Carrying the (Paper) Burden: A Portfolio View of Financial

Stability and Optimal Bank Size

Jaap W.B. Bos∗ Martien Lamers† Victoria Purice‡

Abstract
We investigate how big banks should be from a financial stability point of view by imagining
the supervisor of a banking system as an ‘investor’ holding a portfolio of banks. Allowing
this mean-variance supervisor to choose (i) bank size, (ii) business model or (iii) both si-
multaneously, we find that the largest banks are consistently overrepresented in the U.S.
banking system between 1984 and 2013. From a portfolio point of view, a more stable bank-
ing system appears to favor a significant reduction in the size of the largest banks, as well
as a return to a more traditional intermediation role.

JEL classification: G21; C61; E44; E63.

Keywords: Financial Stability; Modern Portfolio Theory; U.S. banking.

We thank Geert Bekaert, Bob DeYoung, Iftekhar Hasan, Michael Kötter, Lukas Menkhoff, Christophe
Moussu, Rachel Pownall, Koen Schoors, Peter Schotman, David Veredas, Rudi Vander Vennet and seminar
participants at Ghent University, Maastricht University School of Business and Economics, University of Gronin-
gen, the International Conference on Money, Banking and Finance (2012), the Annual Conference of the Royal
Economic Society (2013), the Spring Meeting of Young Economists (2013), the International Conference of the
Financial Engineering and Banking Society (2013), the European meeting of the Financial Management As-
sociation (2013), the Conference of the International Federation of Operational Research Societies (2014), the
International Workshop on Measuring Banking Performance at Loughborough University (2014), the 1st IWH-
FIN-FIRE Workshop on “Challenges to Financial Stability” (2015), and the “Conference on Contemporary Issues
in Banking” (2015) for helpful comments.

∗
Maastricht University School of Business and Economics, P.O. Box 616, 6200 MD, Maastricht, The Nether-
lands, [email protected].
†
University of Groningen, Department of Economics, Econometrics and Finance, Faculty of Economics and
Business, P.O. Box 800, NL-9700 AV Groningen, The Netherlands, [email protected]. Corresponding author.
‡
University of Groningen, Department of Economics, Econometrics and Finance, Faculty of Economics and
Business, P.O. Box 800, NL-9700 AV Groningen, The Netherlands, [email protected].
Portfolio. ORIGIN early 18th century: from Italian portafogli, from portare 'carry' + foglio

'sheet of paper' (from Latin folium).

1 Introduction

Every bank supervisory authority strives for a healthy banking system in which individual

banks thrive, fulfill their role in the economy in an adequate manner, and risks are contained.

In such a system - ideally - individual banks refrain from taking excessive risks, and the system

as a whole is diverse and robust, resulting in low systemic risk. But what exactly are the

characteristics of such a healthy banking system? How far removed are we from that system,

and what is the price we may pay for getting there?

Ever since the global financial crisis, answers to these questions have become much in demand,

and for good reason. After all, societal costs of bank failures have been far from trivial (see,

amongst others, Bernanke, 1983; Calomiris and Mason, 2003; Ashcraft, 2005; Dell’Ariccia et al.,

2008; Claessens et al., 2009; Reinhart and Rogoff, 2009). The vast majority of the academic

and policy literature has gone down the route of assessing the causes of the financial crisis.

Recognizing that individual banks may well have become so large and interconnected that they

simply cannot be allowed to fail, they welcome discussions (i) on measures of systemic risk

which incorporate the macroeconomic effects of a shock to the system, and (ii) on the optimal

supervisory setup for dealing with and minimizing the likelihood of such a shock.

Meanwhile, the lessons that we can learn from these exercises are increasingly clear. Not only

have we become much better at identifying which banks are systemically risky, but we know that

systemic risk has not come down to pre-crisis levels. Also, the ‘excessive’ size of some banks is

generally seen as a major contributor to systemic risk, followed by business model characteristics

such as leverage, interconnectedness and non-interest income (see, e.g., De Jonghe, 2010; Pais

and Stork, 2011; Acharya et al., 2012a; Acharya et al., 2012b; Bisias et al., 2012; Huang et al.,

2012; Engle et al., 2015; Weiß et al., 2014; Adrian and Brunnermeier, 2016; Black et al., 2016;

Brownlees and Engle, 2016).

In this paper, we present a modest contribution to this discussion which intends to comple-

ment the previous findings, as well as hopefully move a step further. We go beyond analyzing

the (past and present) status quo and instead try to answer some of the elusive, but important,

‘what if’ questions. Given the aforementioned importance of bank size in determining systemic

1
risk, the question we focus on is how big banks should be to maintain a healthy financial system

that is able to withstand shocks. To do so, we envision the supervisor of a banking system as

an investor holding the market portfolio of banks. We go even further, and imagine that this

investor can change the weights of the banks in this portfolio, i.e., can change the structure of

the banking system. Despite this being a rather daring and somewhat ‘hypothetical’ exercise, we

nevertheless believe it is worth it: for by taking this perspective, we can investigate the role of

large banks in determining systemic (portfolio) risk in much more detail. By viewing individual

bank size as a key element of a much broader portfolio selection problem, we can investigate

what size banks should be, both in absolute terms and relative to the rest of the banking system,

given their current risk-return trade-off.

The contributions of this paper can be easily summarized by three policy-related experiments,

each related to bank size and business models, we subject our data to. In the first experiment,

we ask whether, in a world in which the supervisor could effortlessly adjust her holdings, it

would be possible to reduce portfolio risk. And if so, what would this mean for bank size? We

find that portfolio risk can be reduced, but this requires the largest banks to have a significantly

lower weight compared to the current situation.

Recognizing that this experiment far exceeds the current regulatory capabilities, we test a

more realistic scenario in a second experiment where the supervisor is able to restrict banks’

business models, while keeping their size unchanged. Importantly, activity restrictions as sug-

gested by Paul Volcker, the Liikanen report, and the Vickers report have been put forward based

on the notion that restricting banks’ business models will significantly alter their contribution

to systemic risk. However, given the current size of banks, we find no real improvements in the

performance of the supervisor’s portfolio when she can only change business models.

In a third experiment, we take into account that banks’ business models and profitability

may not be independent of their size, and ask what would happen to the characteristics of

the supervisor’s portfolio if she could choose bank size and business model simultaneously. We

find that even in this final, more realistic, policy experiment, bank size remains the key driver

of portfolio risk: it is the business model favored by the few very large banks that increases

portfolio risk. Depending on the level of risk aversion of the supervisor, concentration levels (as

measured by the market share of the largest 5 percent of banks) should be 40 to 60 percentage

points lower than is currently the case.

In answering these questions, we incorporate the trade-off between diversity and diversifica-

2
tion (Rajan, 2005; Wagner, 2008, 2010; Allen et al., 2012). While portfolio diversification can

be beneficial for individual banks, if all banks diversify in the same manner the system as a

whole becomes more susceptible to common shocks. So far only a few empirical papers have

investigated this trade-off. Most relevant to our paper is Liu et al. (2015), who find an important

role for the diversity (uniqueness) of banks in mitigating systemic risk. In our setting, this lack

of diversity is captured by a high correlation between individual banks, which drives up the

aggregate risk of the supervisor.

Moreover, we also implicitly address the Too-Many-To-Fail problem in which banks have an

incentive to herd, increasing the likelihood that all will fail together and therefore leading to a

bailout of the entire system (Acharya and Yorulmazer, 2007; Claeys and Schoors, 2007; Brown

and Dinç, 2011; Ibragimov et al., 2011). Acharya and Yorulmazer (2007) show theoretically that

this primarily affects smaller banks. In our setting, if many small banks are exposed to the same

risk factor, the increased correlation would make them act as if they are one large bank and

therefore will drive up the risk in the aggregate supervisory portfolio.

While we recognize that our results are of a partial equilibrium nature, they are consistent

with the existing literature on the role of bank size and financial stability. However, contrary

to previous papers, we go beyond the status quo and as such provide important insights into

the optimal design of a banking system. Reducing the size of the largest banks may not only

make individual banks safer and easier to fail or rescue, but can also contribute to a reduction

in the riskiness of the system as a whole. Moreover, activity restrictions, as suggested by Paul

Volcker, the Liikanen report, and the Vickers report, may not be enough to reduce systemic risk

to a sufficiently low level. These results lend support to a renewed focus in regulatory efforts

to reduce the size of the largest banks, either directly via caps on size, or indirectly via, e.g.,

increased capital requirements on the largest banks encouraging them to downsize.

This paper proceeds as follows. In Section 2 we discuss our policy experiments. In Section 3

we introduce our data, followed by our results in Section 4, and Section 5 where we test the

robustness of the results. Finally, we conclude in Section 6.

3
2 Policy Experiments

2.1 A supervisor’s portfolio

During the crisis, both equity holders and (small) debt holders of banks experienced the

consequences of shocks to the banking system, as a result of shocks to individual banks in that

system. Not surprisingly, therefore, Rajan (2005), Wagner (2008), Acharya (2009), Ibragimov

et al. (2011) and Allen et al. (2012), all draw on Modern Portfolio Theory (Markowitz, 1952,

1959) to study the health of the banking system.

We extend this line of thinking by applying elements of portfolio theory to the banking

system in order to investigate the role of large banks in determining systemic (portfolio) risk.

Accordingly, we envision a supervisor managing a portfolio of banks, with the following return

and risk of that portfolio:

rp = w0 R (1)

σp2 = w0 Σw (2)

where w is a column vector representing the weight of each bank in the portfolio and R represents

the return of each bank. Furthermore, Σ represents the covariance matrix of these returns. The

average return of the portfolio is given by rp , while the variance of this set of returns is given

by σp2 and represents the measure of portfolio risk.

The supervisor is faced with a trade-off between risk and return in her portfolio. On the one

hand, she has a mandate to maintain a stable financial system and would therefore like to keep

aggregate risk at a sufficiently low level. On the other hand, she is also interested in achieving a

certain level of return, as high charter values may boost the stability of individual banks in the

long run. To capture this trade-off, we assume that the supervisor has the following quadratic

utility function, which is common in mean-variance frameworks:

λ 2
U (rp ) = rp − r (3)
2 p

where λ is a strictly positive risk aversion coefficient, and marginal utility is increasing in rp
1
for rp < λ. The supervisor maximizes expected utility E [U (rp )], which can be conveniently

4
 expressed as a function of the portfolio return and variance:

 
λ 2
E [U (rp )] = E rp − rp
2
λ 2
σ + µ2


=µ− (4)
2

where µ is the expected return of the portfolio and σ 2 is the variance, which are both obtained

from their sample equivalents.

Since the last crisis has shown that the costs of an unstable banking system are often borne

not just by equity holders, but by various other claimants and stakeholders, we choose the return

on assets as our (broad) measure of return. After all, in the event that the supervisor has to

bail out a bank, saving or guaranteeing its liabilities will be equivalent to saving or guaranteeing

its assets.1 The weight of a bank in the supervisor’s portfolio is its market share based on the

value of total assets.

To measure those assets, as well as their returns, we use both market and book value data,

depending on the policy experiment at hand. Obviously, market prices are forward looking and

available at a much higher frequency than the backward looking book value returns. However, as

shown by Allen and Carletti (2008), in financial crises market prices tend to reflect the amount

of available liquidity instead of future earnings. Moreover, market prices are only available for

a small subset of relatively large, listed banks and we know that small banks are potentially a

source of (liquidity) contagion through the interbank market (see e.g. Furfine, 1999, 2003; van

Lelyveld and Liedorp, 2006; Degryse and Nguyen, 2007).

Using this framework, we investigate three policy experiments regarding the manner in which

the supervisor could achieve her utility maximizing portfolio (UMP).

2.2 Experiment 1: the role of bank size in reducing portfolio risk

In the first experiment, we ask whether, in a world in which the supervisor could effortlessly

adjust her holdings, it would be possible to reduce portfolio risk. And if so, what would this

mean for bank size? To answer these questions, we maximize the supervisor’s objective function

given a certain level of risk aversion, and adjust the weights of the banks in order to compare the

differences in design between the actual portfolio and the hypothetical UMP, with the addition

1
Moreover, return on assets is a cleaner measure of the underlying profitability than return on equity, since the
latter also reflects management choices with regards to leverage.

5
 of several further constraints. First, the supervisor cannot short a bank, so no bank can have a

negative weight. Second, the weights of banks have to add up to one, as the existing assets are

merely reshuffled, without any of them being created or destroyed. Finally, a supervisor is also

assumed to choose the weights such that the portfolio does not have negative expected returns.

The portfolio optimization of the supervisor is therefore:

argmax E [U (rp )] (5)

w

s.t. w≥0

10 w = 1

w0 µ ≥ 0

which we can solve for different levels of risk aversion, as graphically depicted in Figure 1. In

this Figure we can see tangency portfolios based on the utility function, with low levels of risk

aversion favoring higher returns and higher risk, and high levels of risk aversion leading to a

portfolio choice close to the minimum variance portfolio.

[Insert Figure 1 near here]

When discussing the results, we investigate how the current portfolio risk and return compare

to that of the UMP, and how these relate to the risk aversion of the supervisor. Most importantly,

we compare how the two portfolios differ in the way the supervisor allocates her assets. However,

rather than analyzing individual bank sizes or weights, we look at the share of the largest five

percent of banks as an intuitive and simple measure of concentration which allows us to compare

the size of banks between the different portfolios, as well as over time.2 Finally, we ask whether

the largest banks in the actual portfolio have retained their relative importance by comparing

their initial market share with the weight they receive in the UMP. Results to this experiment

will tell us whether there is a role for large banks in a stable financial system, how large they

can be, and whether these are the same large banks as we currently see.

2
An alternative measure of concentration would be the Herfindahl-Hirschman Index, although using portfolio
weights of the largest banks is more intuitive in this setting. The five percent concentration measure is preferred
since it allows for a better comparison in different-sized banking systems (see Alegria and Schaeck, 2009), making
it easier to compare the results over time.

6
 2.3 Experiment 2: the role of the business model

We recognize that the first experiment, where the supervisor has the ability to effortlessly

reweight her portfolio, goes beyond the mechanisms currently in place.3 What regulators can

do, and have done in the past, is to impose restrictions on the type and mix of activities that

banks are allowed to undertake. Therefore, in a second experiment we determine what business

model the supervisor would ideally assign to each bank in order to reach her optimal portfolio,

while keeping banks at their current size. To do so, we change the balance sheet composition of

each bank as well as its income statement. More specifically, we define a bank’s α as the ratio
Loans
of loans to total earning assets, α = Total Assets - Fixed Assets . Naturally, (1 − α) is then the ratio

of other earning assets to total earning assets. Banks with a high α can be classified as more

traditional banks, whereas banks with a lower α are more akin to universal/investment banks.

We assume that the net interest income of a bank is generated fully by its loans (α), while

the non-interest income can be entirely attributed to other earning assets (1 − α). We can then

express a bank’s return as a function of α, the net interest income and non-interest income:

Net interest incomei Non-interest incomei

ri = αi + (1 − αi ) (6)
Loansi Non-loan assetsi

while keeping the actual implied tax rate, profits from minority interests and other expenses

unchanged. We can subsequently obtain the UMP by maximizing the expected utility of the

supervisor:

argmax E [U (rp )] (7)

α

s.t. α ∈ [0, 1]

with the constraint that α can only take values between zero and one. Changing α affects both

the returns of the banks as well as their correlation with other banks, leading to potentially

improved diversification benefits for the supervisor. As in the previous experiment, we will

compare the risk and return of the UMP with the actual portfolio. In addition, of course, we

are interested in the business models that are preferred in the UMP. The goal of this experiment

is to find out whether we can have a stable banking system with banks as large as they are

currently, and, if so, which business models they should have.

3
Currently, a supervisor can only reweight her portfolio through resolution, (hidden) bailouts, approving mergers
and acquisitions, or forcing banks to sell off assets to maintain a competitive environment.

7
2.4 Experiment 3: size and the business model

In the previous two experiments we evaluate the impact of changing banks’ size and their

business model separately. However, both characteristics of the banks in a supervisor’s portfolio

are likely to be correlated, as larger banks often have a different mix of activities compared with

smaller banks. Therefore, in the third experiment we investigate a more realistic setup in which

bank size and business model change simultaneously. To do so, we first estimate the relationship

between balance sheet composition and bank size, and then use our findings to adjust a bank’s

business model every time its size - through a change of its weight in the portfolio - is changed.

For example, when downsizing a large (universal) bank, not only will its balance sheet shrink,

but also the composition of its balance sheet will resemble that of a more traditional bank.

In this way, the UMP is both more realistic and takes into account that size and business

model are likely related. We again compare the risk and return of this UMP with the actual

portfolio, as well as the distribution of their weights and business models, with the intention of

analyzing whether there is a role large banks in a stable financial system when business models

are endogenous with respect to bank size.

While this last experiment adds more realism to the previous experiments, we recognize that

even in this setting we are analyzing a partial equilibrium outcome. More specifically, we only

analyze the size and composition of banks’ balance sheets, and disregard issues ranging from

changes in customer demand for bank products and services, to locational decisions of the banks

themselves. As a consequence, the results based on these experiments should not be taken at

face value, but rather as ‘crude’ evidence and a first step towards understanding how big banks

should optimally be.

2.5 Assessing the assumptions behind our experiments

In addition to each of our three experiments, we perform two analyses aimed at further

enhancing our understanding of the relationship between bank size and the systemic risk in the

banking system.

Our first additional analysis addresses the fact that a supervisor is most likely concerned

with downside risk rather than ‘symmetric’ risk (see, e.g., De Jonghe, 2010; Acharya et al.,

2012a; Acharya et al., 2012b; Huang et al., 2012; Engle et al., 2015; Adrian and Brunnermeier,

2016; Black et al., 2016; Brownlees and Engle, 2016). So far, we have followed the standard

portfolio theory assumptions and assumed that the return distribution is fully characterized by

8
the first two moments, disregarding any skewness, kurtosis or tail dependence. In a robustness

test we use an alternative measure of risk, namely the semi-variance. In doing so, we focus on

the variance of below-average returns and therefore zoom in on downside risk. A drawback of

this approach is that semi-variance is not compatible with maximizing the expected utility of

the supervisor, which is why we present it as a robustness test.

Our second additional analysis addresses concerns that a supervisor of a banking system is

limited (by transaction costs, regulation, or otherwise) in her ability to change the weight of a

bank in her portfolio. So far, we have been relatively agnostic regarding these concerns, but in

another robustness test we analyze the characteristics of the UMP under scenarios of limited

reweighting.

3 Data

3.1 Sources

We use bank data from multiple sources. We obtain balance sheet and income statement

data on a quarterly basis from the Call Reports for Income and Condition provided by the

Federal Reserve System. Moreover, we use the CRSP-FRB link provided by the Federal Reserve

Bank of New York (2014) to obtain stock market data from CRSP for listed banks.

We retrieve balance sheet and income statement data from the FFIEC 031/041 forms for

the period 1984Q1 and 2013Q4. We only use data on commercial banks, excluding savings,

cooperative and industrial banks, as well as non-deposit trust companies. Banks that are owned

by a holding company are consolidated into their highest owner using the ownership information

provided in the Call Reports. Banks that are not affiliated with a holding company are treated

as independent. For convenience, we classify both independent commercial banks and bank

holding companies as banks. We use a GDP deflator to express all dollar amounts in 2013Q1

dollars.

We clean the data by applying a number of selection criteria. First, we drop banks with

negative or missing equity, and missing total assets or net income. Next, we follow DeYoung

and Hasan (1998) and remove DeNovo banks that have been in existence less than 9 years.

Moreover, to make the analysis more tractable, we drop small banks with less than $100 million

in total assets. Finally, we truncate the following variables at the 1st and 99th percentile: return

on assets, equity/total assets, non-interest income/total income, net interest income/total loans,

9
and non-interest income/non-loan assets. This leaves us with 333,035 bank-quarter observations

for 9,392 unique banks. Summary statistics for selected variables are shown in Panel A of Table 1.

We can see that the return on assets and return on equity are relatively normally distributed,

although slightly skewed to the left because of episodes of financial distress. The weight based

on the total assets of each bank has a much more skewed distribution, with the largest bank in

the sample having a market share of 21 percent, corresponding to total assets of $1.8 trillion,

while the average bank has a market share of 0.03 percent or total assets of $1.8 billion. Finally,

we note that there is considerable variation in terms of business model. The average share of

loans in non-fixed assets, α, is 63 percent, with some banks showing a traditional model with

an α of 98 percent, and others are more akin to investments banks with an α below 10 percent.

For listed banks we obtain market capitalization data using their PERMCO identifier from

the CRSP-FRB link. The sample runs from 1990 to the end of 2013 as the link is only available

from the beginning of the 1990s onward. We calculate the market value of assets as the sum of the

market capitalization and the book value of a bank’s liabilities, which we obtain separately from

the FR Y-9C forms. The market value return on assets is then defined as the daily percentage

change in the market value of assets, and the weights are based on the asset values as well. To

ensure that we include actively traded banks, only banks with a non-zero return on 80 percent

of the trading days are considered. We adjust the data for mergers and acquisitions by creating

a new entity when this type of event is reported in CRSP. This leaves us with 716 unique banks

over 1,786,587 daily observations. Summary statistics are shown in Panel B of Table 1. The

returns seem relatively normally distributed and centered around zero, with a mean daily return

of 0.01 percent. Similar to the book value data, there is a large inequality in terms of size. While

the average bank has a weight of 0.33 percent, the largest bank has a weight of 23 percent and

a market value of assets of $2.5 trillion.

[Insert Table 1 near here]

3.2 Window size and frequency

Because of the high frequency of market data, we perform this analysis on a yearly basis

using daily observations. This means that for every available year we use 252 trading days to

estimate the sample expected return and the sample covariance matrix. Banks with missing

data are discarded for the year, as are banks that are not actively traded on at least 80 percent

of the trading days. In any given year in the sample, this leaves us with between 169 and 393

10
 listed banks.

As book value data are more sparse, we choose to estimate the sample expected return and

the sample covariance matrix with a rolling window approach. Our choice of window length

is based on the way that the Z -score risk measure is usually calculated, with studies using

between 3 and 6 yearly observations for estimating the standard deviation of the return on

assets (see, e.g., Laeven and Levine, 2009; Koetter et al., 2012; Beck et al., 2013). Since we have

quarterly data, we increase the length to 8 observations to lower the noise in the estimation of

the covariance matrix, while still allowing for sufficient time variation. As a consequence, banks

need to post data in each consecutive quarter of the window to be included in the analysis. This

means that in any given quarter in the sample between 2,651 and 3,153 banks are available to

run the analysis on.4

All in all, we can generate results based on market data for each of the 24 available years,

while we estimate the results based on book value data for 113 out of 120 available quarters

since we lose the first seven quarters to build the rolling windows.

3.3 Risk aversion

A final issue to address before we can discuss our results concerns the choice of the risk

aversion coefficient λ. On the one hand, λ is theoretically constrained by the properties of the

quadratic utility function. We need to choose λ such that marginal utility is still increasing in

the return, implying values of λ between 0 and 1r . Given the sample averages of r from Table 1,

the values of λ should be below 384 for book value data, and below 10,000 for market data.

On the other hand, λ should also strike a balance between risk and return. We turn to the

Sharpe ratio σr to investigate this balance. Using the sample averages from Table 1, the book

value Sharpe ratio is 2.89 and the market value Sharpe ratio is 0.30. These values are of greater

magnitude than, for instance, the S&P500 over the same period of time. As the monthly S&P500

Sharpe ratio was 0.15, the market data Sharpe ratio was two times higher and the book value

Sharpe ratio even twenty times higher! For the book value returns, this can be caused by, e.g., a

difference in data frequency (quarterly versus monthly), or by earnings smoothing which lowers

4
Note that we therefore have a covariance matrix estimated in a small T large N setting, which can give rise
to unstable solutions for small changes in returns. To check the robustness of this covariance matrix, we ran
several analyses. First, we used the methodology developed by Ledoit and Wolf (2003, 2004) to minimize errors
in estimating covariance matrices in this setting. Second, we ran Monte Carlo simulations with different starting
weights to check if these influence the stability of the solution. In both cases the results did not change materially,
and we therefore exclude them for reason of brevity. The results are available on request.

11
 the variation in returns. In the case of market returns, potential Too-Big-To-Fail considerations

can play a role (Gandhi and Lustig, 2015; Kelly et al., 2016). Faced with these wildly different

Sharpe ratios, we choose to adjust conventional risk aversion coefficients to balance risk and

return for the assets under consideration. For more details on why this matters and how it

impacts the choice of risk aversion coefficient, we refer to the Appendix.

The usual risk aversion levels lie between 1 and 10, where 1 is a low risk aversion case and

10 indicates high risk aversion. Given the Sharpe ratio of the benchmark S&P500 of 0.15, and

our market data Sharpe ratio of 0.30, a risk aversion coefficient normally associated with high

risk aversion of 10 translates into an equivalent value of around 40. Similarly, for the book value

Sharpe ratio of 2.89, this high risk aversion coefficient of 10 translates into an equivalent value

of roughly 400. Given these properties as well as the positive marginal utility constraint, we

choose values for λ of 10, 50 and 100 for market data, and 100, 200 and 300 for book value data,

representing low, medium and high levels of risk aversion, respectively.5

In the next section we present the results of our analysis.

4 Results

What does the optimal portfolio look like in each of our three policy experiments? Recall

from Section 2 that we first obtain this hypothetical, utility maximizing portfolio (UMP) as-

suming that we can change the weights that banks have without changing their business models.

Next, we follow the opposite approach and change the business models while keeping weights

unchanged. Finally, we combine the two approaches by allowing the business model to be de-

pendent on bank size. For our first experiment, we can obtain the UMP for both market and

book value data. For the second and third experiment, we can only use book value data, since

these experiments require us to manipulate balance sheets and income statements of the banks.

4.1 What role does bank size play in the risk/return trade-off ?

Our main interest in the first policy experiment is whether a supervisor can reduce the risk

of her portfolio of banks if she is able to effortlessly adjust her holdings. Of course, we are not

only interested in the overall portfolio risk, but in particular also want to know what the move

5
We have tested several values of the risk aversion coefficient for the first experiment, and the three values chosen
here seemed to best represent the low, medium and high risk aversion cases. More elaborate results are available
upon request.

12
from the actual portfolio (FED portfolio) to the UMP implies for bank size.

We present the results in Figures 2 and 3 for book value data and market value data,

respectively. Both figures contain three panels, corresponding to low (panel a), medium (panel

b) and high (panel c) risk aversion levels. In the top row, we show the portfolio risk of both

the FED portfolio and the respective UMP. Similarly, the returns for both portfolios are shown

in the middle row, while the bottom row shows the concentration level in the two portfolios as

well as the weight of the largest 5 percent of banks from the FED portfolio in the UMP.

We first discuss the book value results before proceeding with the market data results. As

can be seen in the top row of Figure 2, the risk in the FED portfolio spikes during the Savings

and Loans crisis as well as during the recent crisis, suggesting that our measure of risk, despite it

being symmetric, is able to pick up episodes of systemic distress. During the recent crisis returns

also drop, but overall they are relatively stable at around 0.2 percent. In the UMP, as expected,

we see that portfolio risk and return decrease for increasing levels of risk aversion. Even in the

case where the supervisor has a high degree of risk aversion, risk is not completely eliminated

and riskier episodes are still present. However, we do note that these riskier episodes are nowhere

near as severe as in the FED portfolio, indicating that (systemic) risk is significantly lower in the

UMP. Combined with the observation that, even in the high risk aversion case, returns of the

UMP are generally higher than those of the FED portfolio, we conclude that the FED portfolio

is not on the efficient frontier.

Regarding the composition of the two portfolios, we observe that the concentration in the

FED portfolio gradually increases over time, starting at around 64 percent and reaching 88

percent at its peak. Meanwhile, for a low degree of risk aversion the UMP concentration is on

average similar in magnitude to that of the FED portfolio, albeit more irregular. As risk aversion

increases, however, the average concentration levels drop significantly. For instance, in the high

risk aversion case, concentration fluctuates between 5 and 35 percent, with peaks coinciding with

periods of financial distress. These peaks are possibly the result of an increase in correlation of

(negative) returns, leading to less diversification possibilities for the supervisor and a resulting

spike in concentration. More interestingly, the banks that are currently the largest banks of the

FED portfolio see their cumulative weight reduced to at most 5 percent in the UMP, regardless

of the degree of risk aversion of the supervisor.

In terms of optimal portfolio design, it seems that these large banks are consistently over-

weighted given their risk-return trade-off and that, for medium to high levels of risk aversion,

13
 the UMP does not allow for high levels of concentration to begin with. Note however that

lower concentration does not necessarily imply that there are no large banks in absolute dollar

amounts. The largest banks in the FED portfolio, Bank of America and JPMorgan Chase,

each have assets totaling a little more than 1.7 trillion in 2013 dollars. When looking at the

high-risk aversion UMP, a back of the envelope calculation shows that the largest bank in the

entire sample has roughly 800 billion in total assets in 2013 dollars. Our results should therefore

not be interpreted as direct evidence against large banks per se, but more as an indication that

the current banks are excessively large to the point that there are potential financial stability

improvements if bank size is decreased.

[Insert Figure 2 near here]

Of course, our analysis so far has made use of book value data that are backward looking

in nature, and not available at a frequency higher than quarterly. Does that matter for our

first policy experiment? Figure 3, based on market value data, shows that it hardly does. The

risk measure is again able to pick up episodes of distress, showing spikes around the Savings

and Loans crisis, the LTCM/Asian crisis and the recent Subprime crisis. As with book value

data, the risk and return of the UMP decrease with higher risk aversion. Concentration in

the FED portfolio starts out relatively low in the beginning of the sample, but increases over

time, reflecting the consolidation in the US banking industry. In the UMP, concentration again

decreases when the supervisor has a higher degree of risk aversion. For instance, in the high risk

aversion case the concentration in the UMP is on average 50 percent and reaches at most 62

percent, whereas the average of the FED portfolio is 70 percent with a maximum of 86 percent.

Finally, the current largest banks in the FED portfolio see their weight reduced to no more than

15 percent at any time. These results hold even when we use stock returns instead of the returns

on the market value of assets.6

[Insert Figure 3 near here]

What type of banks are then favored by the supervisor in this first policy experiment? To

answer that question, we construct a crude industry level balance sheet of both the FED portfolio

and the high degree of risk aversion UMP in Figure 4. The Figure shows the broad composition

of assets (panels a and b) and liabilities (panels c and d) of the two portfolios over time, as well

as the reliance on non-interest income (panel e).

6
These additional results are available upon request.

14
 A first observation is that, compared to the FED portfolio, the disintermediation trend,

whereby bank lending becomes less of a core business, is much less prevalent in the UMP. While

banks in the FED balance sheet invest around 50 percent of assets in loans at the end of the

sample period, the UMP favors banks with a higher share of loans in total assets. Compared

to the FED, the share of loans in the UMP is roughly 10 percentage points higher. Second,

this difference in asset composition is also reflected in the earnings streams of banks in the FED

portfolio and the UMP: whereas in the former banks rely increasingly on non-interest income,

in the latter large banks balance non-interest income with interest income much more steadily

over the sample period, as shown in Figure 4e. Third, on the liabilities side banks in the FED

portfolio are more reliant on non-core deposit funding compared to the UMP, although no real

changes in leverage are observed.

[Insert Figure 4 near here]

When we look deeper into the composition of the loan portfolio in Figure 5, we see that the

UMP is geared towards more real estate lending than the FED portfolio. This comes mainly at

the expense of consumer loans, and to a lesser extent of C&I lending. As these real estate loans

are potentially more risky, we compare the share of non-performing loans between the FED

portfolio and the UMP in Panel 5c. There we see that delinquency rates are actually lower for

banks favored in the UMP, even though they do increase somewhat during the Subprime crisis.

[Insert Figure 5 near here]

Given the results in Figures 2-5, according to our first policy experiment the hypothetical

banking industry in the UMP is characterized by smaller banks and a less concentrated banking

system. Why does the supervisor want to hold more of these smaller banks compared to the large

ones? As it turns out, the banks favored by the UMP have a business model that favors retail

banking, since they are funded mainly by deposits, make loans and rely less on non-interest

income. Moreover, the loans are directed more at real estate, but this does not increase the

overall risk of the loan portfolio.

In the results so far, the UMP appears to give a bigger weight to banks with traditional

business models, even though we have not conditioned the portfolio optimization problem this

way. This suggests that besides size, business models ultimately also play a role, as they influence

both the riskiness of individual banks and the riskiness of the entire portfolio. Indeed, this is

15
 in agreement with a growing literature emphasizing the role of business models in financial

instability. For instance, Stiroh (2004) and Stiroh and Rumble (2006) find that non-interest

income reduces aggregate profits while increasing risk. More recent evidence by De Jonghe (2010)

shows that systemic risk is exacerbated by banks diversifying into activities other than lending,

due to increasing correlations between income streams. This finding was also corroborated by

DeYoung and Roland (2001), Brunnermeier et al. (2012), DeYoung and Torna (2013) and Adrian

and Brunnermeier (2016). Similarly, Boot and Ratnovski (2016) find that although there are

initial benefits for banks from starting trading activities, beyond a critical point inefficiencies

dominate and trading becomes increasingly risky. Our results are in line with the notion that a

safer banking system is characterized by more traditional activities and less trading. Regarding

the funding of banks, Huang and Ratnovski (2011) argue that wholesale lenders have lower

incentives for costly monitoring, leading to large (and inefficient) fluctuations of loans on negative

public signals, a problem not encountered in relationship banks. Furthermore, Fahlenbrach et al.

(2012) emphasize that banks with increasing balance sheets through the use of short term non-

deposit liabilities performed poorly during the last crises. This is consistent with our findings

as well, as the risk-averse supervisor prefers banks with more core deposits and less wholesale

funding.

4.2 What role do business models play in the risk/return trade-off ?

The findings so far suggest that besides size there is a role for business models when con-

sidering the optimal design of the banking system. A natural question is then how important

that role is. Next, we ask whether changing bank size is necessary in the first place, or whether

activity restrictions are sufficient to reduce portfolio risk. We try to answer this question by

obtaining UMPs in which banks are kept at their current size, but experience a change in their

business model. As described in Section 2, we do this by altering the composition of loans and

non-loan assets on a bank’s balance sheet.

In doing so, we make a few further simplifying assumptions. First, we assume that changing

the business model does not impact the relative profitability of each business line.7 As a result,

changing the proportion of assets used for lending changes the profitability of a bank, as well

7
We recognize that the set-up in this experiment and the next one ignores potential economies of scale and scope
that banks might have, since we keep the profitability of each business line (loans or investment assets) constant
despite business model focus and size. In unreported analyses we allowed the net interest income and non-interest
income to change with size, but adding this layer of complexity did not quantitatively or qualitatively change
our results and we therefore choose to proceed with the most simple set-up.

16
 as its covariance with other banks in the portfolio. Second, we keep each bank’s non-interest

expenses, profits on minority interests, and implied tax rate constant and equal to their original

values.8 We then obtain the UMP by optimizing the proportion of loans, α, where α = [0, 1],

for each bank.

Since this analysis is heavily based on the balance sheets and income statements of banks,

we perform this analysis only on book value data. The results are reported in Figure 6. We

again report the return and risk of both FED portfolio and UMP in the top two rows, as well

as the weighted business model composition in the bottom row. As Figure 6 shows, on average

there is little difference between the UMP and the FED portfolio in terms of risk, even for

increasing levels of risk aversion. Examining the optimal portfolio design, we see that with low

and moderate levels of risk aversion, banks have a lower proportion of loans on their balance

sheet compared to the FED portfolio. When the risk aversion is high, the two portfolios are

relatively similar in terms of the business model of the banks in each portfolio. We interpret

these results as evidence that given the current concentration levels, risk cannot be reduced

beyond that of the FED portfolio. Instead, to maximize utility, the UMP favors banks with a

higher return and less traditional business models. When risk aversion is high, at best the UMP

mimics the FED portfolio characteristics.

[Insert Figure 6 near here]

Comparing the two policy experiments so far, we observe a significant reduction in risk only

in the experiment in which we allow for concentration to change. As a consequence, based on

this second policy experiment, we conclude that risk in the banking portfolio is more affected

by changing concentration than by purely changing the business models of banks. In a financial

system dominated by large banks, bank size magnifies the effect of risk that is inherent to any

business model. As already shown in Gabaix (2011) and Carvalho and Gabaix (2013), shocks to

individual firms have the potential to lead to aggregate volatility when the size distribution of

an economy is heavy tailed, something that also holds for the banking system (see e.g. Janicki

and Prescott, 2006; Blank et al., 2009). We believe that the structure of the banking system is

the main driver of our findings so far: every bank business model is susceptible to shocks, but

these shocks are amplified by the presence of very large banks, which can potentially lead to

excessive aggregate volatility.

8
We thus limit the investigation of the business models to the assets side of the balance sheet, and forego potential
funding implications due to computational limitations in obtaining the UMP.

17
 All in all, these results therefore suggest that activity restrictions, on which regulators have

focused their efforts so far, are not sufficient to reduce portfolio risk unless they are supplemented

with corresponding reductions in bank size.

4.3 What role does concentration play in the risk/return trade-off when
business models change with bank size?

So far, we have investigated the effect of size and business models on portfolio risk in isolation,

and found evidence suggesting that bank size is the main driver of portfolio risk. However, we

have had to assume that there is no relationship between bank size and bank business model,

as we changed one while keeping the other one constant. This is quite a strict assumption since

business models tend to grow organically with size, and not all bank sizes may be able to sustain

a certain business model equally well. Therefore, in a third and final policy experiment we relax

this assumption and investigate a more realistic policy experiment where business models change

with bank size when obtaining the optimal portfolio.

To do so, we combine the previous two policy experiments. In every iteration of the op-

timization process, when we change a bank’s weight, we change its α correspondingly. As a

result, the change in the relative importance of this bank in the portfolio affects the expected

return matrix and the covariance matrix, both of which are inputs in the utility function. We

parametrize the relationship between the weight of the bank and its α by running a fixed effects

regression of banks’ α on their weight in the FED portfolio. The coefficient on weight then

indicates how much the balance sheet composition of a bank changes as its size changes. The

results of this regression are shown in Table 2, where we can see that an increase in a bank’s

weight of 1 percentage point results in a roughly 0.5 percentage point decrease in the proportion

of loans with respect to total assets.9

[Insert Table 2 near here]

We show the results for this final policy experiment in Figure 7, where we can see the risk

and return of the two portfolios (the top two rows) as well as their concentration (the third row)

and business models (bottom row). The risk and return are quite similar to those found in the

first policy experiment, as both decrease with higher levels of risk aversion. When the degree
9
We further check whether there is time-variation in this slope coefficient by running the regression on different
time samples. However, the results are stable with coefficients around -0.5 and we therefore use the value as
reported in Table 2.

18
of risk aversion is low, the UMP is characterized by episodes of severe distress, which are even

larger in size than those in the FED portfolio. Once the degree of risk aversion increases, risk is

almost always lower in the UMP compared to the FED portfolio. Meanwhile, the concentration

in these UMPs decreases as the degree of risk aversion of the supervisor managing the portfolio

increases. Most importantly, however, the business models in the UMPs and the FED portfolio

are very similar.

[Insert Figure 7 near here]

As a result, we conclude that, even in this final policy experiment where bank size and

business model change simultaneously in a more realistic manner, bank size remains the key

driver of portfolio risk. This finding underlines the main takeaway from our analysis, that

concentration levels should be 40 to 60 percentage points lower than is currently the case,

depending on how risk averse regulators are. This reduction in concentration would still allow

for the existence of large banks, as another back of the envelope calculation shows that the

largest bank in the high risk aversion UMP has assets totaling above 1 trillion dollars, while

still operating under a more traditional business model with an α of 0.8. It seems as if the

existing diversity in business models in the US financial sector is not a problem in itself, but

rather it is the business model favored by the few very large banks that increases portfolio risk.

As a consequence, activity restrictions as suggested by Paul Volcker, the Liikanen report and

the Vickers report are likely not enough to reduce portfolio risk to a lower level. Instead, these

results lend support to a renewed focus in regulatory efforts to reduce the size of the largest

banks, either directly via caps on size, or indirectly via incentives to downsize, such as increasing

capital requirements for the largest banks.

5 Robustness

The three policy experiments in this paper are intended to help us understand the role of

bank size in managing the risks of a banking system. They are ‘thought experiments’, as they

rely - sometimes heavily - on assumptions. Nevertheless, to see how robust our results are, we

now put several of the key assumptions to the test. First, as excessive reweighting of banks

can be costly and inefficient, we begin by exploring a more realistic scenario under which the

reweighting (both in terms of size and business model) is limited. Second, since portfolio risk

should primarily capture downside risk, we replace the symmetric variance-covariance matrix

19
 with one that takes this asymmetry into account. Third, we explore the robustness of our results

with respect to the size of the rolling window used to perform the analysis.

5.1 Adjusting the portfolio

In the baseline analyses, we have obtained the UMPs by allowing unlimited reshuffling of

assets amongst banks. Although reshuffling also occurs naturally in the FED portfolio, either

via mergers and acquisitions, bank entry and exit, or bailouts, potentially it happens more

frequently in the UMP. In order to assess how stable the UMP is over time compared to the

FED portfolio, we calculate the turnover of both portfolios. Turnover is defined as the sum of

absolute weight changes in the portfolio between period t − 1 and t, taking values ranging from

zero (no change) to two (where all assets that were held are sold, and the assets that were not

held are bought). Figure A.1 in the Internet Appendix plots the turnover for both portfolios in

the case where business models are allowed to change with bank size. We observe that turnover

in the UMP decreases with the degree of risk aversion. However, with an average of 0.18 in the

high risk aversion case it is still 35 percent higher than in the FED portfolio. Of course, this

turnover comes at a substantial cost to society, which may include switching costs for depositors,

borrowers as well as other bank stakeholders, and possible liquidation costs when a bank sees its

weight reduced to zero. To limit unnecessary reweighting, and potentially reduce the associated

costs, we perform two sets of robustness tests. First, we allow banks to grow/shrink by no

more than 20 percent of their initial weight. Second, we keep the largest 5 percent of banks at

their current cumulative size, allowing unlimited reweighting of the remaining banks while still

adhering to the no-shorting and no-loss constraints.10 The results for both tests are shown in

Section A in the Internet Appendix.

Interestingly, regardless of the specification, tightening the reweighting in our portfolio leads

to no real improvement in the risk-return trade-off in the UMP compared to the FED portfolio.

In fact, the outcomes of the UMP closely follow those of the FED portfolio. These results are

consistent with findings in the previous section and suggest that without a sufficient decrease

in concentration, which requires a more drastic reweighting, it is not possible to improve the

portfolio risk in this setting.

10
Note that adding these additional constraints is roughly equivalent to adding a cost for turnover directly in the
objective function, and is computationally less intensive.

20
 5.2 Downside risk

In the second robustness test we revisit our definition of risk. Following directly from the

utility function of the supervisor, we have defined risk as the variance based on both gains and

losses in the portfolio. A supervisor, however, may be mainly concerned about downside risk:

the risk of losses. This is also reflected in the current strand of literature on systemic risk, which

focuses on tails of the return distribution (De Jonghe, 2010; Acharya et al., 2012a; Acharya et

al., 2012b; Huang et al., 2012; Engle et al., 2015; Adrian and Brunnermeier, 2016; Black et al.,

2016; Brownlees and Engle, 2016). Following this view, we also measure portfolio risk using a

semi-covariance matrix instead of the symmetric variance-covariance matrix. We calculate the

semi-variance as the variance of the set of returns that is below its mean, and semi-covariances

in a similar manner.11 However, this approach has two downsides. First, any two banks may not

necessarily post below-average returns simultaneously, meaning that the semi-covariance of (i, j)

will be different from that of (j, i). The result can be a non-symmetric covariance matrix, which

may potentially fail to be positive semi-definite. Second, this measure of risk is not compatible

with the expected utility framework as defined in Section 2. Given these two downsides, we

present the results as a robustness test only. The results are presented in Section B in the

Internet Appendix, and are quantitatively and qualitatively similar to the baseline scenarios,

further underlining the importance of concentration as the driving force behind our findings.

5.3 Rolling windows

Finally, we test the robustness of the book value results to the choice of the length of the

rolling window. Given that in the baseline scenario we use only eight quarters (observations)

of data, adding or removing observations can very quickly lead to changes in optimal portfolio

design. First, to match the time frames in the analysis on market value data, we perform the

analysis on sets of all four quarters in a given year. Second, we increase the original window

size to sixteen quarters in order to eliminate the impact of outliers. The results are shown

in Section C in the Internet Appendix, and follow similar patterns to the baseline analyses.

11
Most systemic risk measures rely on using stock return data and can therefore focus on returns that are far
deeper in the tail. Often the 5 percent worst returns are chosen to obtain a measure of downside risk. While this
could potentially work in our setting when using daily market returns, we would have to extend our quarterly
book value return on assets to far beyond what is available. For example, only when we observe a bank’s returns
for 25 straight years would we have 5 observations on which we can calculate the book value tail variance and
covariance. Our analysis relies mainly on book returns as this allows us to manipulate the balance sheets and
income statements, we therefore only use the semi-variance for both book value and market value returns to
perform this robustness test.

21
Therefore, the results are robust to using different window lengths.

6 Conclusion

In this paper we describe three policy experiments aimed at exploring the relationship be-

tween bank size and systemic risk. With these experiments we try to provide a first attempt to

answer the question how big banks should be to maintain a healthy financial system that is able

to withstand shocks.

We view the supervisor as a constrained investor in a portfolio of banks. By applying elements

of Modern Portfolio Theory, we derive a hypothetical distribution of bank sizes and business

models that the supervisor should have held to arrive at an optimal portfolio in order to give us

insights, albeit of a partial equilibrium nature, into the ideal design of the banking system.

To isolate the effect of bank size and bank business model, we run three policy experiments

in which we obtain a supervisors hypothetical optimal portfolio. First, to focus on the effect of

size alone, we keep the business model and earnings of a bank as they are and change only its

size. We find that under a medium to high level of risk aversion the supervisor would prefer to

reduce bank concentration and overall portfolio risk. Moreover, in the optimal portfolio greater

weight is given to more traditional banks that make loans, are funded by core deposits and have

a lower level of non-interest income. Second, to investigate the role that business models play,

we keep the banks at their current size and let the supervisor choose optimal business models

(ranging from more traditional banks to universal/investment banks). However, in this setting

the portfolio risk cannot be reduced below actual values given the current level of concentration.

Third, we combine both approaches by letting the bank’s business model and risk/return profile

depend on its size. That way, when a bank’s size is altered, not only will its balance sheet size

change, but so will the composition of its earning assets. We again find that concentration is

lowered in order to achieve the optimal portfolio, whereas bank business models are roughly

the same as today. The results show that the largest banks are consistently overrepresented in

the current portfolio compared with the supervisor’s optimal portfolio and that in this setting

concentration is the driving force of portfolio risk.

So where do these findings leave us, and how can our paper help to guide policy? Placing

these results in broader perspective is not easy to do given the hypothetical nature of our exercise,

however the results fit into a narrative regarding the role of bank size that both regulatory and

22
 academic efforts have focused on. Calls to ‘break up the banks’ started directly in the aftermath

of the crisis, and came from a wide range of academics (amongst others, Joseph Stiglitz, Paul

Krugman, Ed Prescott, Luigi Zingales and Glenn Hubbard), former chairmen of the Federal

Reserve Alan Greenspan and Paul Volcker, Federal Reserve Board governor Daniel Tarullo and

president of the Dallas Federal Reserve Paul Fisher, and even from former CEOs of large banks

such as Sandy Weill.12 Some critics have pointed out that breaking up the banks could come at

considerable costs, as recent literature has found that even large banks can still enjoy economies

of scale (see, e.g., Wheelock and Wilson, 2012; Hughes and Mester, 2013). However, it is still

debated whether the funding advantages that banks receive from being Too-Big-To-Fail are

properly accounted for in estimating these scale economies (Demirgüç-Kunt and Huizinga, 2013;

Ueda and Weder di Mauro, 2013; Davies and Tracey, 2014).

Nonetheless, most seem to agree that banks should be broken up in one way or another.

However, there is considerable disagreement as to how we should break up the banks (Barth

and Prabha, 2013). Practical limitations center around the questions of whether banks should

be capped in size, but allowed to perform both commercial banking and investment banking, or

whether investment banks should be completely separated from the commercial bank. Moreover,

if a ‘pure’ cap is instituted, what should it be based on: total assets of the bank, its insured

deposit liabilities, or its size relative to GDP? And if a measure is agreed upon, what is the

correct size at which banks should be capped? One suggestion was proposed by Senators Brown

and Kauffman as an amendment (the SAFE Banking Act) to Dodd-Frank, but was ultimately

dropped. Their amendment came closest to the experiments we performed in our paper, and

intended to limit the relative size of banks with regards to GDP. Outside of the U.S. and

Switzerland, where a cap was discussed in 2009, no other countries seriously considered the

option or seem to be heading along this route. Nonetheless, the results in this paper argue for

a renewed focus on this idea of a cap on size, as we found that restrictions based on size can

potentially reduce the level of risk to a lower level.

An avenue that regulators did choose is to levy a systemic surcharge on large banks, such

that they are required to hold more regulatory capital. The Financial Stability Board has

implemented the Total Loss Absorption Capacity (TLAC), which forces global systemically

important banks to hold up to 16 percent regulatory capital to risk-weighted assets by January

2019, increasing to 18 percent by January 2022. At worst this regulation should make banks

12
See Ritholtz (2013) for an extensive overview.

23
 safer, and at best it could raise their cost of capital to the extent that they endogenously decide

to downsize. If banks do decide to downsize via this avenue, regulators could increase these

additional capital requirements until the banks are at a size where they no longer pose a threat

to financial stability. However, it remains to be seen whether the current surcharges are high

enough to incentivize banks to do so (Admati et al., 2010; Haldane, 2012).

A second avenue regulators have taken is that of activity restrictions, i.e. breaking up the

commercial and investment banking parts to ensure that insured deposits are not at risk due

to speculative investments. In the U.S., Dodd-Frank included the Volcker Rule, prohibiting

banks from engaging in proprietary trading. In Europe there have been similar proposals,

with the Vickers report calling to completely ring-fence commercial banking activities from

investment banking activities as well as ring-fence domestic and foreign operations, and the

Liikanen report suggesting to separate proprietary trading. So far, only the Volcker rule has

come into effect, in 2015. Nonetheless, under this rule banks can still perform trading activities

for their own customers, engage in market-making activities, as well as proprietary trading under

certain exceptions. Questions regarding the feasibility of the Volcker rule have been posed in

the meantime, with outgoing Federal Reserve Board governor Daniel Tarullo arguing that it is

too complicated and should be amended (Tarullo, 2017). Moreover, the banks in question have

recently asked for an extended 5 year grace period to sell-off highly illiquid assets. In the U.K.,

banks need to be compliant with the Vickers report by 2019, and while efforts are underway to

so, it remains to be seen to what extent they will manage.

The Volcker Rule, and Vickers/Liikanen report mimic the the Glass-Steagall regulatory set-

up, which separated trading and banking activities, and seemed to work in a stabilizing manner

until it was undone at the end of the 1990s.13 Our results, however, show that simply restricting

the activities of banks might not be enough to reduce systemic risk as long as bank size is not

lowered sufficiently. In particular, in our second and third experiments we have seen that the

existing diversity in business models does not seem to be the main driver of systemic risk in U.S.

banking, but rather it is the business model favored by the few very large banks that increases

risk. While the analyses provided in this paper are relatively crude and not without drawbacks,

one potential way of interpreting our findings and reconciling them with the stability of the

Glass-Steagall period, is that Glass-Steagall could have been so successful simply because the

13
DeYoung and Torna (2013) show, however, that the Gramm-Leach-Bliley Act that essentially removed Glass-
Steagall was not the main driver of bank failures in the recent crisis.

24
banking system was not as heavily concentrated as it is today, since banks operated under a

regulatory set-up which also restricted geographical expansion. Therefore, it is questionable

whether activity restrictions alone can lead to a more stable banking system.

Despite the experimental nature of this paper, we hope to contribute to the discussions on

the optimal size of banks by providing a different angle on the issue of Too-Big-To-Fail and how

to possibly address it. As always, further research is required to better understand how we could

move towards a more resilient banking system.

25
References
Acharya, V. V. (2009). A theory of systemic risk and design of prudential bank regulation.
Journal of Financial Stability 5 (3), pp. 224–255.

Acharya, V. V., R. F. Engle, and M. P. Richardson (2012a). Capital Shortfall: A new approach
to ranking and regulating systemic risks. American Economic Review 102 (3), pp. 59–64.

Acharya, V. V., L. H. Pedersen, T. Philippon, and M. P. Richardson (2012b). Measuring systemic

risk. CEPR Discussion Papers 8824, C.E.P.R. Discussion Papers.

Acharya, V. V. and T. Yorulmazer (2007). Too many to fail – An analysis of time-inconsistency

in bank closure policies. Journal of Financial Intermediation 16 (1), pp. 1–31.

Admati, A. R., P. M. DeMarzo, M. F. Hellwig, and P. Pfleiderer (2010). Fallacies, irrelevant

facts, and myths in the discussion of capital regulation: Why bank equity is not expensive.
Research Papers 2065, Stanford University, Graduate School of Business.

Adrian, T. and M. K. Brunnermeier (2016). CoVaR. American Economic Review 106 (7), pp.
1705–1741.

Alegria, C. and K. Schaeck (2009). On measuring concentration in banking systems. Finance

Research Letters 5, pp. 59–67.

Allen, F., A. Babus, and E. Carletti (2012). Asset commonality, debt maturity and systemic
risk. Journal of Financial Economics 104 (3), pp. 519–534.

Allen, F. and E. Carletti (2008). Mark-to-market accounting and liquidity pricing. Journal of
Accounting and Economics 45 (2-3), pp. 358–378.

Ashcraft, A. B. (2005). Are Banks Really Special? New Evidence from the FDIC-Induced
Failure of Healthy Banks. American Economic Review 95 (5), pp. 1712–1730.

Barth, J. R. and A. Prabha (2013). Breaking (banks) up is hard to do: New perspective on Too
Big to Fail. In The Social Value of the Financial Sector Too Big to Fail or Just Too Big?,
World Scientific Book Chapters, Chapter 20, pp. pp. 377–400. World Scientific Publishing Co.
Pte. Ltd.

Beck, T., O. De Jonghe, and G. Schepens (2013). Bank competition and stability: Cross-country
heterogeneity. Journal of Financial Intermediation 22 (2), 218–244.

Bernanke, B. S. (1983). Nonmonetary effects of the financial crisis in propagation of the Great
Depression. American Economic Review 73 (3), pp. 257–276.

Bisias, D., M. Flood, A. W. Lo, and S. Valavanis (2012). A survey of systemic risk analytics.
Annual Review of Financial Economics 4 (1), pp. 255–296.

Black, L., R. Correa, X. Huang, and H. Zhou (2016). The systemic risk of European banks
during the financial and sovereign debt crises. Journal of Banking & Finance 63, pp. 107–125.

26
Blank, S., C. M. Buch, and K. Neugebauer (2009). Shocks at large banks and banking sector
distress: The banking granular residual. Journal of Financial Stability 5 (4), pp. 353–373.

Boot, A. W. A. and L. Ratnovski (2016). Banking and trading. Review of Finance 20 (6), pp.
2219–2246.

Brown, C. O. and I. S. Dinç (2011). Too many to fail? Evidence of regulatory forbearance when
the banking sector is weak. Review of Financial Studies 24 (4), pp. 1378–1405.

Brownlees, C. T. and R. F. Engle (2012). Volatility, correlation and tails for systemic risk
measurement. Working paper. <http://ssrn.com/abstract=1611229>.

Brunnermeier, M. K., G. Dong, and D. Palia (2012). Banks’ non-interest income and systemic
risk. Working paper. <http://ssrn.com/abstract=1786738>.

Calomiris, C. W. and J. R. Mason (2003). Fundamentals, panics, and bank distress during the
depression. American Economic Review 93 (5), pp. 1615–1647.

Carvalho, V. and X. Gabaix (2013). The great diversification and its undoing. American
Economic Review 103 (5), pp. 1697–1727.

Claessens, S., M. A. Kose, and M. E. Terrones (2009). What happens during recessions, crunches
and busts? Economic Policy 24, pp. 653–700.

Claeys, S. and K. Schoors (2007). Bank supervision Russian style: Evidence of conflicts between
micro- and macro-prudential concerns. Journal of Comparative Economics 35 (3), pp. 630–657.

Davies, R. and B. Tracey (2014). Too big to be efficient? The impact of implicit subsidies on
estimates of scale economies for banks. Journal of Money, Credit and Banking 46 (s1), pp.
219–253.

De Jonghe, O. (2010). Back to the basics in banking? A micro-analysis of banking system

stability. Journal of Financial Intermediation 19 (3), pp. 387–417.

Degryse, H. and G. Nguyen (2007). Interbank exposures: An empirical examination of contagion

risk in the Belgian banking system. International Journal of Central Banking 3 (2), pp. 123–
171.

Dell’Ariccia, G., E. Detragiache, and R. Rajan (2008). The real effect of banking crises. Journal
of Financial Intermediation 17, pp. 89–112.

Demirgüç-Kunt, A. and H. Huizinga (2013). Are banks too big to fail or too big to save? Inter-
national evidence from equity prices and CDS spreads. Journal of Banking & Finance 37 (3),
pp. 875–894.

DeYoung, R. and I. Hasan (1998). The performance of de novo commercial banks: A profit
efficiency approach. Journal of Banking & Finance 22, pp. 565–587.

27
DeYoung, R. and K. P. Roland (2001). Product mix and earnings volatility at commercial banks:
Evidence from a degree of total leverage model. Journal of Financial Intermediation 10 (1),
54–84.

DeYoung, R. and G. Torna (2013). Nontraditional banking activities and bank failures during
the financial crisis. Journal of Financial Intermediation 22 (3), pp. 397–421.

Engle, R., E. Jondeau, and M. Rockinger (2015). Systemic risk in Europe. Review of Fi-
nance 19 (1), pp. 145–190.

Fahlenbrach, R., R. Prilmeier, and R. M. Stulz (2012). This time is the same: Using bank
performance in 1998 to explain bank performance during the recent financial crisis. The
Journal of Finance 67 (6), pp. 2139–2185.

Federal Reserve Bank of New York (2014). CRSP-FRB Link.

Furfine, C. H. (1999). The microstructure of the Federal Funds market. Financial Markets,
Institutions & Instruments 8 (5), pp. 24–44.

Furfine, C. H. (2003). Interbank exposures: Quantifying the risk of contagion. Journal of Money,
Credit and Banking 35 (1), pp. 111–128.

Gabaix, X. (2011). The granular origins of aggregate fluctuations. Econometrica 79 (3), pp.
733–772.

Gandhi, P. and H. Lustig (2015). Size anomalies in U.S. bank stock returns. Journal of Fi-
nance 70 (2), pp. 733–768.

Haldane, A. (2012). On being the right size. Speech Given at the Institute of Economic Affairs’
22nd Annual Series of The Beesley Lectures, London.

Huang, R. and L. Ratnovski (2011). The dark side of bank wholesale funding. Journal of
Financial Intermediation 20 (2), pp. 248–263.

Huang, X., H. Zhou, and H. Zhu (2012). Systemic risk contributions. Journal of Financial
Services Research 42 (1), pp. 55–83.

Hughes, J. P. and L. J. Mester (2013). Who said large banks don’t experience scale economies?
Evidence from a risk-return-driven cost function. Journal of Financial Intermediation 22 (4),
pp. 559–585.

Ibragimov, R., D. Jaffee, and J. Walden (2011). Diversification disasters. Journal of Financial
Economics 99 (2), pp. 333–348.

Janicki, H. P. and E. S. Prescott (2006). Changes in the size distribution of U.S. banks: 1960-
2005. Economic Quarterly, pp. 291–316.

Kelly, B., H. Lustig, and S. V. Nieuwerburgh (2016). Too-Systemic-to-Fail: What option markets
imply about sector-wide government guarantees. American Economic Review 106 (6), pp.
1278–1319.

28
Koetter, M., J. W. Kolari, and L. Spierdijk (2012). Enjoying the quiet life under deregula-
tion? Evidence from adjusted lerner indices for U.S. banks. The Review of Economics and
Statistics 94 (2), pp. 462–480.

Laeven, L. and R. Levine (2009). Bank governance, regulation and risk taking. Journal of
Financial Economics 93 (2), pp. 259–275.

Ledoit, O. and M. Wolf (2003). Improved estimation of the covariance matrix of stock returns
with an application to portfolio selection. Journal of Empirical Finance 10 (5), pp. 603–621.

Ledoit, O. and M. Wolf (2004). Honey, I shrunk the sample covariance matrix. The Journal of
Portfolio Management 30 (4), pp. 110–119.

Liu, H., L. Norden, and F. Spargoli (2015). Why banks want to be complex. Working Paper.
<http://ssrn.com/abstract=2697220>.

Markowitz, H. (1952). Portfolio selection. The Journal of Finance 7 (1), pp. 77–91.

Markowitz, H. (1959). Portfolio Selection: Efficient Diversification of Investments. New York:

John Wiley & Sons.

Pais, A. and P. A. Stork (2011). Bank size and systemic risk. European Financial Manage-
ment 19 (3), pp. 429–451.

Rajan, R. G. (2005). Has financial development made the world riskier? Proceedings Aug, pp.
313–369.

Reinhart, C. M. and K. S. Rogoff (2009). The aftermath of financial crises. American Economic
Review 99 (2), pp. 466–472.

Ritholtz (2013). Top economists and financial experts say we must break up the giant banks. Ac-
cessed on March 31, 2017 from <http://ritholtz.com/2013/03/top-economists-and-financial-
experts-say-we-must-break-up-the-giant-banks/>.

Stiroh, K. J. (2004). Diversification in banking: Is noninterest income the answer? Journal of

Money, Credit and Banking 36 (5), pp. 853–882.

Stiroh, K. J. and A. Rumble (2006). The dark side of diversification: The case of US financial
holding companies. Journal of Banking & Finance 30, pp. 2131–2161.

Tarullo, D. (2017). Departing thoughts. Speech Given at The Woodrow Wilson School, Princeton
University, Princeton, New Jersey.

Ueda, K. and B. Weder di Mauro (2013). Quantifying structural subsidy values for systemically
important financial institutions. Journal of Banking & Finance 37 (10), pp. 3830–3842.

van Lelyveld, I. and F. Liedorp (2006). Interbank contagion in the Dutch banking sector: A
sensitivity analysis. International Journal of Central Banking 2 (2), pp. 99–133.

29
Wagner, W. (2008). The homogenization of the financial system and financial crises. Journal of
Financial Intermediation 17 (3), pp. 330–356.

Wagner, W. (2010). Diversification at financial institutions and systemic crises. Journal of

Financial Intermediation 19 (3), pp. 373–386.

Weiß, G. N., S. Neumann, and D. Bostandzic (2014). Systemic risk and bank consolidation:
International evidence. Journal of Banking & Finance 40, pp. 165–181.

Wheelock, D. C. and P. W. Wilson (2012). Do large banks have lower costs? New estimates of
returns to scale for U.S. banks. Journal of Money, Credit and Banking 44 (1), pp. 171–199.

30
Tables and Figures

Table 1: Summary statistics

Variable Obs Mean Std. Dev Min Max

A. Book value
Return on assets 333,035 0.26 0.18 -1.18 0.85
Weight 333,035 0.03 0.30 0.00 21.62

Return on equity 333,035 2.84 2.16 -26.94 18.12

Equity/Total assets 333,035 9.44 2.43 4.20 22.22
Non-interest income/Total income 333,035 16.53 7.33 3.35 45.08
Non-interest income/Non-loan assets 333,035 0.70 0.56 0.06 4.01
Net interest income/Total loans 333,035 1.71 0.49 0.89 4.12
Loans/(Total assets - Fixed assets) 333,035 63.21 14.19 7.66 98.83

B. Market value
Return on assets 1,786,587 0.01 0.41 -34.94 43.15
Weight 1,786,587 0.33 1.46 0.00 23.37
This table shows summary statistics for the main variables used in the analysis. Panel A shows the book
value variables, while Panel B shows the market value variables. The book value variables are constructed as
follows. Return on assets an Return on equity are calculated using Net Income (RIAD4079) and Total Assets
(RCFD2170) and Total Equity (RCFD3210), respectively. Weight is the market share that each bank has in
a given quarter based on the total assets. Other variables are obtained as: Non-interest income (RIAD4079),
Total income (RIAD4079 + RIAD4107 - RIAD4073), Non-loan assets (RCFD2170 - RCFD2122), Total loans
(RCFD2122), Fixed assets (RCFD2145 + RCFD3163 + RCFD2160). The market value Return on assets is
calculated as the daily percentage change in the market value of assets, which is obtained as the sum of market
capitalization and book value of liabilities. The weight is the daily market share based on the market value of
assets.

31
 Table 2: How does balance sheet composition depend on bank size?

α
Constant 0.564***
(0.003)
w -0.455**
(0.183)

Bank fixed effects Included

Time fixed effects Included
R2 0.227
N 239,520
This table shows the relationship between bank size (w) and the share of loans in assets (α), which is used in
Section 4.3 to investigate the role of concentration in determining systemic risk when bank business models
change with size. Robust standard errors in parentheses. * significant at 10 percent; ** significant at 5 percent;
*** significant at 1 percent.

Figure 1: The supervisory view: Markowitz efficient frontier and risk aversion

Low risk aversion

Expected return: µ = w r
T

Medium risk aversion

H   sk aversion

2 T
Expected risk:
= w Σw

32
 Figure 2: What role does concentration play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Return (in %)
Return (in %)
Return (in %)
0.20 0.20 0.20

33
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel 2a shows the difference in risk,
return and concentration of each portfolio for low levels of risk aversion, while panels 2b and 2c show this for medium and high levels of risk aversion, respectively. The concentration in each
portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The levels of risk aversion are chosen based on the price of
risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
 Figure 3: What role does concentration play in the risk/return trade-off when using market values?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

1.50 1.50 1.50

1.00 1.00 1.00

0.50 0.50 0.50

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP FED UMP FED UMP FED

0.30 0.30 0.30

0.20 0.20 0.20

0.10 0.10 0.10

Return (in %)
Return (in %)
Return (in %)

34
0.00 0.00 0.00

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using stock market value data. Panel 3a shows the difference in
risk, return and concentration of each portfolio for low levels of risk aversion, while panels 3b and 3c show this for medium and high levels of risk aversion, respectively. The concentration
in each portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The levels of risk aversion are chosen based on the
price of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 10, while medium and high levels of risk aversion are λ = 50 and λ = 100, respectively.
 Figure 4: What happens to the intermediary role of banks in a more stable banking system?
(a) FED - Assets (b) UMP High risk aversion - Assets

Assets Assets
100 100 100 100
Balance Sheet Composition (in %)

Balance Sheet Composition (in %)

80 80 80 80

60 60 60 60

40 40 40 40

20 20 20 20

0 0 0 0
1990q4 1995q4 2000q4 2005q4 2010q4 1990q4 1995q4 2000q4 2005q4 2010q4

Fixed Investment Assets Fixed Investment Assets

Loans Loans

(c) FED - Liabilities (d) UMP High risk aversion - Liabilities

Liabilities Liabilities
100 100 100 100
Balance Sheet Composition (in %)

Balance Sheet Composition (in %)

80 80 80 80

60 60 60 60

40 40 40 40

20 20 20 20

0 0 0 0
1990q4 1995q4 2000q4 2005q4 2010q4 1990q4 1995q4 2000q4 2005q4 2010q4

Equity Other Liabilities Equity Other Liabilities

Wholesale Funds Core Deposits Wholesale Funds Core Deposits

(e) Is there a difference in income characteristics?

35.0
Non−Interest Income/Total Income (in %)

30.0

25.0

20.0

15.0

10.0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

FED UMP

The figure compares the intermediary role of the FED portfolio and high risk aversion UMP, by showing weighted average
industry level balance sheets in panels 4a-4d and the different weighted non-interest income/total income ratios in panel 4e.
To construct the balance sheet we use the following data series: Fixed & Other Assets (RCFD2145 + RCFD3163 +
RCFD2160); Loans (RCFD2122); Investment Assets (RCFD2170 - Loans - Fixed & Other Assets); Equity (RCFD3210);
Core Deposits (RCON2215 + RCON6810 + RCON0352 + RCON6648); Wholesale Funding (RCON2604 + RCFN2200
+ RCFD3200 + RCFD2800 + RCONb993 + RCFDb995 + RCFD3190); Other Liabilities (RCFD2948 - Core Deposits -
Wholesale Funding).

35
 Figure 5: Does individual bank risk increase in a more stable banking system?
(a) FED - Loan portfolio (b) UMP High risk aversion - Loan portfolio
100 100 100 100
Loan Portfolio Composition (in %)

Loan Portfolio Composition (in %)

80 80 80 80

60 60 60 60

40 40 40 40

20 20 20 20

0 0 0 0
1990q4 1995q4 2000q4 2005q4 2010q4 1990q4 1995q4 2000q4 2005q4 2010q4

C&I Loans Real Estate Loans C&I Loans Real Estate Loans
Consumer Loans Other Loans Consumer Loans Other Loans

(c) Is there a difference in nonperforming loans?

3.0
Nonperforming Loans / Total Loans (in %)

2.5

2.0

1.5

1.0

0.5

0.0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

FED UMP

The figures compares how individually risky banks are in the FED portfolio compared to the UMP. In panels 5a and 5b
we show the loan portfolio composition for the FED portfolio and the high risk aversion UMP, by using weighted average
industry level loan portfolios. In panel 5c we show the weighted nonperforming loans (as a ratio of total assets) for both
portfolios. To construct the loan portfolio we use the following data series: C&I Loans (RCFD1766); Real Estate Loans
(RCFD1410); Consumer Loans (RCFD1975); Other Loans (RCFD1400 - RCFD1766 - RCFD1410 - RCFD1975). All series
are scaled by total loans (RCFD1400). Nonperforming loans are constructed as RCFD1403 + RCFD1407 and are scaled
by total assets (RCFD2170).

36
 Figure 6: What role does the business model of banks play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.60 0.60 0.60

0.40 0.40 0.40

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)

37
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel 6a shows the difference in risk, return
and business model (α) of each portfolio for low levels of risk aversion, while panels 6b and 6c show this for medium and high levels of risk aversion, respectively. The expected utility of the
supervisor is maximized by only changing the business models of banks while keeping the weight equal to the actual weight. The business model of a bank is a weight, α, on the proportion
of loans that a bank has, while (1 − α) is the proportion of non-loan (investment) related activities. The banks are assumed to keep the same level of profitability on each of these business
lines as α changes. Here, the weighted average α is shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion are chosen based on the price of risk (Sharpe
ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
 Figure 7: What role does concentration play in the risk/return trade-off when business models change with bank size?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion
0.20 0.20 0.20

0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

1.00 1.00 1.00

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Return (in %)
Return (in %)
Return (in %)
0.20 0.20 0.20

0.00 0.00 0.00

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

38
60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP 5% FED 5% UMP 5% FED 5% UMP 5% FED 5%

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel 7a shows the difference in risk, return,
concentration and business model (α) of each portfolio for low levels of risk aversion, while panels 7b and 7c show this for medium and high levels of risk aversion, respectively. The expected
utility of the supervisor is maximized by changing the weights of banks. However, as banks are made smaller or larger, the business model of the banks (α) change in the following way: as
banks receive a 1 percent increase in weight, the proportion of loans that a bank has decreases by 0.45 percent. The profitability on both business lines (loans and investments) is assumed
to stay the same as α changes. Due to the change in proportion α, the total income of the bank changes and with it the covariance matrix as well. Here, the outcome and concentration and
the weighted average α are shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion are chosen based on the price of risk (Sharpe ratio) of the data, and in
this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
Appendix
The coefficient of risk aversion influences the balance between risk and return in the utility
function. However, the coefficient itself is sensitive to the ratio of risk and return (i.e., the Sharpe
ratio) for assets under consideration, thus affecting the portfolio allocation and its expected
utility. In order to make these allocation decisions, risk aversion values between 1 and 10 are
usually chosen for assets of firms, such as those in the S&P500. The main difference between
these firms and our sample of banks is the fact that the latter have higher Sharpe ratios, both for
book value returns and market value returns. For book value returns, this can be caused by, e.g.,
a difference in data frequency (quarterly versus monthly), or by earnings smoothing which lowers
the variation in returns. In case of market returns, potential Too-Big-To-Fail considerations can
play a role here (Gandhi and Lustig, 2015; Kelly et al., 2016).
To see why this difference in Sharpe ratios matters, consider again the expected utility
function we use:
λ 2
µ + σ2 .


E [U ] = µ − (1)
2
To arrive at the same level of expected utility under a different µ over σ ratio requires a
different value for λ. We illustrate this with a simple example. First, we define µi , σi and
µi
SRi = σi as the return, standard deviation and Sharpe ratio of asset i. In the case where we
have 2 assets, and where the Sharpe ratios are quite different, we expect different values for the
risk aversion coefficients, λ1 and λ2 . In order to use the same expected utility framework, we
first equate both expected utility levels:

λ1 2 λ2 2
µ1 + σ12 = µ2 − µ2 + σ22



µ1 − (2)
2 2

Substituting the Sharpe ratio into this expression, and equating the returns, µ1 = µ2 , we obtain
the following expression for λ2 as a function of λ1 , SR1 and SR2 :
1
1+ SR12
λ2 = λ1 1 . (3)
1+ SR22

Conventional levels of risk aversion usually range from 1 (low risk aversion) to 10 (high risk
aversion). We use the Sharpe ratio of the benchmark S&P500 to adjust the risk aversion levels
for the assets we consider in our analysis. The monthly Sharpe ratio of the S&P500 for the
sample under consideration, SR1 , is roughly 0.15, while the (quarterly) book value Sharpe ratio,
SR2 is 2.89. Using a value of λ1 = 10, the equivalent risk aversion value for the book value data
is:

1
1+ 0.152
λ2 = 10 × 1 = 405. (4)
1+ 2.892

Similarly, our market value data have a (monthly) Sharpe ratio of 0.3, such that the equivalent
risk aversion value is:

39
 1
1+ 0.152
λ2 = 10 × = 37.5. (5)
1 + 0.31 2
In the table below we illustrate that, for a given value of µ, the utilities are indeed the same:

Table 1: Comparison risk aversion levels

U.S. banks
S&P500 Book value data Market value data
λ 10 405 37.5
SR 0.15 2.89 0.3
µ 0.0026 0.0026 0.0026
σ 0.0173 0.0009 0.0087

E [U ] 0.001064 0.001064 0.001064

Therefore, the values for λ we use in the paper are equivalent to conventional values, once we
take into account the differences in the Sharpe ratios due to, e.g., different data frequency and/or
earnings smoothing.

40
 For Online Publication

Internet Appendix
This Internet Appendix contains results that were omitted from the body of the paper for
brevity. In Section A, we explore different scenarios in which the UMP is obtained by only
allowing limited reweighting of the actual portfolio. Section B reports the results when using
the semi-variance as our measure of risk. Finally, Section C presents the results for different
sized windows for the analyses with book value data.

Appendix A Limited reweighting

Figure A.1 plots the turnover for both portfolios in the case where business models are
allowed to change with bank size. We observe that the turnover in the UMP decreases with the
level of risk aversion, but with an average value of 0.18 in the high risk aversion case it is still 35
percent higher than in the FED portfolio. Since this turnover is potentially costly to society, we
explore two scenarios in which the turnover is limited. First, we allow banks to grow/shrink by
no more than 20 percent of their initial weight. Second, we keep the largest 5 percent of banks
at their current cumulative size, allowing unlimited reweighting of the remaining banks while
still adhering to the no-shorting and no-loss constraint. Figures A.2-A.5 show the results for the
scenarios when banks assets or business models cannot change more than 20 percent, whereas
Figures A.6-A.9 show the results where the largest 5 percent of banks are kept equal to their
current value. However, regardless of the specification, the risk-return trade-off in the UMP is
practically unchanged compared to the FED portfolio in all time periods.

Figure A.1: How much more intervention would be required for the UMP?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion
2.0 2.0 2.0

1.5 1.5 1.5

Turnover

Turnover

Turnover

1.0 1.0 1.0

0.5 0.5 0.5

0.0 0.0 0.0

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

FED UMP FED UMP FED UMP

The figure shows the difference in reweighting between the FED and the UMP by plotting the turnover of
each portfolio. The levels of risk aversion are chosen based on the price of risk (Sharpe ratio) of the data,
and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk
aversion are λ = 200 and λ = 300, respectively.

41
 Figure A.2: What role does concentration play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.40 0.40 0.40

0.30 0.30 0.30

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)
0.10 0.10 0.10

42
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel A.2a shows the difference in risk,
return and concentration of each portfolio for low levels of risk aversion, while panels A.2b and A.2c show this for medium and high levels of risk aversion, respectively. The concentration in
each portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The levels of risk aversion are chosen based on the price
of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
 Figure A.3: What role does concentration play in the risk/return trade-off when using market values?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

1.00 1.00 1.00

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.20 0.20 0.20

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP FED UMP FED UMP FED

0.08 0.08 0.08

0.06 0.06 0.06

0.04 0.04 0.04

0.02 0.02 0.02

Return (in %)
Return (in %)
Return (in %)

43
0.00 0.00 0.00

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP 5% FED 5% UMP 5% FED 5% UMP 5% FED 5%

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using stock market value data. Panel A.3a shows the difference in
risk, return and concentration of each portfolio for low levels of risk aversion, while panels A.3b and A.3c show this for medium and high levels of risk aversion, respectively. The concentration
in each portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The levels of risk aversion are chosen based on the
price of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 10, while medium and high levels of risk aversion are λ = 50 and λ = 100, respectively.
 Figure A.4: What role does the business model of banks play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.20 0.20 0.20

0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.40 0.40 0.40

0.30 0.30 0.30

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)
0.10 0.10 0.10

44
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel A.4a shows the difference in risk,
return and business model (α) of each portfolio for low levels of risk aversion, while panels A.4b and A.4c show this for medium and high levels of risk aversion, respectively. The expected
utility of the supervisor is maximized by only changing the business models of banks while keeping the weight equal to the actual weight. The business model of a bank is a weight, α, on
the proportion of loans that a bank has, while (1 − α) is the proportion of non-loan (investment) related activities. The banks are assumed to keep the same level of profitability on each of
these business lines as α changes. Here, the weighted average α is shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion are chosen based on the price of
risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
 Figure A.5: What role does concentration play in the risk/return trade-off when business models change with bank size?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion
0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.40 0.40 0.40

0.30 0.30 0.30

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)
0.10 0.10 0.10

0.00 0.00 0.00

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

45
60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP 5% FED 5% UMP 5% FED 5% UMP 5% FED 5%

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel A.5a shows the difference in risk,
return, concentration and business model (α) of each portfolio for low levels of risk aversion, while panels A.5b and A.5c show this for medium and high levels of risk aversion, respectively.
The expected utility of the supervisor is maximized by changing the weights of banks. However, as banks are made smaller or larger, the business model of the banks (α) change in the
following way: as banks receive a 1 percent increase in weight, the proportion of loans that a bank has decreases by 0.45 percent. The profitability on both business lines (loans and
investments) is assumed to stay the same as α changes. Due to the change in proportion α, the total income of the bank changes and with it the covariance matrix as well. Here, the outcome
and concentration and the weighted average α are shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion are chosen based on the price of risk (Sharpe
ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
 Figure A.6: What role does concentration play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.35 0.35 0.35

0.30 0.30 0.30

0.25 0.25 0.25

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)

46
0.15 0.15 0.15

0.10 0.10 0.10

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel A.6a shows the difference in risk,
return and concentration of each portfolio for low levels of risk aversion, while panels A.6b and A.6c show this for medium and high levels of risk aversion, respectively. The concentration in
each portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The levels of risk aversion are chosen based on the price
of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
 Figure A.7: What role does concentration play in the risk/return trade-off when using market values?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

1.00 1.00 1.00

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.20 0.20 0.20

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP FED UMP FED UMP FED

0.25 0.25 0.25

0.20 0.20 0.20

0.15 0.15 0.15

0.10 0.10 0.10

Return (in %)
Return (in %)
Return (in %)

47
0.05 0.05 0.05

0.00 0.00 0.00

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP 5% FED 5% UMP 5% FED 5% UMP 5% FED 5%

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using stock market value data. Panel A.7a shows the difference in
risk, return and concentration of each portfolio for low levels of risk aversion, while panels A.7b and A.7c show this for medium and high levels of risk aversion, respectively. The concentration
in each portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The levels of risk aversion are chosen based on the
price of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 10, while medium and high levels of risk aversion are λ = 50 and λ = 100, respectively.
 Figure A.8: What role does the business model of banks play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.20 0.20 0.20

0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.40 0.40 0.40

0.30 0.30 0.30

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)
0.10 0.10 0.10

48
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel A.8a shows the difference in risk,
return and business model (α) of each portfolio for low levels of risk aversion, while panels A.8b and A.8c show this for medium and high levels of risk aversion, respectively. The expected
utility of the supervisor is maximized by only changing the business models of banks while keeping the weight equal to the actual weight. The business model of a bank is a weight, α, on
the proportion of loans that a bank has, while (1 − α) is the proportion of non-loan (investment) related activities. The banks are assumed to keep the same level of profitability on each of
these business lines as α changes. Here, the weighted average α is shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion are chosen based on the price of
risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
 Figure A.9: What role does concentration play in the risk/return trade-off when business models change with bank size?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion
0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.40 0.40
0.40

0.30 0.30
0.30

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)
0.10 0.10 0.10

0.00 0.00 0.00

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

49
60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP 5% FED 5% UMP 5% FED 5% UMP 5% FED 5%

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel A.9a shows the difference in risk,
return, concentration and business model (α) of each portfolio for low levels of risk aversion, while panels A.9b and A.9c show this for medium and high levels of risk aversion, respectively.
The expected utility of the supervisor is maximized by changing the weights of banks. However, as banks are made smaller or larger, the business model of the banks (α) change in the
following way: as banks receive a 1 percent increase in weight, the proportion of loans that a bank has decreases by 0.45 percent. The profitability on both business lines (loans and
investments) is assumed to stay the same as α changes. Due to the change in proportion α, the total income of the bank changes and with it the covariance matrix as well. Here, the outcome
and concentration and the weighted average α are shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion are chosen based on the price of risk (Sharpe
ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
Appendix B Semi-variance as measure of risk
In this section we examine our definition of risk. In the baseline experiment, the risk is
calculated on all data (gains and losses). We run extra analyses where the relevant risk is based
on downside risk. To do so, we define the semi-variance as the variance of the set of returns
that is below its mean, and semi-covariances in a similar manner. While the covariance matrix
is not symmetric anymore, potentially failing to be positive semi-definite, we find no instances
of negative portfolio variances.
We do however adjust the size of the risk aversion coefficients, as the Sharpe ratios based
on the semi-variance are much higher than in the baseline experiment. For the market returns,
the Sharpe ratio is now roughly 10 times higher, whereas for the book returns it is roughly 2.5
times compared to the baseline values. We adjust λ correspondingly, and use values of 100, 200
and 300 for the market values, and 100, 500, 1000 for the book values indicating low, medium
and high risk aversion, respectively.
Figures B.1-B.4 show the results, which are quantitatively and qualitatively similar to the
baseline experiment.

50
 Figure B.1: What role does concentration play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.10 0.10 0.10

0.08 0.08 0.08

0.06 0.06 0.06

0.04 0.04 0.04

0.02 0.02 0.02

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Return (in %)
Return (in %)
Return (in %)
0.20 0.20 0.20

51
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel B.1a shows the difference in risk,
return and concentration of each portfolio for low levels of risk aversion, while panels B.1b and B.1c show this for medium and high levels of risk aversion, respectively. The concentration in
each portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The levels of risk aversion are chosen based on the price
of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 500 and λ = 1000, respectively.
 Figure B.2: What role does concentration play in the risk/return trade-off when using market values?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.30 0.30 0.30

0.20 0.20 0.20

0.10 0.10 0.10

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP FED UMP FED UMP FED

0.40 0.40 0.40

0.30 0.30 0.30

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)
0.10 0.10 0.10

52
0.00 0.00 0.00
1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using stock market data. Panel B.2a shows the difference in risk,
return and concentration of each portfolio for low levels of risk aversion, while panels B.2b and B.2c show this for medium and high levels of risk aversion, respectively. The concentration in
each portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The levels of risk aversion are chosen based on the price
of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.
 Figure B.3: What role does the business model of banks play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

1.00 1.00 1.00

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Return (in %)
Return (in %)
Return (in %)
0.20 0.20 0.20

53
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel B.3a shows the difference in risk,
return and business model (α) of each portfolio for low levels of risk aversion, while panels B.3b and B.3c show this for medium and high levels of risk aversion, respectively. The expected
utility of the supervisor is maximized by only changing the business models of banks while keeping the weight equal to the actual weight. The business model of a bank is a weight, α, on
the proportion of loans that a bank has, while (1 − α) is the proportion of non-loan (investment) related activities. The banks are assumed to keep the same level of profitability on each of
these business lines as α changes. Here, the weighted average α is shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion are chosen based on the price of
risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 500 and λ = 1000, respectively.
 Figure B.4: What role does concentration play in the risk/return trade-off when business models change with bank size?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion
0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1987q1 1993q3 2000q1 2006q3 2013q1

UMP FED UMP FED UMP FED

1.00 1.00 1.00

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Return (in %)
Return (in %)
Return (in %)
0.20 0.20 0.20

0.00 0.00 0.00

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1987q1 1993q3 2000q1 2006q3 2013q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

54
60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1987q1 1993q3 2000q1 2006q3 2013q1

UMP 5% FED 5% UMP 5% FED 5% UMP 5% FED 5%

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1987q1 1993q3 2000q1 2006q3 2013q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data. Panel B.4a shows the difference in risk,
return, concentration and business model (α) of each portfolio for low levels of risk aversion, while panels B.4b and B.4c show this for medium and high levels of risk aversion, respectively. The
expected utility of the supervisor is maximized by changing the weights of banks. However, as banks are made smaller or larger, the business model of the banks (α) change in the following
way: as banks receive a 1 percent increase in weight, the proportion of loans that a bank has decreases by 0.45 percent. The profitability on both business lines (loans and investments)
is assumed to stay the same as α changes. Due to the change in proportion α, the total income of the bank changes and with it the covariance matrix as well. Here, the outcome and
concentration and the weighted average α are shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion are chosen based on the price of risk (Sharpe ratio) of
the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 500 and λ = 1000, respectively.
Appendix C Different estimation windows
Finally, we explore if the choice of an 8 quarter window on which the covariance matrices
are estimated matters for the optimization. In order to see if our results are robust, we run the
analysis using two different windows. First, to match the time frames in the analysis on market
value data, we perform the analysis on sets of all 4 quarters in a given year. Second, we increase
the original window size to 16 quarters in order to eliminate the impact of outliers.
The results for the 4 quarter window are shown in Figures C.1-C.3, whereas the results for
the 16 quarter window are shown in Figures C.4-C.6. Again, the results are robust to using
different windows sizes.

55
 Figure C.1: What role does concentration play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion
0.12 0.12 0.12

0.10 0.10 0.10

0.08 0.08 0.08

0.06 0.06 0.06

0.04 0.04 0.04

0.02 0.02 0.02

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Return (in %)
Return (in %)
Return (in %)
0.20 0.20 0.20

56
0.00 0.00 0.00
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1985q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data with a yearly, 4 quarter non-overlapping,
window. Panel C.1a shows the difference in risk, return and concentration of each portfolio for low levels of risk aversion, while panels C.1b and C.1c show this for medium and high levels
of risk aversion, respectively. The concentration in each portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The
levels of risk aversion are chosen based on the price of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk
aversion are λ = 200 and λ = 300, respectively.
 Figure C.2: What role does the business model of banks play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.15 0.15

0.15
0.10 0.10

0.10

0.05 0.05
0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP FED UMP FED UMP FED

0.60 0.60 0.60

0.40 0.40 0.40

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)

57
0.00 0.00 0.00
1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP FED UMP FED UMP FED

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data with a yearly, 4 quarter non-overlapping,
window. Panel C.2a shows the difference in risk, return and business model (α) of each portfolio for low levels of risk aversion, while panels C.2b and C.2c show this for medium and high
levels of risk aversion, respectively. The expected utility of the supervisor is maximized by only changing the business models of banks while keeping the weight equal to the actual weight.
The business model of a bank is a weight, α, on the proportion of loans that a bank has, while (1 − α) is the proportion of non-loan (investment) related activities. The banks are assumed to
keep the same level of profitability on each of these business lines as α changes. Here, the weighted average α is shown for both the FED portfolio and the hypothetical UMP. The levels of
risk aversion are chosen based on the price of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion
are λ = 200 and λ = 300, respectively.
 Figure C.3: What role does concentration play in the risk/return trade-off when business models change with bank size?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP FED UMP FED UMP FED

1.00 1.00 1.00

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Return (in %)
Return (in %)
Return (in %)
0.20 0.20 0.20

0.00 0.00 0.00

1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP FED UMP FED UMP FED

100 100 100

80 80 80

58
60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP 5% FED 5% UMP 5% FED 5% UMP 5% FED 5%

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data with a yearly, 4 quarter non-overlapping,
window. Panel C.3a shows the difference in risk, return, concentration and business model (α) of each portfolio for low levels of risk aversion, while panels C.3b and C.3c show this for
medium and high levels of risk aversion, respectively. The expected utility of the supervisor is maximized by changing the weights of banks. However, as banks are made smaller or larger,
the business model of the banks (α) change in the following way: as banks receive a 1 percent increase in weight, the proportion of loans that a bank has decreases by 0.45 percent. The
profitability on both business lines (loans and investments) is assumed to stay the same as α changes. Due to the change in proportion α, the total income of the bank changes and with it
the covariance matrix as well. Here, the outcome and concentration and the weighted average α are shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion
are chosen based on the price of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200
and λ = 300, respectively.
 Figure C.4: What role does concentration play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.10 0.10 0.10

0.08 0.08 0.08

0.06 0.06 0.06

0.04 0.04 0.04

0.02 0.02 0.02

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP FED UMP FED UMP FED

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Return (in %)
Return (in %)
Return (in %)
0.20 0.20 0.20

59
0.00 0.00 0.00
1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP FED UMP FED UMP FED

100 100 100

80 80 80

60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3 1987q3 1994q1 2000q3 2007q1 2013q3

UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP UMP 5% FED 5% FED 5% in UMP

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data with a 16 quarter rolling window. Panel C.4a
shows the difference in risk, return and concentration of each portfolio for low levels of risk aversion, while panels C.4b and C.4c show this for medium and high levels of risk aversion,
respectively. The concentration in each portfolio is shown by plotting their concentration ratios, as well as the weights that the current largest banks have in the UMP. The levels of risk
aversion are chosen based on the price of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are
λ = 200 and λ = 300, respectively.
 Figure C.5: What role does the business model of banks play in the risk/return trade-off?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion

0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

0.60 0.60 0.60

0.40 0.40 0.40

0.20 0.20 0.20

Return (in %)
Return (in %)
Return (in %)

60
0.00 0.00 0.00
1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data with a 16 quarter rolling window. Panel C.5a
shows the difference in risk, return and business model (α) of each portfolio for low levels of risk aversion, while panels C.5b and C.5c show this for medium and high levels of risk aversion,
respectively. The expected utility of the supervisor is maximized by only changing the business models of banks while keeping the weight equal to the actual weight. The business model
of a bank is a weight, α, on the proportion of loans that a bank has, while (1 − α) is the proportion of non-loan (investment) related activities. The banks are assumed to keep the same
level of profitability on each of these business lines as α changes. Here, the weighted average α is shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion
are chosen based on the price of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200
and λ = 300, respectively.
 Figure C.6: What role does concentration play in the risk/return trade-off when business models change with bank size?
(a) Low risk aversion (b) Medium risk aversion (c) High risk aversion
0.20 0.20 0.20

0.15 0.15 0.15

0.10 0.10 0.10

0.05 0.05 0.05

Standard Deviation (in %)

Standard Deviation (in %)
Standard Deviation (in %)
0.00 0.00 0.00
1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

1.00 1.00 1.00

0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

Return (in %)
Return (in %)
Return (in %)
0.20 0.20 0.20

0.00 0.00 0.00

1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

100 100 100

80 80 80

61
60 60 60

40 40 40

Weight (in %)
Weight (in %)
Weight (in %)

20 20 20

0 0 0
1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP 5% FED 5% UMP 5% FED 5% UMP 5% FED 5%

1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

Portfolio Weighted α
Portfolio Weighted α
Portfolio Weighted α

0.0 0.0 0.0

1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1 1990q1 1995q1 2000q1 2005q1 2010q1 2015q1

UMP FED UMP FED UMP FED

The figure shows the comparison between the FED portfolio and the hypothetical UMP for different levels of risk aversion using book value data with a 16 quarter rolling window. Panel C.6a
shows the difference in risk, return, concentration and business model (α) of each portfolio for low levels of risk aversion, while panels C.6b and C.6c show this for medium and high levels of
risk aversion, respectively. The expected utility of the supervisor is maximized by changing the weights of banks. However, as banks are made smaller or larger, the business model of the
banks (α) change in the following way: as banks receive a 1 percent increase in weight, the proportion of loans that a bank has decreases by 0.45 percent. The profitability on both business
lines (loans and investments) is assumed to stay the same as α changes. Due to the change in proportion α, the total income of the bank changes and with it the covariance matrix as well.
Here, the outcome and concentration and the weighted average α are shown for both the FED portfolio and the hypothetical UMP. The levels of risk aversion are chosen based on the price
of risk (Sharpe ratio) of the data, and in this case low level of risk aversion is proxied by a λ of 100, while medium and high levels of risk aversion are λ = 200 and λ = 300, respectively.