Regulatory efforts to reduce the likelihood of commercial bank failures suffer from an

apparent mismatch between the scope of the problem and the scope of the solution currently in

use. The financial health of large, interconnected banks is now widely understood to depend

directly on the financial health of the other large banks to which they are connected. Despite the

first-order importance of systemic risk in determining the likelihood of bank failures, however,

the capitalization of large commercial banks is still governed largely by microprudential

regulation.1

To date, most theoretical analyses of microprudential capital regulation have evaluated its

costs and benefits in the context of isolated, individual banks. 2 The costs and benefits of

microprudential capital regulation in a system of interconnected financial institutions have been

largely unexamined.

Similarly, in the years since the financial crisis, most analyses of systemic risk have

naturally focused on measuring the contribution of network topology to the risks faced by

1
Basel III, codified in 2010 in response to the financial crisis, represents a shift in focus toward macroprudential
standards, designed to internalize the systemic externalities of individual institutions’ risk-taking. The innovation in
Basel III most relevant to this study is the countercyclical capital buffer, which is discussed in Section III.b.
2
There is no clear consensus in the theoretical literature as to the efficacy of microprudential capital regulations.
The theoretical costs of capital regulation include an increased cost of capital (Baker and Wurgler (2015)),
reductions in lending (Barajas et al. (2010)) and bank profitability (Osborne et al. (2012)), as well as the potential
creation of the perverse incentive for banks to rationally increase risk-taking and the expected return on their asset
portfolios in response to these constraints (Kahane (1977), Koehn and Santomero (1980), Calem and Rob (1999),
Berger et al. (1995), Blum (1999), and Diamond and Rajan (2000), among others). The theoretical benefits of capital
regulation are equally intuitive: a reduction in the number of insolvent banks and the size of banks’ losses (Furlong
and Keeley (1989), Berger and Bouwman (2013), Osborne et al. (2012)), and are most effective in conjunction with
complementary restrictions on overall leverage (Hugonnier and Morellec (2017)). These studies, many of which
predate the financial crisis, largely focus on the relationship between capital regulation and the costs or benefits to
an individual bank, rather than the banking system as a whole.

Electronic copy available at: https://ssrn.com/abstract=3564267

interconnected financial institutions.3 Less research has been conducted with respect to the

efficacy of the current microprudential regime for capital regulation (which was largely

engineered prior to the crisis) in mitigating risk which has a significant systemic component.

This paper links these two strands of literature by developing a theoretical framework for

evaluating the effects of microprudential capital regulation in a complex system of

interconnected financial institutions. Given the importance to the real economy of maintaining

the stability of the commercial banking system, understanding the nature of the relationship

between systemic risk and microprudential capital regulation is critical to designing an effective

regulatory regime.4

A complex system is characterized by two main features: (a) a network of interconnected

nodes of some kind, where each node has a value that depends in some way on its neighbors in

the network; and (b) a time-series of shocks that disturb these nodes in some way, with the added

requirement that the disturbance created by each shock persists until such time as a new shock

disturbs the same node again. Most shocks have little effect on the system; the effect on the

individual node being shocked is small, and the ripple effect to the nodes’ neighbors in the

network is minimal. On rarer occasions, however, the cumulative effect of past shocks causes the

network to reach a “critical state”, such that even a small shock to the right node in the network

3
Silva et al. (2017) identify 266 studies of systemic risk published prior to September 2016. Of these studies, about
one-fourth (71) investigate regulation, but only a few narrowly focus on the relationship between bank capital and
systemic risk. Acharya et al. (2012) propose a regulatory capital regime (“capital shortfall”) benchmarked to a
bank’s contribution to systemic risk. Similarly, Gauthier et al. (2012)) simulate bank failures under five different
candidate macroprudential capital regimes, and find that implementing a macroprudential capital regime could
substantially reduce both individual and systemic failure rates.
4
Hart and Zingales (2014) discuss the rationale for government intervention in the banking sector specifically:
because banks serve as stores of liquidity in the real economy, bank failures are disproportionately damaging to
agents in need of liquidity, reducing aggregate demand and creating a larger welfare loss than failures of other
corporate entities.

Electronic copy available at: https://ssrn.com/abstract=3564267

can cause a massive cascade of disturbances to other nodes, transmitted by the connections of the

network (a “phase transition”). The resulting frequency distribution of the combined magnitudes

of disturbances follows a power-law: there are many small disturbances and few large ones, with

the probability of observing a given disturbance in inverse proportion to its size.5

The complex systems framework lends itself naturally to the context of evaluating the

solvency of large commercial banks, the balance sheets of which are often heavily

interconnected (creating a network of nodes) and are subjected to continuous disturbances in

asset and equity values resulting from fluctuations in asset markets. Widespread loss contagion

among financial institutions can be modeled as a “phase transition” of a complex system,

resulting from rare occurrences of small negative shocks propagating through the network

(Acemoglu et al. (2015)).

I develop a simulation model in which the actions taken by an individual bank to

maintain compliance with capital regulations directly affect the balance sheets of the banks to

which it is connected in the interbank debt network. In this model, banks are required to maintain

a ratio of capital to risk-weighted assets above a threshold established by the regulator. The

efficacy of microprudential capital regulation is evaluated as the effect of varying the threshold

5
The canonical example of this phenomenon, called “self-organized criticality”, is the “sandpile” model of Bak,
Tang, and Wiesenfeld (1987). A simulated square grid of 1x1 cells forms the floor of a virtual sandbox. Grains of
sand with height 1 and base 1x1 fall, one at a time, on random locations on the grid. In this simple model, the nodes
are the cells of the grid, connected into a network by their spatial proximity, and the random shocks result from the
falling grains of sand. When the pile on a given square reaches a height of 4 grains, it becomes unstable and topples,
distributing one grain each to four of its neighbors. If any of those neighbors is of height 3, it too becomes unstable
and topples, and the process continues until the system is at rest. The distribution of “avalanche” sizes, measured as
the total distance fallen by all grains of sand in a given interval, follows a power-law. Intuitively, this distribution
obtains because most shocks to the grid occur before the pile reaches a critical state, and therefore have only a small
effect, and only a small number of shocks strike the right spot in the pile while the pile reaches a critical state,
causing a large avalanche.

Electronic copy available at: https://ssrn.com/abstract=3564267

capital ratio on the rate of insolvencies across the entire banking system produced by the

simulation model.

The results of this simulation exercise are perhaps counterintuitive: increasing the

threshold capital ratio unambiguously results in larger cascades of insolvencies and higher

overall rates of bank insolvency. The primary insight that results from the simulation is that, in a

complex system of banks that are networked by bilateral debt exposures, microprudential capital

regulation that binds for a distressed bank has the effect of reducing that bank’s probability of

immediate insolvency, but increasing its counterparty banks’ probability of eventual insolvency.

This effect occurs as the result of the “fire sale” distribution of troubled assets from the

distressed banks to its counterparties as it sells risky assets to its creditor banks in order to reduce

the denominator of its risk-weighted capital ratio and comply with the capital regulation.

The simulation model has several free parameters, which correspond to the means and

volatilities of the returns on the banks’ risky asset portfolios, the severity of the discount applied

to troubled assets during a fire sale, and the average degree of the interbank debt network. The

adverse effects of microprudential capital regulations on the size on failure rates and the size of

failure cascades are more pronounced for parameter constellations that mimic economic

downturns (risky bank asset portfolios with low or negative means and high volatilities, more

severe fire sale discounts) and network fragility (intermediate average degree values that are

large enough to increase the likelihood of loss contagion in the interbank debt network, but too

small to reduce the risk of failure through debt portfolio diversification). These results provide

theoretical support for a countercyclical capital buffer (“CCyB”), a more recent innovation in

bank regulation that allows for microprudential capital regulations to be relaxed during periods

of widespread financial distress among banks, and enhanced during periods of growth when bank

Electronic copy available at: https://ssrn.com/abstract=3564267

leverage is expanding. I find that these results are mitigated only slightly when the model is

extended so that banks respond to lower threshold capital ratios by increasing the volatility of the

assets in their portfolios.6

The remainder of the paper is organized as follows. I provide empirical evidence for the

suitability of the complex systems framework in explaining the distribution of U.S. commercial

bank failures in Section II. The simulation model that I implement is described in Section III. I

present the main results of the simulation exercises in Section IV, and conclude in Section V.

II. The Distribution of Bank Failures

Evaluating the ability of bank regulations to mitigate the frequency of large-scale

“cascades” of many failures in a short period of time requires a more nuanced understanding of

how these cascades arise. The post-crisis systemic risk literature offers ample evidence that

cascades are determined in large part by the topology of the network that links financial

institutions (Battiston et al. (2012), Glasserman and Young (2016)). The complexity of financial

networks is likely to hinder the abilities of regulators and individual banks to correctly price the

risk of failure cascades and take steps to mitigate their likelihood (Battiston et al. (2016)).

The FDIC maintains a publicly available database of U.S. commercial bank failures

dating back to 19347. (Here, “failure” means that the bank ceases to exist as a going concern, and

its either liquidated or sold to another entity. In this case, the bank is either insolvent or in clear

danger of insolvency.) The data include the identifying characteristics of the failed bank, such as

6
These results build on the work of several recent papers that bridge the theoretical literatures in systemic risk and
prudential regulation. Notably, Gai et al. (2011), who implement a simulation model in which interconnections
between banks can foment a systemic liquidity crisis. My paper seeks to answer a similar question with respect to
system-wide insolvencies. Battiston et al. (2013) demonstrate the potential for minor reductions in the capitalization
of systemically important banks to initiate widespread failure cascades.
7
https://www.fdic.gov/bank/individual/failed/banklist.html

Electronic copy available at: https://ssrn.com/abstract=3564267

its name and the location of its headquarters, as well as the date of failure. For most entries after

1986, two additional data points are also reported: the assets held by the bank during the quarter

prior to its reported failure, and the estimated losses resulting from the failure.8

To estimate the size of a “typical” bank failure, I use the FDIC data series from 1934 to

2019, inclusive. I adjust both of the variables pertaining to failure size (prior-quarter assets and

estimated losses) for inflation, assuming a rate of 3.5% per year to convert to year-2019 dollars,

and report the distributions in Figures I and II. Both variables are lognormally distributed. The

geometric mean of the prior-quarter assets of failed banks in the sample is $151.4 million in

2019 USD, and the geometric mean of the estimated equity value lost is $29.5 million in 2019

USD.

These figures obscure an important characteristic of bank failures, which becomes

apparent only when analyzing the sizes of failure cascades, rather than the sizes of individual

failures. I analyze the distribution of failure cascades, measured as the number of failures

recorded in the FDIC data within a given interval of time (one month, one quarter, or one year).

The distribution of the sizes of failure cascades among U.S. commercial banks, measured

as the number of failures 𝑥 in a given period of time, follows a clear power-law distribution of

the form Pr⁡(𝑥) ∝ 𝑥 −𝛾 ,, which obtains regardless of the interval of time used as the unit of

analysis (Figure III). The power-law distribution is evident in a wide variety of natural

phenomena, including the magnitudes of earthquakes, the diameters of craters on the moon, and

the intensity of solar flares (Newman, 2005), and strongly suggests (although by itself does not

prove) that the underlying mechanism that produces the distribution is a “complex system”.

8
Estimated losses are calculated as the value of deposits in the failed bank reimbursed by the FDIC, minus the asset
value recovered by the FDIC from the liquidation of the bank’s assets.

Electronic copy available at: https://ssrn.com/abstract=3564267

 Much like with earthquakes or the magnitudes of the sandpile avalanche simulated by

Bak et al. (1987), there is no “typical” failure cascade size. At an annual frequency, the exponent

of the power-law distribution is -1.52; in context, this means that doubling the size of a given

failure cascade being analyzed (for example, from ten failed banks per year to twenty) makes the

larger cascade about one-third as likely to occur as the smaller cascade.9 The persistence of the

power-law distribution at multiple measurement frequencies, or “scale invariance”, is another

hallmark of complex systems.

Although the linear relationship between cascade size and frequency evident from the

log-log plots in Figure III, Panel A is not conclusive evidence of a power-law relationship

between the size and frequency of bank failure cascades, the Kolmogorov-Smirnov test statistics

reported in Figure III, Panel B support the validity of the fit of a power-law distribution. More

importantly, the simulation model that I implement in Section II shows that a similar power-law

distribution of bank failure cascade frequencies can be reproduced as the result of a complex

system of interconnected banks with plausible asset dynamics (Figure VII, VIII).

The two largest cascades of commercial bank failures in the data (the S&L crisis of the

1980s and the subprime mortgage crisis of 2007-2008) are mostly attributable to losses in banks’

risky asset portfolios, rather than loss contagion.10 The simulation model I implement in this

paper attempts to capture the mechanism by which losses on banks’ asset portfolios are

redistributed when banks are undercapitalized, including the extreme case of insolvency. The

large cascades of U.S. commercial bank failures were not caused by asset redistribution in

9
If Pr(𝑥) = 𝑥 −1.52 is the probability of a cascade of size x, then Pr(2𝑥) = (2𝑥)−1.52 = 2−1.52 Pr(𝑥) = 0.35 Pr(𝑥).
10
In explaining the bank failure cascade of 2007-2008, Jackson and Pernoud (2019) discuss the relative roles of
commonality in bank asset returns and contract-level contagion in the interbank debt network. They conclude that
the failure cascade represents a combination of both effects: failing banks had portfolios of high-risk,
underperforming mortgage loans, but were also exposed to each other’s debt losses.

Electronic copy available at: https://ssrn.com/abstract=3564267

response to capital regulation. Importantly, however, the loan portfolios of the banks that failed

during both waves were sold to other financial institutions, including other (larger) commercial

banks, at a discount (for example, in the wake of the recent financial crisis, Wachovia’s assets

were acquired by Wells Fargo, Colonial Bank’s assets by BB&T, and FBOP Corporation’s by

BBVA USA). Almost all of the 537 post-crisis bank failures in the FDIC’s database were

acquired by another commercial bank or commercial bank holding company.

Acquisitions following insolvencies represent the most extreme and most visible case of

the redistribution of the troubled assets of commercial banks. The purpose of this study is to first

develop a tractable model for bank failure cascades that incorporates both a network of interbank

debt contracts and commonality in asset returns (Section III.a.). I then use that model to assess

the consequences for systemic risk of the less visible asset redistribution channel that obtains

through banks selling risky assets to one another to remain compliant with capital regulations

(Section III.b.)

III. Simulation Model

a. The interbank network as a complex system

To simulate a distribution of bank failures resulting from a complex system of

interdependent bank values, two ingredients are necessary: (1) a model that captures the

interconnectedness of the values of items on banks’ balance sheet (the “adjacency” of the banks,

or nodes, in the network), and (2) a model for disturbances to the values of those balance sheet

items (the “shocks” to the nodes’ values).

Electronic copy available at: https://ssrn.com/abstract=3564267

 For the first component, I use the model of Jackson and Pernoud (2019) as the basis for

my model of interconnected bank values, with a few minor modifications.11 The model consists

of a set of banks indexed bank 𝑖 = 1 … . 𝑛 and a set of risky bank assets indexed 𝑘 = 1 … 𝑚. The

book value of a bank’s assets is given by:

(1) 𝑉𝑖𝑡𝐴 = ∑𝑘 𝑄𝑖𝑘𝑡

𝐴
𝑝𝑘 (𝑡) + 𝑄𝑖𝑡0 𝑝0 (𝑡) + ∑𝑗 𝐷𝑖𝑗𝑡
𝐴

𝐴
Where 0 ≤ 𝑄𝑖𝑘𝑡 ≤ 1 describes the fraction of a risky asset 𝑘 = 1 … 𝑚 that bank 𝑖 =

1 … . 𝑛 owns at time 𝑡 = 1 … 𝑇. At the beginning of each iteration of the simulation model, each

risky asset has a price of 𝑝𝑘 (0) = 1.

The risky assets represent the combined balance sheet assets of large financial

institutions, which could include personal and commercial loans, marketable investment

securities held in the bank’s portfolio, and any other tangible assets. For the sake of simplicity, I

abstract away from intangible assets in this model, and assume that all tangible balance sheet

items can be transferred, in part or in total, to other large financial institutions in the model.

To initialize the simulation, I follow Jackson and Pernoud (2019) in setting 𝑄 𝐴 to an

identity matrix of size 𝑛 = 𝑚, so that each bank starts with ownership of a “proprietary”

representative risky asset, which can then be divided and resold as necessary throughout the

simulation in response to capital adequacy constraints.12

11
Namely, I introduce the matrix of “primitive liabilities” 𝑄𝐿 to complement the matrix of “primitive assets” 𝑄 𝐴 . I
also model changes in risky asset values as a discrete approximation of a continuous-time process, rather than as a
single-period failure with an exogenously determined failure cost. Finally, because banks’ positions in peer banks’
equity are trivially small in practice, I eliminate banks’ cross-holdings in equity positions by setting the matrix of
equity cross-holdings in Jackson and Pernoud (2019) to an n x n identity matrix.
12
In this model, banks are simply portfolio managers that seek to maximize shareholders’ utility over the mean and
volatility of the return on the banks’ assets. I abstract away from behavioral considerations: managerial incentives
are irrelevant and always comply with regulations. Strategic default is not an option, as banks are never bailed out,
and managers are junior claimants on banks’ assets, receiving nothing in the event of an insolvency.

Electronic copy available at: https://ssrn.com/abstract=3564267

 The second necessary component of the model for the complexity of the system is the

method for “disturbing” or “shocking” the values of items on the balance sheet. Here, I use

standard asset pricing dynamics, with parameters estimated from publicly reported balance sheet

data for large financial institutions.

The price of the risky asset 𝑝𝑘 (𝑡) is time-varying, with dynamics given by the following

stochastic differential equation:

Δ𝑝𝑘
(2) = 𝜇𝑘𝐴 𝑑𝑡 + 𝜎𝑘𝐴 𝑑𝑊
𝑝𝑘

Where 𝜇𝑘𝐴 is the expected return of the 𝑘 𝑡ℎ risky asset over the short interval of time 𝑑𝑡,

and 𝜎𝑘𝐴 is the volatility of the 𝑘 𝑡ℎ risky asset’s returns, which multiplies 𝑑𝑊, an increment of a

geometric Brownian motion.

To allow for correlations −1 ≤ 𝜌𝑘1,𝑘2 ≤ 1 between any two risky assets 𝑘1 and 𝑘2, I

estimate the vector SDE

Δ𝑝
(3) = 𝜇𝑝𝐴 𝑑𝑡 + Σ𝑝A 𝑑𝑊𝑚
𝑝

Where 𝜇𝑝𝐴 is the vector of expected returns on all risky assets in 𝑚, Σ𝑝A is the variance-

𝐴 𝐴 𝐴
covariance matrix of risky asset returns, with elements Σ𝑘1,𝑘2 = 𝜌𝑘1,𝑘2 𝜎𝑘1 𝜎𝑘2 , and 𝑑𝑊𝑚 is the

increment of an 𝑚-dimensional geometric Brownian motion. I assume that asset price dynamics

are unaffected by the profitability or solvency of any of the banks in the model.

For each iteration of the simulation model, I generate the means and volatilities of the

banks’ risky asset returns, as well as the correlations between each pair of returns, as random

draws. The means of these random draws are free parameters, 𝜇̅ 𝐴 and 𝜎̅𝐴 . In order to choose

plausible ranges of values for these parameters, I analyze the observed returns of the 25 largest

financial institutions for which the Federal Reserve conducted supervisory stress tests in 2018

Electronic copy available at: https://ssrn.com/abstract=3564267

and for which ten years of continuous quarterly balance sheet data were available from the third

quarter of 2009 to the second quarter of 2019.13

The log asset returns of these 25 institutions (measured as 𝑟𝐴 (𝑡) = ln⁡(𝐴(𝑡)/𝐴(𝑡 − 1))

were distributed approximately normally, with a mean of 3.3% and a standard deviation of 8.0%.

The 25 bank-level return asset return means are distributed approximately normally around the

mean of the means, 3.3%, with a standard deviation of means of 3.0%. Accordingly, I generate

the 25 risky asset return means 𝜇𝐴 for each iteration of the simulation as random draws from a

normal distribution centered at 𝜇̅𝐴 . Similarly, because the variances of a sample of normally-

distributed random variables follow a chi-square distribution, I generate the 25-bank-level return

variances as random draws from a chi-squared distribution centered at 𝜎̅𝐴2 . In each case, I choose

reasonable values for the distributional parameters so that the resulting set of risky asset return

means and volatilities is comparable to the empirically observed distribution.

Correlations are generated by first taking a draw from a normal distribution, 𝑧~𝑁(𝑧̅, 𝜎𝑍2 ),

and transforming the resulting value to a correlation coefficient using the method of Fisher

(1915):

e2z −1
(4) 𝜌 = 𝑒 2𝑧 +1

Among the 25 large financial institutions that make up the sample, the distribution of

correlation coefficients (transformed to a normally distributed variable, inverting the method

above) is centered at 𝑧̅ = 0.03. Accordingly, I simulate correlations resulting from random draws

from a standard normal distribution centered on 𝑧̅ = 0. The resulting covariance matrix Σ𝐴 ,

13
Of these, only 20 are registered as commercial banks, so the remaining five are therefore not subject to the FDIC’s
capital adequacy regulatory regime, or eligible for FDIC insurance on any deposit liabilities. I use these banks’ asset
dynamics to stand in for those of other large commercial banks with incomplete data. Because the simulation only
uses the dynamics of these banks’ assets as an input, and seeks to explain the failure distribution of a hypothetical
universe of banks with these asset dynamics and a given degree of interconnectedness via debt obligations, the
results of the simulation should be unaffected by this distinction.

Electronic copy available at: https://ssrn.com/abstract=3564267

which provides a reasonable approximation of the observed covariance matrix, is populated with

the products of the simulated pairwise correlation coefficients 𝜌 and the simulated risky asset

return volatilities 𝜎𝐴 .14

The bank also holds some amount 𝑄𝑖𝑡0 of a risk-free “0th” asset, which has a price 𝑝0 (𝑡). I

treat this risk-free asset as a cash equivalent, such that its dynamics are given by 𝜇0𝐴 = 0 (so that

the “risk-free rate” in the system is normalized to 0) and 𝜎0𝐴 = 0, with a correlation equal to

𝜌0,𝑘 = 0⁡∀⁡𝑘 = 1 … 𝑚. I initialize the value of the risk-free asset to 𝑝0 (0) = 1 and the value of

each bank’s cash holdings to 𝑄𝑖𝑡0 = 0.15

The final category of assets is interbank debt obligations, where 𝐷𝑖𝑗𝐴 represents the face

value of debt owed by bank 𝑗 to bank 𝑖. These obligations could include short-term (overnight)

interbank loans, long-term unsecured interbank loans, reverse repurchase agreements, and any

other forms of fixed-term contracts between banks.16 I assume that all interbank debt assets and

liabilities on a bank’s balance sheet represent contracts with other commercial banks within the

model. This assumption is not intended to reflect empirical reality, but to allow for this model to

sketch the results of an edge case where losses are always contained and redistributed within the

14
Because the simulation method requires that the covariance matrix Σ𝐴 be positive definite, in some cases, it is
necessary to perform a Cholesky decomposition on the simulated covariance matrix, identify negative eigenvalues,
replace them with an arbitrarily small number, and recompose the matrix.
15
The model can be adjusted to allow for a non-zero risk-free rate; for simplicity, I assume away the role of inflation
over the short run and choose a cash equivalent risk-free asset with zero return.
16
For simplicity, I normalize the effective maturity of all debt items to 𝑇 = 12 months, the length of the simulation
period. Although shorter-term debt obligations with maturities less than 6 months make up the vast majority of
observed interbank loan transaction volume, these loans often represent continuing lending relationships with
repeated engagements between pairs of banks (Afonso et al, 2014), and can therefore plausibly be modeled as a
“revolving” short-term credit facility that extends over the period from 𝑡 = 1 … 𝑇. The simulation model can also be
estimated over shorter horizons (ex. 𝑇 = 30 days) without much difficulty; my approach in this study is to split the
difference between the frequencies with which interbank debt obligations mature (daily or monthly), bank capital
adequacy is assessed from SEC filings (quarterly), and the Fed conducts bank stress tests (annually).

Electronic copy available at: https://ssrn.com/abstract=3564267

system, rather than allowed to leave the system and appear on the balance sheets of other

financial institutions, such as investment banks, that are not subjected to capital adequacy

requirements. The opposite extreme edge case, a banking system with no internal interbank debt

contracting, results in an ambiguous relationship between capital regulations and failure

distributions that varies with the severity of the fire-sale discount (Figure X). Empirical reality

lies somewhere between these two poles; the purpose of this model is to examine the relationship

between capital regulation and the distribution of bank failures under the assumption that the

network structure has the potential to differentially effect the size of failure cascades.

The book value of a bank’s liabilities at time t is given by:

(5) 𝑉𝑖𝑡𝐿 = ∑𝑙 𝑄𝑖𝑙𝑡

𝐿 𝐿
𝑑𝑙 (𝑡) + ∑𝑗 𝐷𝑖𝑗𝑡

𝐿
Where 0 ≤ 𝑄𝑖𝑙𝑡 ≤ 1 describes the fraction of some liability 𝑙 = 1 … 𝑧 on the balance

sheet of bank 𝑖 at time 𝑡. These are “outside” liabilities due to an entity other than another bank

in the network, which could include deposits, loan obligations to be repaid to the Federal

Reserve, and nonbank debts (such as private loans or bond issues). I assume that these liabilities,

as well as the interbank debt obligations in 𝐷 𝐴 , return a uniform risk-free rate of 𝑟 = 0.17 The

dynamics of these outside liabilities 𝑑 are therefore identical to those of the risk-free asset 𝑝0

(i.e. zero mean and zero volatility), so that 𝑑𝑙 (𝑡) = 0.75 for all liabilities 𝑙 in 𝑧 at all times 𝑡 ≤ 𝑇.

𝐿
The debt obligations owed by bank 𝑖 to bank 𝑗 measured at time 𝑡 are given by 𝐷𝑖𝑗𝑡 , so

𝐿
that ∑𝑗 𝐷𝑖𝑗𝑡 represents the sum total of bank 𝑖’s interbank debt liabilities. I set the matrix 𝐷𝐿

17
This has been a reasonable approximation for much of the post-crisis period; the Fed’s discount rate, repo rates,
and the short-term T-bill rate all hovered close to zero from 2009 through 2016, and have returned to near-zero
levels in the early part of 2020. The unsecured interbank loans in 𝐷 𝐴 should earn a nonzero prime rate two to three
points above the discount rate in practice; in this model, I again implicitly assume that this rate is zero, simply
because the interesting variation in failure rates created by interconnectedness in bank values is driven by defaults
on delivery of the face values of these debts, not the rates of interest they could return.

Electronic copy available at: https://ssrn.com/abstract=3564267

equal to the transpose of the matrix 𝐷 𝐴 so that the model is internally consistent; a debt owed by

bank 𝑗 to bank 𝑖 appears as an asset on the balance sheet of bank 𝑖 and as a liability on the

balance sheet of bank 𝑗, and so on.

Taking the expressions for the values of the bank’s assets and liabilities together, I write

the book value of the bank’s equity at time 𝑡 as:

(6) 𝑉𝑖𝑡𝐸 = 𝑉𝑖𝑡𝐴 − 𝑉𝑖𝑡𝐿

For simplicity, I use the value 𝑉𝑖𝑡𝐸 as bank i’s Tier 1 equity capital at time t, which forms

the numerator for the threshold capital ratio. The calculation of the threshold capital ratio is

described in the next section.

b. Capital regulation

The focus of this study is the effect of microprudential capital regulation on the overall

rate of bank failure across the banking system. The financial crisis of 2007-2008 was a crisis of

insolvency, not illiquidity (Thakor (2018)); accordingly, the output variable of interest in this

simulation models pertain to the rate of insolvencies, rather than effects on asset prices or bank

profitability.

To operationalize the simulation, I begin by choosing values of four free parameters: the

average degree (number of counterparties) in the interbank debt network 𝛿; the required capital

adequacy ratio 𝑘̅, below which banks are required to enter a regime of “prompt corrective

action” through a “fire sale” of a fraction of its representative risky asset; the weight 𝑤𝐴 on the

bank’s representative risky asset in the calculation of the capital adequacy ratio 𝑘̅; and a “fire

sale” discount factor 𝛽. I also specify the capital ratio threshold for a “critically

undercapitalized” bank 𝑘, below which a bank must be liquidated; for tractability, I normalize

Electronic copy available at: https://ssrn.com/abstract=3564267

𝑘 = 0 for all simulations, which corresponds to the point at which the bank is insolvent and has

zero remaining equity value. (All simulation parameters are reported in Table V.)

With 25 banks in the simulation, the number of possible interbank debt counterparties for

a given bank is in the range 𝛿̅ = {0, 1, … ,24}. I choose a value of 𝛿̅ for each simulation and then

𝑇
generate a random interbank debt matrix 𝐷 𝐴 = 𝐷𝐿 for each simulation with 𝛿̅ as its average

degree. This matrix stays fixed for each iteration of the simulation. This matrix, which contains

elements equal to 𝐷𝑖𝑗𝐴 = 1 if bank i is owed a debt by bank j, and 𝐷𝑖𝑗𝐴 = 0 otherwise, is scaled so

that its rows sum to 0.5 to initialize the simulation. These rescaled elements correspond to the

values of the interbank debt assets held by each bank.

In all simulations with non-zero interbank debt obligations, the sum total of a bank’s
𝐴
interbank debt assets is normalized to ∑𝑖 𝐷𝑖𝑗𝑡 = ⁡0.5, while the average total of a bank’s

𝐿
interbank debt liabilities ∑𝑗 𝐷𝑖𝑗𝑡 ⁡is also 0.5, although individual banks’ values vary.18 The total

value of each bank’s assets at the initiation of the simulation is 𝑉 𝐴 (0) = 1.0 + 0.5 = 1.5. The

average total value of each bank’s liabilities at the initiation of the simulation is 𝑉 𝐿 (0) = 0.75 +

0.5 = 1.25. Book equity value is therefore 𝑉 𝐸 (0) = 0.25. The simulation results are not

especially sensitive to reasonable initial choices of asset and liability values. 19

18
The distinction results from the method for operationalizing 𝐷𝐿 as the transpose of 𝐷 𝐴 , the rows of which are
normalized to sum to 0.5.
19
In my simulation, all banks are initialized with an identical asset value of 1.5 units. This abstraction facilitates the
implementation and interpretation of the model’s results, but does not match the empirical reality; even among the
large financial institutions I use to parameterize my simulation, the largest is 20 times the size of the 25 th-largest in
terms of balance sheet assets. Incorporating differential equity values into the simulation is an avenue for future
research.

Electronic copy available at: https://ssrn.com/abstract=3564267

 The regulatory regime for capital adequacy is defined by the other three free parameters

̅ and 𝛽.20 If a bank’s Tier 1 capital ratio falls below 𝑘̅, it is considered “undercapitalized”,
𝑤𝐴 , 𝑘,

and must respond by taking what the FDIC calls “prompt corrective action” to improve its

capital position. It does this selling a fraction 1 − 𝛼 of its proprietary risky asset, 0 < 𝛼 ≤ 1. It

sells this fraction of its risky asset (in exchange for the riskless “cash-like” asset) to its creditor

banks, which are the banks to which the undercapitalized bank has outstanding debt obligations

(the non-zero elements of the row of 𝐷𝐿 corresponding to the undercapitalized bank).21 In this

model, all interbank loans are thus functionally collateralized by the bank’s proprietary risky

asset.

This asset sale is being conducted at a time of financial distress for the undercapitalized

bank; as a result, the creditor banks purchasing the asset may demand a “fire-sale” discount of

20
The regulatory regime I implement in this simulation is a stylized representation of the regime used by the FDIC.
The FDIC distinguishes between five categories of capitalization based on Tier 1 capital ratios, ranging from “well
capitalized” (Tier 1 ratio of 8% or greater) to “critically undercapitalized” (Tier 1 ratio of 2% or less). The
numerator of the ratio is “Tier 1” equity capital, which is the sum of the values of the bank’s common stock and
retained earnings. The denominator of the ratio is a risk-weighted average of the bank’s asset values. The risk
weights that I apply to the simulation banks’ assets are likewise simplified for the sake of tractability; in practice, the
risk weights assigned to balance sheet assets range from 0% (U.S. and OECD government bonds) to 1,250% or more
(structured or securitized products). Again for simplicity, I also abstract away from the additional restrictions on
bank balance sheets imposed by (a) fractional reserve ratio requirements for banks’ cash holdings relative to
deposits, (b) non risk-weighted leverage restrictions imposed under Basel III, and (c) additional “capital buffer”
requirements that are combined with the Tier 1 ratio. The model I develop is designed to capture the salient features
of the regulatory regime applied to commercial banks, while still being empirically tractable via simulation.
21
The bank could also respond by raising new equity capital to increase its Tier 1 capital ratio to exceed the required
threshold; FDIC guidance discusses both possibilities in terms of banks’ actions over time (“increased capital
formation” versus “reduced asset growth”). Some empirical evidence suggests that undercapitalized banks favor
shifting their payout policies away from dividends and toward earnings retention, as opposed to selling risky assets
(Cohen (2013)). In a version of the simulation model that allows for raising equity capital, resolving financial
distress is trivial, and the resulting effects on the distribution of bank failures are uninteresting; banks simply raise as
much equity capital as necessary each time they become “undercapitalized”. In this model, I assume that banks are
financially constrained, pay no dividends, and cannot raise new capital at an affordable rate of required return. This
has some intuitive appeal: undercapitalized banks have, by definition, experienced a negative rate of return on equity
over the simulation period, and are unlikely to be able to provide an adequate return on equity to compensate for
their current level of risk in the near future. The literature on troubled banks provides ample evidence of banks
responding to undercapitalization with asset fire sales (Campello et al. (2010); de Haan and van den End (2013);
Ramcharan (2020), Diamond and Rajan (2011)), which may be optimal when compared to issuing new equity
(Hyun and Rhee (2011)).

Electronic copy available at: https://ssrn.com/abstract=3564267

0 < 𝛽 ≤ 1 that multiplies the asset’s current price 𝑝𝑘 (𝑡). Conceptually, a lower value of 𝛽 could

represent greater risk aversion on the part of creditor banks’ shareholders, an increased level of

desperation to remain solvent on the part of the troubled bank, greater negotiating leverage for

creditor banks, or some combination of the three. In implementing my simulation, I use values of

𝛽 ranging between 0.9 (an asset is sold for 90 cents on the dollar) and one (an asset is sold with

no fire-sale discount).22

In my model, the bank’s Tier 1 capital ratio at time t is given by

(7) 𝑘𝑖𝑡 = 𝑉𝑖𝑡𝐸 /[𝑤0 𝑄𝑖𝑡0 𝑝0 (𝑡) + 𝑤𝐴 (∑𝑘 𝑄𝑖𝑘𝑡

𝐴 𝐴
𝑝𝑘 (𝑡) + ∑𝑗 𝐷𝑖𝑗𝑡 )]

Where 𝑤0 is the regulatory risk weight assigned to the bank’s position in the risk-free

(cash) asset, and 𝑤𝐴 is the regulatory risk weight assigned to both the bank’s proprietary risky

asset and its interbank debt assets. For the sake of simplicity, I set 𝑤0 = 0 and 𝑤𝐴 = 1 for all

simulations. The risk-free asset does not count against the denominator for the bank’s Tier 1

capital ratio, and so an asset fire sale has the effect of reducing the denominator of 𝑘𝑖𝑡 by

exchanging a risky asset for a risk-free one. For any fire-sale discount 𝛽 < 1, the asset sale also

reduces the bank’s equity value; accordingly, the bank identifies the smallest possible fraction

(1 − 𝛼) that it can sell in order to become compliant with 𝑘̅. After the asset sale, the bank’s

newly compliant Tier 1 capital ratio is given by:

′
(8) 𝑘̅ ≤ 𝑘𝑖𝑡 = 𝑉𝑖𝑡𝐸 /[𝑤0 (𝑄𝑖𝑡0 + 𝛽(1 − 𝛼) ∑𝑘 𝑄𝑖𝑘𝑡
𝐴 𝐴
𝑝𝑘 (𝑡))𝑝0 (𝑡) + 𝑤𝐴 (𝛼 ∑𝑘 𝑄𝑖𝑘𝑡 𝑝𝑘 (𝑡) +

𝐴
∑𝑗 𝐷𝑖𝑗𝑡 )]

22
I present a representative balance sheet of an undercapitalized bank in my simulation during a fire sale in Figure
VI. Although each bank begins the simulation with zero units of the risk-free (cash) asset for the sake of simplifying
the model, the creditor banks to the undercapitalized bank nonetheless purchase a fraction of the undercapitalized
bank’s risky asset using cash. I record the cash used to finance this sale as a risk-free “outside” debt liability on the
creditor bank’s balance sheet, so the purchase of the troubled asset is effectively being financed with a zero-interest
loan from a non-network bank (which could include the Federal Reserve). This has an identical effect on the creditor
bank’s equity position and its capital adequacy as reducing some existing store of cash.

Electronic copy available at: https://ssrn.com/abstract=3564267

 ′
Where 𝑉𝑖𝑡𝐸 is equal to 𝑉𝑖𝑡𝐸 reduced by the difference between the value of the fraction of

the risky asset sold and the cash received in the fire sale:
′
(9) 𝑉𝑖𝑡𝐸 = 𝑉𝑖𝑡𝐸 + (𝛽 − 1)[(1 − 𝛼) ∑𝑘 𝑄𝑖𝑘𝑡
𝐴
𝑝𝑘 (𝑡))]

If a bank is in such extreme distress that its Tier 1 capital ratio falls below 𝑘 ≤ 𝑘̅, it is

considered “critically undercapitalized”, and ceases to be a going concern. At this point, the

bank’s assets are distributed to its creditors according to an “orderly liquidation” regime, as in

Jackson and Pernoud (2019), such that each creditor bank 𝑗 ∈ 𝑐 receives a fraction of the

liquidated bank i’s assets (including cash, the representative risky asset, and interbank debt

assets) whose total value is in proportion to the creditor bank’s share of the liquidated bank’s

debts, 𝐷𝑖𝑗𝐿 / ∑𝑐 𝐷𝑖𝑐

𝐿
. Shareholders are junior claimants on the liquidation value of the firm and

receive nothing.

The liquidated bank’s deposits must also be relocated. In my model, the deposits of the

liquidated bank i are divided proportionally among the liquidated bank’s debtors 𝑗 ∈ 𝑑,

according to the fraction 𝐷𝑖𝑗𝐴 / ∑𝑑 𝐷𝑖𝑑

𝐴
. This is conceptually equivalent to the depositors of the

failed bank being reimbursed in full by FDIC insurance, and then choosing to relocate their

deposits in proportion to their previous bank’s debt obligations.23 The liquidated bank’s

interbank debt liabilities are zeroed out on the asset side of the balance sheets of creditor banks.

To model the dynamics of the complex system, I simulate 1,000 iterations of the risky

asset price paths for each combination of values for a set of free parameters and observe the

23
While this solution does not map neatly to the real-world setting, it has some intuitive appeal: depositors with full
insurance should be indifferent between a set of identical commercial banks paying zero interest on deposits.
Furthermore, the bank’s debt to the depositor is functionally held by the bank’s own debtors: if Bank A owes a
dollar to Depositor X, but Bank B owes a dollar to Bank A, then Depositor X is ultimately a creditor to Bank B.

Electronic copy available at: https://ssrn.com/abstract=3564267

distributions of (1) bank equity returns and (2) bank failures, defined as the fraction of banks that

become insolvent (𝑘 = 𝑘 = 0). I report the results in Section IV.

The power-law distribution is immediately apparent in the log-log plot the distribution of

failures from a typical simulation (Figures VII, VIII). The consistency between the power-law

distribution produced by the simple simulation model and the observed power-law distribution of

commercial bank failures strongly suggests that the complex systems framework is the right lens

through which to conduct analyses of bank failures.

IV. Results

a. Capital thresholds

If bank-level capital adequacy ratios generally function as intended by regulators, there

should be a negative relationship between capital ratio thresholds for prompt corrective action

and the overall failure rate in the banking system. Under this hypothesis, there would also be a

positive relationship between the prompt corrective action threshold and the value of the power-

law exponent in the distribution of failures (where a larger exponent value corresponds to a

lower frequency of large failure cascades). Conversely, if bank-level capital adequacy

requirements have unintended, adverse consequences, it may be that the opposite results (a

higher frequency of failures and of large failure cascades) are associated with higher capital

thresholds for prompt corrective action.

Implementing the simulation algorithm described in Section III over a range of capital

thresholds 𝑘̅ = {0, … , 0.15}, I find that higher required capital ratios 𝑘̅ are associated with a

greater likelihood of large failure cascades, as measured by both the total failure rate per

simulation and by the exponent of the power-law distribution of failure frequencies estimated

from the simulation (Figures VIII, XI, XII, XIII).

Electronic copy available at: https://ssrn.com/abstract=3564267

 This seemingly counterintuitive result has a very intuitive explanation in the context of

complex systems. The prompt corrective action regime serves to temporarily reduce the

immediate likelihood of an individual undercapitalized bank’s insolvency, but increases the

eventual likelihood of a cascade of multiple banks insolvencies.

When a bank removes a troubled asset from its books through a fire sale, the risk

associated with the asset is not eliminated, but displaced elsewhere within the banking system.

Comparing the characteristics of the assets of surviving and failing banks from the simulation

model results, failed banks’ risky assets are significantly more likely to have a low (or negative)

mean return and/or a high volatility when compared with surviving banks (Figure XIV). As a

result of the asset sale necessitated by the individual troubled bank breaching the capital

threshold, shares of this troubled asset have now been distributed to the undercapitalized bank’s

creditor institutions, in exchange for some (possibly discounted) amount of the safe asset. With a

troubled asset making up some portion of their balance sheets, the creditor banks are now each

more likely to see a further reduction in the asset’s value and become insolvent.

Failure cascades are also more likely with lower values of 𝛽, which represent steeper

discounts for the fire sale of the undercapitalized bank’s risky asset (Figure XI). Receiving a

lower fire sale price for its risky asset means that the undercapitalized bank receives a smaller

infusion of cash, and its book values of assets and equity are correspondingly lower than in the

case with no discounting. This puts the bank at greater risk of falling below the threshold for

prompt corrective action again in the near future, which in turn results in further fire sales and

further redistribution of the undercapitalized bank’s toxic asset throughout the system.

The adverse consequences of higher capital thresholds with even a small fire sale

discount are also evident in the frequency of large cascades, not just the overall failure rate

Electronic copy available at: https://ssrn.com/abstract=3564267

(Figure VIII). Comparing the top left and bottom left panels of Figure VIII, which correspond to

the simulation environment under 𝛿̅ = 0, the effect of transitioning from no capital threshold

(top left panel) to a capital threshold of 𝑘̅ = 0.075 is minimal; the slope of the log-log plot is

basically unchanged (-2.9065 in the top left panel to 2.8533 in the bottom left panel). The top

right and bottom right panels correspond to a fragile network topology (𝛿̅ = 12) with a modest

fire sale discount (𝛽 = 0.95). Here, the adverse effect of introducing capital thresholds is

apparent from the shift in the slope of the log-log plot (-2.0097 in the top right, -1.4767 in the

bottom right panel).24 Larger cascades are much more frequent as losses are redistributed within

the system under prompt corrective action.

b. Network structure

A heavily interconnected financial system is frequently equated to a structurally fragile

financial system, which obscures the importance of network topology and the role of

diversification.

In my simulation model, the failure rate is maximized at low- to intermediate values of

the average network degree. Using a single-period model with random, exogenous negative

shocks to individual institutions’ solvency, Elliott et al. (2014) identify a “sweet spot” for

maximizing failure cascades among interconnected institutions at intermediate values of average

network degree. In this “sweet spot”, institutions are sufficiently interconnected to allow for

contagion of losses through defaults on debt, but insufficiently interconnected to reap the

benefits of debt portfolio diversification. I obtain the same “sweet spot” using a multi-period

implementation of the simulation model (Figure IX).

24
Note that a value of gamma closer to zero corresponds to a higher probability of large failure cascades.

Electronic copy available at: https://ssrn.com/abstract=3564267

 In this simulation model, the negative relationship between bank debt portfolio

diversification and the likelihood of insolvency is true at both the level of the individual bank as

well as at the level of the banking system. For any given value of the average interbank debt

network degree 𝛿̅, the random allocation of the distribution of the number of counterparties for

an individual bank’s debt obligations (both as a creditor and a debtor) within the model produces

variation in the extent to which banks are diversified against counterparty risk. Banks with few

debtors in the interbank debt network are poorly diversified against the risk that a given debtor

becomes insolvent and is unable to repay some or all of its debt, while banks with many debtors

are relatively well diversified against this risk. I find that less-diversified banks are also more

likely to fail in this simulation model than their better-diversified counterparts (Figure XIV).

The effect of network topology is also apparent in the exponent of the power-law

distributions of failures produced by the simulation (Figure VIII). Comparing the top left and top

right panels of Figure VIII demonstrates the effect of introducing a fragile network topology on

the likelihood of large failure cascades in the absence of capital regulations. Increasing the

average network degree from 𝛿̅ = 0 to 𝛿̅ = 12 causes the slope of the log-log plot to shift from

-2.9065 to -2.0097, corresponding to a greater frequency of large failure cascades. The same

effect appears in the presence of capital thresholds (bottom left and bottom right panels).

c. Asset return parameters

Unsurprisingly, decreasing the mean return on the risky asset for an average bank in the

simulation 𝜇̅𝐴 increases the likelihood of failure cascades (Figure XII). Because bank-level risky

asset return means are a draw from a normal distribution centered at 𝜇̅ 𝐴 , a lower mean leads to a

larger proportion of banks with low or negative means. Assets with low or negative means are

likely to drive a bank below the capital threshold and trigger a fire sale, and are correspondingly

Electronic copy available at: https://ssrn.com/abstract=3564267

less likely to produce positive returns over subsequent periods, spreading losses to creditor

banks. Similarly, increasing the volatility of the risky asset for an average bank in the simulation

𝜎̅𝐴 also increases the likelihood of failure cascades (Figure XIII). More volatile assets naturally

produce more extreme negative returns, and so are more likely to cause a bank to breach the

capital threshold and trigger a fire sale, which in turn increases the likelihood that a creditor bank

that purchases the asset experiences extreme negative returns and is itself pushed below the

capital threshold.

Taken together, these conditions (low or negative mean returns on risky assets, increased

volatility of risky assets, and a steeper fire sale discount) are representative of the banking

environment during economic downturns.25 Ideally, these are precisely the conditions under

which microprudential regulations such as capital thresholds should fulfill their intended purpose

and be most effective in preventing failure cascades.

The results of this simulation model suggest that exactly the opposite result is possible:

microprudential capital regulations have a positive effect on the probability of bank solvency

when asset liquidations are conducted at full price (Figure XI) and when banks operate without

interbank debt (Figure X), and have a smaller negative effect on the probability of bank solvency

when asset returns are larger and less volatile (Figure XII, XIII). Under less-ideal conditions,

however, the distribution of risky assets throughout the system that results from prompt

corrective action is counterproductive to the goal of mitigating bank failure cascades over the

remainder of the simulation period.

25
The effects on banks’ risk of insolvency from changes to variables including the Dow Jones Total Stock Market
Index, the yield on U.S. BBB corporate bonds, and the U.S. Market Volatility Index (VIX), the growth rate in real
GDP, and the unemployment rate are used in the annual stress tests of systemically important financial institutions
conducted by the Federal Reserve as a requirement of the Dodd-Frank Act.

Electronic copy available at: https://ssrn.com/abstract=3564267

 These results provide theoretical support for implementing the countercyclical capital

buffer (“CCyB”) established as a part of Basel III. The CCyB allows for the central banks that

voluntarily agree to comply with the Basel accords to require that banks retain an additional

percentage of equity capital during leverage expansions, reserving the ability to turn the buffer

requirement back off in the event of a downturn. The primary rationale for the CCyB is that

turning the buffer on during economic expansions allows system banks to reap the accumulated

benefits of additional risk-weighted capital once a downturn occurs. The results of this study

suggest a complementary rationale for turning the CCyB switch back off during contractions: the

adverse consequences of prompt corrective action are most severe under conditions

approximating a downturn. The costs and benefits of the CCyB in the U.S. banking system are

still purely theoretical: the Federal Reserve’s Board of Governors has left countercyclical capital

buffers at zero percent in each of its annual votes since the implementation of the CCyB under

Basel III.

d. Endogenous bank risk-taking

The basic model described in II.b. and the analysis of the effect of capital adequacy

requirements discussed so far in this section both treat the dynamics of banks’ asset values as

exogenously-determined parameters that are unaffected by the capital adequacy requirement or

the financial health of the lending bank. It should be noted, however, that these parameters were

estimated from historical data that reflect the choices made by banks with respect to leverage and

asset portfolio volatility under a given set of capital requirements. The observed distribution of

reported Tier 1 capital ratios among solvent banks is therefore truncated from below at a ratio

greater than the minimum of 6 percent required under Basel III (see Figure IV).26

26
Another explanation is that risk-weighted capital ratios are incorrectly specified, and do not form a binding
constraint on banks’ asset portfolio selection (Cathcart et al. (2015)). U.S. commercial banks keep capital in excess

Electronic copy available at: https://ssrn.com/abstract=3564267

 It is unlikely, but still possible, that these capital regulations represented a binding

constraint on the banks’ optimization problem when selecting portfolio assets. 27 Under less-

stringent capital requirements, the likelihood of decreasing asset values causing a bank to drop

below the threshold for regulatory action is lower, and so a given bank might choose a larger

weight on higher-volatility loans in its asset portfolio. More volatile asset portfolios would

mechanically increase the likelihood that an individual bank would fail.28

It is therefore possible that additional bank risk-taking induced by weaker capital

requirements could also result in more failure cascades. This stands in direct opposition to the

simulation results discussed so far., which suggest that the prompt corrective action mechanism

serves to spread toxic assets more broadly throughout the system, temporarily saving an

individual troubled bank, but increasing the likelihood that many other banks will fail in the near

future. Accordingly, I use my simulation model to test whether one of these two effects of

weaker capital requirements (induced risk-taking or lessened risk-spreading) dominates.

of required regulatory minimums, and empirical evidence suggests that they manage these ratios toward
predetermined targets (Berger et al. (2008)).The bank capital literature offers several explanations for this behavior,
including reducing the expected costs of bankruptcy (Allen et al. (2015)), reducing agency problems with respect to
loan monitoring (Allen et al. (2011)), and loss aversion induced by the bank’s experiences during previous crises
(Bouwman & Malmendier (2015)).
27
The concern raised in this section is only of first-order importance if the leverage constraint imposed by the Tier 1
capital requirement is binding with respect to the bank’s portfolio optimization problem. The distribution in Figure
IV is truncated from below at 6%, and has a mean of 13.02% and a standard deviation of 2.59%. Assuming
normality and adjusting for this truncation to find the “true” moments of the distribution yields a mean of 12.99%
and a standard deviation of 2.63%. If banks’ optimal choices over asset portfolios are reflected by a non-truncated
normal distribution with these moments, the probability of a bank choosing a portfolio that results in a Tier 1 capital
rat io of less than 6% is approximately 0.4%, indicating that the constraint on asset selection posed by the Tier 1
capital adequacy requirements under Basel III is generally non-binding. The early years of the sample period (2010
to 2019) correspond to the Basel II regime and the phase-in of Basel III implementation; under Basel II, the
minimum Tier 1 ratio was even lower, at 4.5%, which provides further evidence that required Tier 1 capital ratios do
not represent a binding constraint with respect to banks’ portfolio optimization strategies.
28
Consider the edge case where every bank held only the risk-free asset. In this version of the model, the failure rate
would be zero by definition. For each bank, the rate of return on assets would be 𝜇 𝐴 = 𝑟 = 0, the rate of return on
equity would be 𝜇 𝐴 ∗ (𝑉 𝐴 /𝑉 𝐸 ) = 𝑟 = 0, and the risk-weighted Tier 1 capital ratio would be undefined.

Electronic copy available at: https://ssrn.com/abstract=3564267

 To do so, I need a tractable model of the relationship between banks’ risk-taking and

required capital ratios. I begin by assuming that each bank has chosen a set of portfolio weights

on all of the potential individual assets (mortgages, commercial loans, student loans, credit cards,

and so on) within its investment opportunity set, such the weighted-average mean and volatility

of its portfolio asset returns are 𝜇 𝐴 and 𝜎𝐴 , respectively.

I next assume that each bank in the sample is currently choosing an optimal set of

weights on assets that maximizes a utility function over 𝜇 𝐴 and 𝜎 𝐴 belonging to a representative

bank shareholder, subject to a given tolerance for falling below 𝑘̅, the capital threshold for

prompt corrective action.29 This representative bank shareholder desires the maximum attainable

Sharpe ratio with respect to the bank’s asset returns, (𝜇 𝐴 − 𝑟)/𝜎 𝐴 , subject to the means and

covariances of the available assets and the constraint on leverage imposed by 𝑘̅. I therefore

assume that the observed moments of the bank’s asset return distribution correspond to this

maximum attainable Sharpe ratio for each bank.

Assume that the leverage constraint imposed by 𝑘̅ on the bank’s portfolio selection
∗
problem is binding, so that the observed level of asset portfolio risk 𝜎 𝐴 is below the level 𝜎 𝐴

that the bank would optimally choose in the absence of the constraint. Further assume that, in the

familiar CAPM setting (Sharpe, 1964), this unconstrained optimum corresponds to a point on the
∗
capital allocation line with 𝜇 𝐴 > 𝜇 𝐴 . Conceptually, the bank is increasing its leverage to invest

more in the maximally efficient portfolio corresponding to (𝜇 𝐴 , 𝜎 𝐴 ). Because the Sharpe ratio

gives the slope of the capital allocation line, this new optimal portfolio must have the same

29
This tolerance corresponds to a value-at-risk probability. For example, if a given bank is willing to tolerate a 5%
chance of dropping below the threshold for prompt corrective action in one year, it would choose a set of individual
asset weights over the set of feasible “representative” asset return distribution parameters (𝜇 𝐴 , 𝜎 𝐴 ) such that 𝑍 =
(𝑥 𝐴 − 𝜇 𝐴 )/𝜎 𝐴 ≤ -1.645, where 𝑥 𝐴 = −(𝑘̅ − 𝑘)/(𝑘̅ − 1)⁡ is the return on the bank’s assets that results in 𝑘 ≤ 𝑘̅ .

Electronic copy available at: https://ssrn.com/abstract=3564267

 ∗ ∗
Sharpe ratio as the portfolio chosen in the constrained setting: (𝜇 𝐴 − 𝑟)/𝜎 𝐴 > (𝜇 𝐴 − 𝑟)/𝜎 𝐴 . If

the risk-free rate is assumed to be zero, then the ratio of the mean and volatility of each portfolio

is identical, and the mean and volatility of the unconstrained portfolio’s asset returns can be

expressed as the corresponding parameters of the constrained portfolio, scaled by a constant 𝜆:

∗
𝜇𝐴 𝜆 𝜇𝐴
(10) ∗ = ( ) ( 𝐴)
𝜎𝐴 𝜆 𝜎

I evaluate the effect of increasing the bank leverage parameter 𝜆 on the failure cascade

distribution parameter 𝛾 in the edge case where capital requirements are non-existent above and

beyond the bank remaining solvent, 𝑘̅ = 𝑘 = 0, using the simulation approach described in

Section II.b. I run the simulation over a set of leverage parameters 𝜆 = {1 … 1.2}, keeping all

other model simulation inputs unchanged. This allows for the cleanest test of whether additional

risk-taking induced by weaker capital adequacy requirements dominates the systemic risk-

mitigating effect.

I find that increased bank leverage induced by lower required minimum capital ratios

indeed accelerates the rate of failures in the banking system, although the incremental effect of

leverage is relatively small (Figure XV). For an average interbank debt network degree of 𝛿̅ =

12, a ten percent increase in leverage (𝜆 = 1.1) corresponds to an increase in the predicted

failure rate from 13.8% to 15.2%, and a twenty percent increase in leverage (𝜆 = 1.2)

corresponds to a predicted failure rate of 16.3%. These estimated increases are comparable to the

effects of introducing capital thresholds of 𝑘̅ = 0.045 and 𝑘̅ = 0.75, respectively, in the base

case with 𝛽 = 0.95 and 𝛿̅ = 12. In this simulation model, the additional failures resulting from

increased risk-taking induced by setting 𝑘̅ = 0 are minimal compared to the failures resulting

from contagion generated by asset fire sales under prompt corrective action.

Electronic copy available at: https://ssrn.com/abstract=3564267

 V. Conclusion

I find that bank failure cascades can plausibly be modeled as the result of a complex

system reaching a critical state that depends heavily on the interconnectedness between the

balance sheets of the banks in the system. As a result, bank-level measures that seek to predict

large failure cascades are prone to sizable errors, especially when the diversification of the

interbank debt network is relatively low.

The complex origins of failure cascades also limit the efficacy of ratio-based capital

adequacy requirements, which can have unintended, adverse consequences for the frequency of

large cascades of failures. These risk-weighted ratios, designed to maintain an adequate store of

equity capital on banks’ balance sheets relative to the volatility of their assets’ market values,

actually increase the likelihood of bank failures and failure cascades when banks’ troubled assets

are sold at even a small fire sale discount, or when the interbank debt network is only moderately

diversified. These effects are attenuated if banks are induced by decreased capital adequacy

requirements to take greater risks in their asset portfolios, but the likelihood of capital adequacy

requirements serving as a binding constraint on banks’ portfolio optimization problems appears

to be very low. Taken together, these results provide a new avenue of theoretical support for the

use of the countercyclical capital buffer established in Basel III.

The complex systems framework for analyzing bank failures is ripe with opportunities for

future research. For the sake of tractability, this paper assumes equally-sized banks, a simplified

risk-weighting scheme, and a single regulatory mechanism (bank-level risk-weighted capital

ratios), leaving considerable room for further inquiries into the effects on the distribution of

failures from altering these and other assumptions. Given the societal and economic importance

Electronic copy available at: https://ssrn.com/abstract=3564267

of bank failures, future research into alternative regulatory techniques that better incorporate the

complexity of the banking system is necessary.

Electronic copy available at: https://ssrn.com/abstract=3564267

References

Acemoglu, D., Ozdaglar, A., & Tahbaz-Salehi, A. (2015). Systemic Risk and Stability in
Financial Networks. American Economic Review, 105(2), 564-608.

Acharya, V., Engle, R., & Richardson, M. (2012). Capital Shortfall: A New Approach to
Ranking and Regulating Systemic Risks. American Economic Review, 102(3), 59-64.

Afonso, G., Kovner, A., & Schoar, A. (2014). Trading Partners in the Interbank Lending Market.
Federal Reserve Bank of New York Staff Reports, 620, 1-52.

Allen, F., Carletti, E., & Marquez, R. (2011). Credit Market Competition and Capital Regulation.
The Review of Financial Studies, 24(4), 983-1018.

Allen, F., Carletti, E., & Marquez, R. (2015). Deposits and Bank Capital Structure. Journal of
Financial Economics, 118(3), 601-619.

Bak, P., Tang, C., & Wiesenfeld, K. (1987). Self-Organized Criticality: An Explanation of 1/f
Noise. Physical Review Letters, 59(4), 381-384.

Baker, M., & Wurgler, J. (2015). Do Strict Capital Requirements Raise the Cost of Capital?
Bank Regulation, Capital Structure, and the Low-Risk Anomaly. American Economic Review:
Papers & Proceedings, 105(5), 315-320.

Barajas, A., Chami, R., Cosimano, T., & Hakura, D. (2010). U.S. Bank Behavior in the Wake of
the 2007-2009 Financial Crisis. Working paper.

Battiston, S., Caldarelli, G., May, R., Roukny, T., & Stiglitz, J. (2016). The Price of Complexity
in Financial Networks. PNAS – Proceedings of the National Academy of Science, 113(36),
10031-10036.

Battiston, S., Delli Gatti, D., Gallegati, M., Greenwald, B., Stiglitz, J. (2012). Default Cascades:
When Does Risk Diversification Increase Stability? Journal of Financial Stability, 8(3), 138-
149.

Battiston, S., di Iasio, G., Infante, L., & Pierobon, F. (2013). Capital and Contagion in Financial
Networks. Working paper.

Berger, A., & Bouwman, C.H.S. (2013). How Does Capital Affect Bank Performance During
Financial Crises? Journal of Financial Economics, 109(1), 146-176.

Berger, A., DeYoung, R., Flannery, M.J., Lee, D., & Oztekin, O. (2008). How Do Large
Banking Organizations Manage Their Capital Ratios? Journal of Financial Services Research,
34(2-3), 123-149.

Electronic copy available at: https://ssrn.com/abstract=3564267

Berger, A., Herring, R., & Szego, G. (1995). The Role of Capital in Financial Institutions.
Journal of Banking & Finance, 19(3-4), 393-430.

Blum, J. (1999). Do Capital Adequacy Requirements Reduce Risks in Banking? Journal of

Banking & Finance, 23(5), 755-771.

Bouwman, C.H.S., & Malmendier, U. (2015). Does a Bank’s History Affect Its Risk-Taking?
American Economic Review, 105(5), 321-25.

Calem, P., & Rob, R. (1999). The Impact of Capital-Based Regulation on Bank Risk-Taking.
Journal of Financial Intermediation, 8(4), 317-352.

Campello, M., Graham, J., & Harvey, C. (2010). The Real Effects of Financial Constraints:
Evidence From a Financial Crisis. Journal of Financial Economics, 97(3) 470-487

Cathcart, L., El-Jahel, L., & Jabbour, R. (2015). Can Regulators Allow Banks to Set Their Own
Capital Ratios? Journal of Banking & Finance, 53, 112-123.

Clauset, A., Shalizi, C., & Newman, M. (2009). Power-Law Distributions in Empirical
Data. SIAM Review, 51(4), 661-703.

Cohen, B. (2013). How Have Banks Adjusted to Higher Capital Requirements? BIS Quarterly
Review: September 2013, 25-41.

de Haan, L., & van den End, J.W. (2013). Banks’ Responses to Funding Liquidity Shocks:
Lending Adjustments, Liquidity Hoarding and Fire Sales. Journal of International Financial
Markets, Institutions & Money, 26, 152-174.

Diamond, D., & Rajan, R. (2000). A Theory of Bank Capital. The Journal of Finance, 55(6),
2431-2465.

Diamond, D., & Rajan, R. (2011). Fear of Fire Sales, Illiquidity Seeking, and Credit Freezes. The
Quarterly Journal of Economics, 126(2), 557-591.

Elliott, M., Golub, B., & Jackson, M. (2014). Financial Networks and Contagion. The American
Economic Review, 104(10), 3115-3153.

Fisher, R.A. (1915). Frequency Distribution of the Values of the Correlation Coefficient in
Samples from an Indefinitely Large Population. Biometrika, 10(4), 507-521.

Furlong, F., & Keeley, M. (1989). Capital Regulation and Bank Risk-Taking: A Note. Journal of
Banking & Finance, 13(6), 883-891.

Gai, P., Haldane, A., & Kapadia, S. (2011). Complexity, Concentration & Contagion. Journal of
Monetary Economics, 58, 453-470.

Electronic copy available at: https://ssrn.com/abstract=3564267

Gale, D., & Ozgur, O. (2005). Are Bank Capital Ratios Too High or Too Low? Incomplete
Markets and Optimal Capital Structure. Journal of the European Economic Association, 3(2-3),
690-700.

Gauthier, C., Lehar, A., & Souissi, M. (2012). Macroprudential Capital Requirements and
Systemic Risk. Working paper.

Glasserman, P., & Young, H.P. (2016). Contagion in Financial Networks. Journal of Economic
Literature, 54(3), 779-831.

Hart, O., & Zingales, L. (2014). Banks Are Where the Liquidity Is. Working paper.

Hugonnier, J., & Morellec, E. (2017). Bank Capital, Liquid Reserves, and Insolvency Risk.
Journal of Financial Economics, 125(2), 266-285.

Hyun, J-S., & Rhee, B-K. (2011). Bank Capital Regulation and Credit Supply. Journal of
Banking & Finance, 35(2), 323-330.

Jackson, M., & Pernoud, A. (2019). What Makes Financial Markets Special? Distorted
Investment Incentives, Regulation, and Systemic Risk Measurement. Working paper.

Newman, M. (2005). Power Laws, Pareto Distributions and Zipf’s Law. Contemporary Physics,
46(5), 323-351.

Osborne, M., Fuertes, A-M., & Milne, A. (2012). Capital and Profitability in Banking: Evidence
from U.S. Banks. Working paper.

Ramcharan, R. (2020). Bank Balance Sheets and Liquidation Values. Working Paper.

Sharpe, W.F. (1964). Capital Asset Prices: A Theory of Market Equilibrium under Conditions of
Risk. Journal of Finance. 19(3), 425-442.

Silva, W., Kimura, H., & Sobreiro, V.A. (2017). An Analysis of the Literature on Systemic
Financial Risk: A Survey. Journal of Financial Stability, 28, 91-114.

Thakor, A. (2018). Post-Crisis Regulatory Reform in Banking: Address Insolvency Risk, Not
Illiquidity! Journal of Financial Stability, 37, 107-111.

VanHoose, D. (2007). Theories of Bank Behavior Under Capital Regulation. Journal of Banking
& Finance, 31, 3680-3697.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure I. Prior-period assets of failed banks, 1986-2019.

The assets of N = 3,943 failed U.S. commercial banks, measured as of the quarter prior to failure, are collected from
the FDIC’s public database. The banks’ assets are reported as of the date of failure, and so are converted to 2019
USD using an assumed 3.5% annual inflation rate to facilitate comparison. Because the data are lognormally
distributed, the frequency distribution is presented using the base-10 logarithms of the inflation-adjusted asset value.
The sample mean is 108.18 = 151.4 million inflation-adjusted dollars.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure II. Estimated losses from bank failures, 1986-2019.

The estimated losses resulting from N = 2,573 U.S. commercial bank failures, measured as of the difference between
the value of the failed banks’ deposits reimbursed by the FDIC and the asset value recovered by the FDIC via
liquidation of the failed bank, are collected from the FDIC’s public database. The banks’ estimated losses are
reported as of the date of failure, and so are converted to 2019 USD using an assumed 3.5% annual inflation rate to
facilitate comparison. Because the data are lognormally distributed, the frequency distribution is presented using the
base-10 logarithms of the inflation-adjusted asset value. The sample mean is 107.47 = 29.5 million inflation-
adjusted dollars.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure III. Log of bank failure count and log of count frequency, 1934-2019.
Panel A. Log-log plots.

The plots show the log number of bank failures in a given period against the log of the frequency of that failure count in the FDIC’s public bank failure database.
Annual, quarterly, and monthly frequencies are reported in order from left to right. The best-fit line is given by Pr(𝑥) ∝ 𝑥 −𝛾 over the domain 𝑥 ≥ 𝑥𝑚𝑖𝑛 , where
the parameter estimate 𝛾̂ for values greater than or equal to 𝑥𝑚𝑖𝑛 are estimated via maximum likelihood, using the method of Clauset et al. (2009). Taken
together, the plots demonstrate “scale invariance” (i.e., similar power-law distributions at varying measurement frequencies) characteristic of the output of
complex systems.

Panel B. Distribution parameter estimates.

Estimation frequency Gamma 𝒙𝒎𝒊𝒏 K-S p-value (power-law)

Yearly 1.52 3 0.2870
Quarterly 3.25 52 0.4950
Monthly 3.37 21 0.4280

The Kolmogorov-Smirnov (KS) test statistic describes the likelihood of rejecting the null hypothesis that an observed data sample was drawn from some
candidate population distribution. A p-value of 0.1 or smaller would suggest rejecting the null and concluding that the sample was not drawn from the candidate
distribution. In this case, the K-S p-values for all estimation frequencies are well in excess of traditional critical values (0.01, 0.05, 0.1) when estimated using a
power-law distribution parameterized with identical values of 𝛾 and 𝑥𝑚𝑖𝑛 , suggesting that they fit a power-law fairly closely.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure IV. Reported Tier 1 capital ratios, largest U.S. commercial banks, 2010-2019.

This figure shows the quarterly Tier 1 capital ratios for solvent institutions among the 25 largest commercial banks
in the U.S., collected from Fed reports from 2010 to 2019 (N = 897). Because the Fed’s reporting requirements for
Tier 1 capital ratios vary slightly during the sample period, I report the largest of the available numbers in the event
of a conflict between multiple reports for the same bank and quarter.

Electronic copy available at: https://ssrn.com/abstract=3564267

Table V. Simulation parameters.
Parameter Description Simulation values
𝑛 Number of banks 𝑖 = 1 … 𝑛 25
𝑚 Number of risky assets 𝑘 = 1 … 𝑚 25
𝑧 Number of outside liabilities 𝑙 = 1 … 𝑧 25
𝑄𝐴 Matrix of risky asset holdings (𝑛 x 𝑚) 25 x 25 identity matrix
𝑄0 Vector of risk-free asset holdings (𝑛 x 1) 25 x 1 zero vector
𝑄𝐿 Matrix of outside liability holdings (𝑛 x 𝑚) 25 x 25 identity matrix
𝑝(𝑡) Vector of risky asset prices at 𝑡 (𝑚 x 1) 𝑝𝑘 (0) = 1⁡∀⁡𝑘 = 1 … 𝑚
𝑝0 (𝑡) Risk-free asset price at 𝑡 𝑝0 (𝑡) = 1⁡∀⁡𝑡 = 1 … . 𝑇
𝑑(𝑡) Vector of outside liability values at 𝑡 (𝑧 x 1) 𝑑𝑙 (𝑡) = 0.75⁡∀⁡𝑙 = 1 … 𝑧
𝐷𝐴 Matrix of interbank debt holdings ∑𝑖 𝐷𝑖𝑗𝐴 = 0.5⁡∀⁡𝑖
′
𝐷𝐿 Matrix of interbank debt liabilities 𝐷𝐿 = 𝐷 𝐴
𝛿 Average degree of interbank debt network 1 ≤ 𝛿 ≤ (𝑛 − 1)
𝑤0 Risk weight attached to risk-free asset 0
𝑤𝐴 Risk weight attached to representative risky asset 1
𝑤𝑥 Risk weight attached to interbank debt 1
𝑘 Upper threshold ratio for bank capitalization 0 to 0.15
𝛽 Fire sale discount factor 0.9 ≤ 𝛽 ≤ 1
𝑘 Lower threshold ratio for bank capitalization 0
𝑇 Number of months in simulation 12
𝜆 Leverage parameter for induced risk-taking 1 ≤ 𝜆 ≤ 1.2
⁡𝑟 Risk free rate 0

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure VI. Representative balance sheet.
Panel A. Initial simulation values (𝑡 = 0).

Assets Liabilities + Equity

Risk-Free (Cash) 𝑄𝑖𝑡0 𝑝0 (𝑡) = 0.00 Outside Debts 𝐿
∑𝑙 𝑄𝑖𝑙𝑡 𝑑𝑙 (𝑡) = 0.75
Risky Asset 𝐴
∑𝑘 𝑄𝑖𝑘𝑡 𝑝𝑘 (𝑡) = 1.00 Interbank Debt ∑𝑗 𝐷𝑖𝑗𝐿 (𝑡) = 0.50
Interbank Debt ∑𝑗 𝐷𝑖𝑗𝐴 (𝑡) = 0.50 Total Liabilities 𝑉𝑖𝐿 (𝑡) = 1.25
Total Assets 𝑉𝑖𝐴 (𝑡) = 1.50 Total Equity 𝑉𝑖𝐸 (𝑡) = 0.25

Banks begin with an average of 1.5 units of assets, comprised of 0 units of the risk-free asset (cash), 1 unit of the
representative risky asset, and 0.5 units of debt obligations owed by other banks. Each bank also begins with 0.75
units of outside debts (deposits and debts to other creditors), and an average of 0.5 units of debt obligations owed to
other banks. The average starting equity value for a bank in the simulation is therefore 0.25 units.

Panel B. After time-t fire sale of risky asset(s).

Assets Liabilities + Equity

Risk-Free (Cash) 𝑄𝑖𝑡0 𝑝0 (𝑡) Outside Debts 𝐿
∑𝑙 𝑄𝑖𝑙𝑡 𝑑𝑙 (𝑡)
⁡+[𝜷(𝟏 − 𝜶) ∑𝒌 𝑸𝑨𝒊𝒌𝒕 𝒑𝒌 (𝒕)] 𝒑𝟎 (𝒕)
Risky Asset 𝐴
∑𝑘 𝑄𝑖𝑘𝑡 𝑝𝑘 (𝑡) Interbank Debt ∑𝑗 𝐷𝑖𝑗𝐿 (𝑡)
−(𝟏 − 𝜶) ∑𝒌 𝑸𝑨𝒊𝒌𝒕 𝒑𝒌 (𝒕)
Interbank Debt ∑𝑗 𝐷𝑖𝑗𝐴 (𝑡) Total Liabilities 𝑉𝑖𝐿 (𝑡)
′
Total Assets 𝑉𝑖𝐴 (𝑡) Total Equity 𝑉𝑖𝐸 (𝑡)

The bank chooses the minimum fraction (1 − 𝛼), 0 ≤ 𝛼 ≤ 1, of its risky asset portfolio 𝑄𝑖𝐴 to sell to other banks so
that it comes back into compliance with the capital adequacy requirement 𝑘𝑖 (𝑡) ≥ 𝑘̅ . The sale occurs at a “fire sale”
discount 𝛽, 0 ≤ 𝛽 ≤ 1. Liabilities are unaffected, so that a full-price sale (𝛽 = 1) has no immediate effect on the
book value of the bank’s assets or equity (the loss in risky asset value is replaced dollar-for-dollar in cash), while
any fire sale with 𝛽 < 1 will result in a reduction in the risk-weighted value of the bank’s assets and a smaller
reduction in the book values of the bank’s assets and equity. If there is no 𝛼 in 0 ≤ 𝛼 ≤ 1 that can satisfy 𝑘𝑖 (𝑡) ≥ 𝑘̅ ,
the bank is considered “critically undercapitalized” and is liquidated according to the “orderly liquidation” regime
for insolvent banks, ceasing to continue as a going concern.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure VII. Simulated failure distributions.

The number of simulated failures per year is plotted against the observed frequency of that number of failures per
year, again on a log scale, collected from 10,000 iterations of the simulation model (100 loops of a simulated
distribution from 100 years of bank data). Parameter constellations are as follows: 𝛿̅ = 0, 𝑘̅ = 0 (top left); 𝛿̅ = 12,
𝑘̅ = 0 (top right); 𝛿̅ = 0, 𝑘̅ = 0.075 (bottom left); 𝛿̅ = 12, 𝑘̅ = 0.075 (bottom right). All simulations use a fire sale
discount 𝛽 = 0.95 where applicable.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure VIII. Simulated failure distributions, log-log plots.
Panel A. Log-log plots.

Panel B. Parameter estimates.

𝛿̅ = 0, 𝑘̅ = 0 𝛿̅ = 12, 𝑘̅ = 0 𝛿̅ = 0, 𝑘̅ = 0.075 𝛿̅ = 12, 𝑘̅ = 0.075

Intercept 4.5210 4.3484 4.5088 4.0703
(0.0318) (0.0369) (0.0327) (0.0377)
Slope -2.9065 -2.0097 -2.8533 -1.4767
(0.0427) (0.0346) (0.0420) (0.0277)

Rsq 94.2% 88.2% 93.9% 81.0%

N 285 452 301 667

The number of simulated failures per year is plotted against the observed frequency of that number of failures per
year, again on a log scale, collected from 10,000 iterations of the simulation model (100 loops of a simulated
distribution from 100 years of bank data). Parameter constellations are as follows: 𝛿̅ = 0, 𝑘̅ = 0 (top left); 𝛿̅ = 12,
𝑘̅ = 0 (top right); 𝛿̅ = 0, 𝑘̅ = 0.075 (bottom left); 𝛿̅ = 12, 𝑘̅ = 0.075 (bottom right). All simulations use a fire sale
discount 𝛽 = 0.95 where applicable. The best-fit lines in Panel A correspond to Pr(𝑥) = 𝐶𝑥 𝛾̂ , estimated via
ordinary least-squares using log transformed variables: ln(Pr(𝑥)) = ln(𝐶 ) + 𝛾̂ ∗ ln⁡(𝑥).

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure IX. Bank failures and interbank debt network structure.

The realized bank failure rates from the simulation model are reported as a function of the average degree of the
interbank debt network, which takes values of 𝛿̅ = 1 (very low diversification in interbank debt obligations) to 𝛿̅ =
24 (very high diversification in interbank debt obligations). In all simulations, capital regulations are given by 𝑘 =
𝑘̅ = 0 (i.e., so long as a bank is solvent, it can maintain any level of capitalization without facing the requirement for
prompt corrective action). The fire-sale discount parameter 𝛽 is therefore irrelevant. The average mean and volatility
of banks’ representative risky assets are 𝜇̅𝐴 = 3.5% and 𝜎̅𝐴 = 8.0% in all simulations. Each panel reports the
simulated bank failure rate for each value of 𝛿̅; a rate of 0.2 indicates that out of 1,000 simulations for 25 banks,
5,000/(1,000*25) = 20% of banks failed. The failure rates produced by the simulation model are highly consistent
with those resulting from a single-period version of the model (Elliott, Jackson, and Golub (2014)).

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure X. Bank failures in isolation.

The plot shows the results of 1,000 iterations of the simulation model with a solvency condition given by 𝑘 = 0, and
prompt corrective action threshold values of 𝑘̅ = {0,0.015, 0.03, 0.045, 0.06, 0.075, 0.09, 0.105, 0.12, 0.135, 0.15}..
The average mean and volatility of banks’ representative risky assets are 𝜇̅𝐴 = 3.5% and 𝜎̅𝐴 = 8.0% in all
simulations. In these simulations, there is no interbank debt network, so the average degree of the interbank debt
network 𝛿̅ is irrelevant. In the event of a bank’s insolvency, its deposit liabilities and shares in its proprietary asset
are equally redistributed among half of the remaining banks at random. OLS best-fit lines connect results for
simulations using values of the fire-sale discount parameter 𝛽 = {0.9, 0.95, 1.00}. Each point in the plot reports the
simulated bank failure rate for each combination of parameter inputs 𝜆 and 𝛿̅; a rate of 0.2 indicates that out of 1,000
simulations for 25 banks, 5,000/(1,000*25) = 20% of banks failed.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure XI. Failure rates by capital threshold and fire sale discount.

All figures represent the results of simulations with a solvency condition given by 𝑘 = 0, and prompt corrective action threshold values of 𝑘̅ =
{0,0.015, 0.03, 0.045, 0.06, 0.075, 0.09, 0.105, 0.12, 0.135, 0.15}. OLS best-fit lines connect results for simulations over average interbank debt network degree
values of 𝛿̅ = {9, 12, 15, 18, 21, 24}. The average mean and volatility of banks’ representative risky assets are 𝜇̅𝐴 = 3.5% and 𝜎̅𝐴 = 8.0% in all simulations.
Each panel reports the simulated bank failure rate for each combination of parameter inputs 𝑘̅ and 𝛿̅; a rate of 0.2 indicates that out of 1,000 simulations for 25
banks, 5,000/(1,000*25) = 20% of banks failed. From left to right, the panels correspond to a fire-sale discount parameter 𝛽 equal to 0.9 (assets are sold at 90
percent of book value), 0.95 (assets are sold at 95 percent of book value), and 1 (no fire-sale discount), respectively.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure XII. Failure rates by capital threshold and average of banks’ mean risky asset returns.

All figures represent the results of simulations with a solvency condition given by 𝑘 = 0, and prompt corrective action threshold values of 𝑘̅ =
{0,0.015, 0.03, 0.045, 0.06, 0.075, 0.09, 0.105, 0.12, 0.135, 0.15}. OLS best-fit lines connect results for simulations over average interbank debt network degree
values of 𝛿̅ = {9, 12, 15, 18, 21, 24}. The fire-sale discount parameter 𝛽 is equal to 0.95 (assets are sold at 95 percent of book value) in all simulations. From left
to right, the panels correspond to a distribution of banks’ representative risky assets means with an average of 𝜇̅𝐴 = 1.5% (high likelihood of negative asset
returns), 𝜇̅𝐴 = 3.5% (base case), and 𝜇̅𝐴 = 5.5% (low likelihood of negative asset returns), respectively. The average volatility of banks’ representative risky
asset returns is 𝜎̅𝐴 = 8.0% in all simulations. Each panel reports the simulated bank failure rate for each combination of parameter inputs 𝑘̅ and 𝛿̅; a rate of 0.2
indicates that out of 1,000 simulations for 25 banks, 5,000/(1,000*25) = 20% of banks failed.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure XIII. Failure rates by capital threshold and average of banks’ risky asset volatilities.

All figures represent the results of simulations with a solvency condition given by 𝑘 = 0, and prompt corrective action threshold values of 𝑘̅ =
{0,0.015, 0.03, 0.045, 0.06, 0.075, 0.09, 0.105, 0.12, 0.135, 0.15}. OLS best-fit lines connect results for simulations over average interbank debt network degree
values of 𝛿̅ = {9, 12, 15, 18, 21, 24}. The fire-sale discount parameter 𝛽 is equal to 0.95 (assets are sold at 95 percent of book value) in all simulations. The
average of banks’ representative risky asset mean returns is 𝜇̅𝐴 = 3.5% in all simulations. From left to right, the panels correspond to a distribution of banks’
representative risky assets volatilities with an average of 𝜎̅𝐴 = 6.0% (low likelihood of negative asset returns), 𝜎̅𝐴 = 8.0% (base case), and 𝜎̅𝐴 = 10.0% (high
likelihood of negative asset returns), respectively. Each panel reports the simulated bank failure rate for each combination of parameter inputs 𝑘̅ and 𝛿̅; a rate of
0.2 indicates that out of 1,000 simulations for 25 banks, 5,000/(1,000*25) = 20% of banks failed.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure XIV. Frequency distributions of selected bank balance sheet parameters by simulated failure status.

The plots show the frequency distributions of selected balance-sheet parameters of 25 banks over 10,000 iterations of the simulation model, with 𝑘̅ = 0.075, 𝛿̅ =
12, 𝛽 = 0.95, 𝜇̅𝐴 = 3.5%, and 𝜎̅𝐴 = 8.0% in all iterations. These balance-sheet parameters are measured at the beginning of the iteration (and so are expected,
rather than realized, values). The frequency distributions are separated by the solvency status of the bank at the end of the iteration (“failed” or “survived”). The
leftmost panel shows the means of individual banks’ risky asset returns 𝜇𝐴 . The average failed bank had an expected mean risky asset return of 2.95%, compared
with an expected mean risky asset return of 3.63% for the average surviving bank. The difference between the means is significant at the 1% level using
Student’s t-test. The central panel shows the volatilities of individual banks’ risky asset returns 𝜎𝐴 . The average failed bank had an expected risky asset return
volatility of 10.55%, compared with an expected risky asset return volatility of 7.65% for the average surviving bank. The difference between the distributions is
significant at the 1% level using both a Kolmogorov-Smirnov two-sample test and a Chi-squared two-sample test. The rightmost panel shows the number of bank
counterparties with debt obligations to the individual bank (i.e., the bank’s debtors in the interbank network). The average failed bank had 10.39 debtors in the
interbank network, compared with 12.32 debtors for the average surviving bank. The difference between the means is significant at the 1% level using Student’s
t-test.

Electronic copy available at: https://ssrn.com/abstract=3564267

Figure XV. Bank failures and induced leverage.

The plot shows the results of 1,000 iterations of the simulation model over varying levels of “induced leverage”, as
measured by the free parameter 𝜆. For all iterations, capital regulations are given by 𝑘 = 𝑘̅ = 0 (i.e., so long as a
bank is solvent, it can maintain any level of capitalization without facing the requirement for prompt corrective
action). The fire-sale discount parameter 𝛽 is therefore irrelevant. The average mean and volatility of banks’
representative risky assets are 𝜇̅𝐴 = 3.5% and 𝜎̅𝐴 = 8.0% in all simulations. OLS best-fit lines connect results for
simulations over average interbank debt network degree values of 𝛿̅ = {9, 12, 15, 18, 21, 24}. Each point in the plot
reports the simulated bank failure rate for each combination of parameter inputs 𝜆 and 𝛿̅; a rate of 0.2 indicates that
out of 1,000 simulations for 25 banks, 5,000/(1,000*25) = 20% of banks failed.

Electronic copy available at: https://ssrn.com/abstract=3564267