Portfolio Similarity and Asset Liquidation
in the Insurance Industry∗
Mila Getmansky†
Giulio Girardi‡
Stanislava Nikolova¶
Kathleen W. Hanley§
Loriana Pelizzonk
This Draft: March 16, 2016
Abstract
The designation of insurance companies as Systemically Important Financial Institutions
(SIFI) explicitly considers the potential for the interconnectedness of their portfolios to affect the asset liquidation channel of systemic risk transmission. In this paper, we study
the determinants and effects of interconnectedness on the selling behavior of insurers using
cosine similarity. We document that greater portfolio similarity between two insurers is significantly related to the similarity in insurers’ asset liquidation decisions. This relationship
is stronger when both insurers are capital constrained but only at the asset class level suggesting that correlated re-balancing to improve capital ratios occurs at a higher level than
individual securities. We find that the relationship between the portfolio holdings of both
illiquid securities and downgraded issuers and sales similarity is greater during the financial
crisis. In addition, we show that portfolio similarity cannot be explained by easily observable characteristics such as return correlation, concentration or assets under management.
We interpret our findings to mean that the co-investment decisions of insurers under many
different economic conditions are an important determinant of selling behavior. Overall, our
paper provides an implementable mechanism to identify and monitor the interconnectedness
of insurance company portfolios.
Keywords: Interconnectedness, Asset liquidation, Similarity, Systemic Risk, Financial Stability, Insurance Companies, SIFI
∗
We thank Mark Flannery and participants at the Temple University Workshop on Systemic Risk and
the Insurance Industry and MIT CSRA Meeting for helpful comments. We thank Max Riedel and Matteo
Sottocornola for excellent research support. The Securities and Exchange Commission, as a matter of policy,
disclaims responsibility for any private publication or statement by any of its employees. The views expressed
herein are those of the author and do not necessarily reflect the views of the Commission or of the authors
colleagues on the staff of the Commission.
†
Isenberg School of Management, University of Massachusetts, Amherst, MA 01003, Email: msherman@isenberg.umass.edu.
‡
Division of Economic and Risk Analysis, U.S. Securities and Exchange Commission, Washington DC
20549-9040, Email: girardig@sec.gov.
§
College of Business and Economics, Lehigh University, Email: kwh315@lehigh.edu.
¶
College of Business Administration, University of Nebraska–Lincoln, Lincoln, NE 68588, Email:
snikolova2@unl.edu.
k
Goethe University Frankfurt - Center of Excellence SAFE and Ca’ Foscari University of Venice, Department of Economics, 30100 Venice, Email: pelizzon@unive.it.
“The severity of the disruption caused by a forced liquidation of Prudentials assets could
be amplified by the fact that the investment portfolios of many large insurance companies
are composed of similar assets, which could cause significant reductions in asset valuations
and losses for those firms. The erosion of capital and potential de-leveraging could result in
asset fire sales that cause significant damage to the broader economy.” (emphasis added)
Basis for the Financial Stability Oversight Council’s Final Determination
Regarding Prudential Financial, Inc.
1
Introduction
The recent global financial crisis of 2007-2009 exposed many vulnerabilities in the fi-
nancial system. One of these was the implicit and explicit interconnectedness of financial
institutions, which led to the collapse of Lehman Brothers, Bear Stearns, Washington Mutual, Wachovia, and AIG, among others. In addition, the events of this period precipitated
huge losses in several financial markets (e.g. stocks, credit default swaps, subprime mortgage,
and money markets). Governments in the United States, United Kingdom, and several other
Western European countries were forced to intervene in order to prevent further declines.
Following the global financial crisis, the U.S. Congress passed the Dodd-Frank Wall Street
Reform and Consumer Protection Act of 2010 (Dodd-Frank Act). The Act created the Financial Stability Oversight Council (FSOC) and endowed the Council with the authority to
designate bank and nonbank Systemically Important Financial Institutions (SIFIs). Companies that are designated as SIFIs are subject to enhanced prudential standards with the
goal of limiting the effect of a SIFI’s distress on financial stability. In designating nonbank
financial institutions, such as insurance companies, asset management companies, savings
and loan holding companies, broker-dealers, and other financial firms as SIFIs, a variety
of factors have been considered by regulators with size and interconnectedness being two
criteria.1
1
As noted in the final rule on the Authority to Require Supervision and Regulation of Certain Nonbank
Financial Companies, “Section 113 of the Dodd-Frank Wall Street Reform and Consumer Protection Act
(Pub. L. 111203, 124 Stat. 1376 (2010).) authorizes the Financial Stability Oversight Council to determine
that a nonbank financial company shall be supervised by the Board of Governors of the Federal Reserve
System and shall be subject to prudential standards... if the Council determines that material financial
distress at the nonbank financial company, or the nature, scope, size, scale, concentration, interconnectedness,
or mix of the activities of the nonbank financial company, could pose a threat to the financial stability of the
1
Although it is relatively straight-forward to measure the size of a financial institution
(global assets, or for a foreign institution, U.S. total consolidated assets), there is no consensus on how to measure interconnectedness. Interconnectedness among financial institutions
can contribute to systemic risk in a variety of ways on both the asset and liability side
of the balance sheet. Prior research that has focused on the contribution of insurers to
systemic risk generally examines interconnectedness through operational risks, reinsurance,
non-traditional investments, and financing. Harrington (2009) suggests that the systemic
risk of property and casualty (P&C) insurers is low, while that of life insurers could be high
because of higher leverage and potential policyholder withdrawals during a financial crisis.
Cummins and Weiss (2014) conclude that the core activities of both life and P&C insurers
are not a source of systemic risk, but the non-core activities of life insurers could make them
systemically relevant. Reinsurance can also be a source of interconnectedness (Cummins and
Weiss (2014) and Park and Xie (2014)), as can be insurers’ increased exposure to derivatives
(Geneva Association (2010) and Grace (2010)) and increased reliance on short-term funding
(Geneva Association (2010)).
The Council defines the asset liquidation channel as occurring when a “nonbank financial company holds assets that, if liquidated quickly, would cause a fall in asset prices and
thereby significantly disrupt trading or funding in key markets or cause significant losses or
funding problems for other firms with similar holdings.”2 In this paper, we focus on the
asset side of the balance sheet and examine whether the interconnectedness among the investment portfolios of insurers could contribute to systemic risk through the asset liquidation
channel. We contribute to the literature in two ways. First, we outline a methodology that
captures interconnectedness between pairs of insurers. Second, we show how our measure of
interconnectedness is related to the asset liquidation channel.
United States.” Similar characteristics are used internationally by the Financial Stability Board to designate
globally systemically important institutions (G-SIFIs), with size and interconnectedness being two criteria
((BIS (2014))).
2
See Basis for the Financial Stability Oversight Council’s Final Determination Regarding Prudential
Financial, Inc. available on the FSOC website (FSOC (2013)).
2
Sales of securities can be triggered by asset valuation shocks (Brunnermeier and Pedersen
(2005)), capital constraints (Ellul, Jotikasthira, and Lundblad (2011), Danielsson, Shin, and
Zigrand (2004) and Brunnermeier and Pedersen (2009)), and sale of collateral (Shleifer and
Vishny (2011)). The price impact of portfolio re-balancing is exacerbated when the securities sold are illiquid (Brunnermeier and Pedersen (2009) and Cont and Wagalath (2015)).
According to Kartasheva (2014), insurers do not need to fail to propagate systemic risk; it
may be sufficient for them to “fire sell” assets to produce a significant systemic impact.
The transmission of systemic risk through the asset liquidation channel is likely to occur if
the portfolios of insurance companies are similar. When faced with a financial shock to either
assets or liabilities, the assumption is that insurers will re-balance their portfolios in the same
fashion thereby transmitting systemic risk throughout the economy. For example, Ben S.
Bernanke (Monitoring the Financial System, 2013) states that “Examples of vulnerabilities
include interconnectedness, and complexity, all of which have the potential to magnify shocks
to the financial system. Absent vulnerabilities, triggers [such as losses on mortgage holdings]
would generally not lead to full-blown financial crises.” However, little research has been
conducted to examine the similarity of insurers’ portfolios and whether it affects their selling
behavior.
We propose a measure of interconnectedness among insurers based on the cosine similarity
of their portfolios and selling behavior. The National Association of Insurance Commissioners
(NAIC) requires that insurance companies disclose their portfolio holdings and transactions
at the individual security level. Therefore, we are able to determine both the yearly portfolio
composition and re-balancing through time. We demonstrate that our interconnectedness
measure can predict asset liquidation at both the security issuer and at the asset class level.
Using the sample of all insurers in Schedule D, we show that their asset allocation decisions can be characterized by only a few different strategies. Cluster analysis reveals that
insurer allocation strategies can be divided into three distinct strategies based upon the primary investment by the insurer: 1) municipal bonds and equity, 2) corporate bonds, and
3
3) U.S. government debt. We find that broad asset allocation strategies among insurance
companies tend to be long-term in nature and do not vary much over the business cycle.3
Although broad asset class allocation strategies may remain similar through time, the
actual composition of the portfolios may exhibit substantial variation either because the
weights of the asset classes or the choice of individual securities within the strategy differ.
Restricting our sample to approximately 99 publicly-traded insurers that comprise around
70% of the book value of securities held by the insurance industry, we first construct a vector
of portfolio weights at the asset class or issuer (6 digit CUSIP) level for each insurer. We then
use cosine similarity to measure how similar are the portfolios of a pair of insurers in terms of
asset classes or issuers.4 One benefit of cosine similarity is that it has an easy interpretation
because it is bounded between zero and one. The more (less) similar are the portfolios of
two insurers, the closer the cosine similarity is to one (zero). Our sample consists of over
41,000 insurer pair-years from 2002 to 2014.
We examine the determinants of pairwise portfolio similarity using a multivariate approach that controls for characteristics of the two insurers and their portfolios. We show
that for both asset class and issuer level similarity, two insurance companies have generally
more similar portfolios if they are both in the same line of business (life or P&C).
Insurers are also more similar if their portfolios are less, not more, concentrated, as
measured by a Herfindahl index suggesting that diversification increases portfolio similarity.
This is consistent with a growing literature (Allen, Babus, and Carletti (2012), Castiglionesi
and Navarro (2008), Ibragimov, Jaffee, and Walden (2011), and Wagner (2010)) that argues
that diversification in the banking industry can lead to greater systemic risk because it leads
banks to invest in similar assets. In support of the adverse effects of diversification, Wagner
(2009) argues that asset allocation in banks is not efficient and proposes that the optimal
3
Cluster analysis is one metric with which to examine systemic risk. Allen, Babus, and Carletti (2012)
propose a model in which asset commonality and short-term debt of banks interact to generate excessive
systemic risk. They show that in a clustered structure, groups of institutions that hold similar assets default
together. Blei and Ergashev (2014) use cluster analysis to construct ACRISK, a measure of systemic risk
based on commonalities in bank’s asset holdings that captures the buildup of systemic risk.
4
Sias, Turtle, and Zykaj (2015) analyze hedge funds’ trading behavior also using cosine similarity.
4
regulatory treatment of banks is to encourage more correlation among bank asset holdings
at already highly correlated banks and lowering the correlation at other banks.
At the asset class level we find that both pairs of potential SIFIs (PSIFIs) and pairs
of non-PSIFIs have greater portfolio similarity although the coefficient for PSIFI pairs is
only marginally significant. (We define a PSIFI as an insurer that either has $50 billion in
consolidated assets in at least one year of our sample period.)5 At the issuer level, portfolio
similarity continues to be greater when the pair of insurers are both non-PSIFI, but we do not
have any evidence that PSIFI pairs are more likely to have higher portfolio similarities. In
addition, the average level of portfolio similarity between pairs of PSIFIs is generally lower
than the average portfolio similarity between pairs of non-PSIFIs throughout the sample
period. Thus, we do not find strong evidence for the assumption that the portfolios of
PSIFIs are more likely to be composed of similar assets than other pairs of insurers.
Next we examine whether public market information such as return correlation is related
to portfolio similarity. A number of papers have proposed return correlation as a measure
of interconnectedness (Billio, Getmansky, Lo, and Pelizzon (2012), Neale, Drake, Schorno,
and Semann (2012), and Brunetti, Harris, Mankad, and Michailidis (2015)). Including return
correlation in the model of pairwise determinants of portfolio similarity does not substantially
improve the model. For asset class similarity, the return correlation coefficient is insignificant
and for issuer similarity it is significantly negative with little improvement in the R2 in any
specification. These findings suggest that return correlation may not be an appropriate
metric to measure the effect of interconnectedness on the transmission of risk through the
asset liquidation channel.
In order for our measure to be useful, it must predict whether insurers are selling similar
assets. When we examine the determinants of quarterly pairwise sales similarity, we do not
find a strong association between business lines and sales similarity. Unlike the determinants
of portfolio similarity, we show that PSIFI pairs are more likely to have similar sales behavior
5
The methodology for determining global systemically important insurers is also presented in (IAIS
(2013))and (IAIS (2015)).
5
at both the asset class and issuer level but non-PSIFIs are not. Furthermore, the sales
similarity between two insurers is stronger if both of them have concentrated portfolios.
Despite the fact that we do not find that PSIFIs are more likely to hold similar portfolios,
our results do provide confirmation that PSIFIs are more likely to make similar selling
decisions when their portfolios are similar.
In order to examine the potential for interconnectedness in portfolio holdings to influence
the asset liquidation channel, we analyze whether portfolio similarity captures commonality
in asset re-balancing. At either the asset class or issuer levels, we find that portfolio similarities at prior year-end are significantly related to quarterly sales similarity regardless of
whether both insurers are PSIFIs or non-PSIFIs. As with portfolio similarity, we present
evidence that return correlations cannot substitute for our measure of interconnectedness.
Other factors may influence selling behavior, most notably, capital regulation. The literature has shown that capital-constrained insurers attempt to improve their regulatory capital
standing primarily by re-balancing their investment portfolios (Ellul, Jotikasthira, and Lundblad (2011), Ambrose, Cai, and Helwege (2008), Manconi, Massa, and Yasuda (2012)), and
these activities can lead to fire sale prices (Merrill, Nadauld, Stulz, and Sherlund (2013)).
We examine whether our findings are driven by both insurers in a pair having low risk-based
capital (RBC) ratios. To do so, we interact a dummy variable, equal to one if both insurers
have low RBC ratios, with our two measures of portfolio similarity. The results indicate that
insurers with low RBC ratios and similar portfolios are more likely to sell the same asset
class. At the issuer level, however, low RBC ratios do not appear to affect sales similarity.
We interpret this finding to mean that insurers with similar portfolios re-balance within the
same asset class but not necessarily by selling the same set of issuers.
In addition, we analyze whether the relationship between portfolio similarity and sales
similarity can be captured by the similarity in holdings of illiquid or liquid securities or
the similarity of downgraded and non-downgraded bonds. Given a broad literature on the
potential for regulated entities to sell liquid securities in times of crisis and to sell downgraded
6
securities in order to avoid an increase in capital, it is possible that the relationship we
document is driven by insurers’ portfolio re-balancing in these types of securities and not
more generally (or vice versa). We decompose insurers’ portfolio similarity into liquid/illiquid
by asset classes and by whether or not an issuer was downgraded in the following year. We
find that the portfolio similarity of all of these subsets predicts sales similarity and therefore,
our results are not simply due to a specific group of securities.
Finally, we analyze whether our findings change in the period surrounding the financial
crisis. We define three dummy variables, equal to one if the time period falls within 2002-2006
(Pre-Crisis), 2007-2009 (Crisis) and 2010-2014 (Post-Crisis), zero otherwise, and interact
them with the portfolio similarities based on liquidity and downgrades. We find that the
relationship between the portfolio similarity of illiquid and downgraded issuers and sales
similarity is stronger during the financial crisis indicating that the co-investment decisions
of these assets are more likely affect selling behavior in times of market stress.
Our analysis complements recent work on corporate bond herding among insurers. Chiang and Niehaus (2016) examine the herding behavior of life insurers and find that buy-side
herding is more pronounced when insurers are part of groups that have been designated as
SIFIs and the bond has been downgraded. Comparing mutual funds, pension funds and
insurers, Fang Cai and Li (2016) also find evidence of herding by insurance companies in
downgraded bonds. Our methodology provides two potential benefits over other measures of
herding. First, we can measure interconnectedness a over a wide variety of asset classes, not
just corporate bonds. This means that we allow insurers to re-balance the portfolio using
any asset class and issuer. Second, our methodology allows for insurers to make the decision
not to sell any particular securities. This is because sales similarity is based on changes in
portfolio holdings that may, in some cases, be zero.
Overall, our results indicate that the interconnectedness of insurers’ portfolios, as measured by cosine similarity, captures attributes of interconnectedness that other simple measures such as portfolio, return correlation, concentration or size do not. Our methodology can
7
be applied more broadly to assess the level of interconnectedness of specific activities among
entities that may contribute to systemic risk. For example, one could measure whether
similarity in mortgage portfolios leads to correlated losses or lending behavior. More importantly, it also captures commonality in asset liquidation across insurers making our measure
relevant to regulators who are tasked with monitoring systemic risk in the economy.
The remainder of the paper is organized as follows. In Section 2 we present how our sample and variables are constructed as well as summary statistics. In Section 3 we describe the
composition of insurers’ portfolios using cluster analysis. In Section 4 we define our pairwise
similarity measure and investigate its determinants. Section 5 presents our examination of
the relationship between portfolio similarity and sales similarity. We examine how capital
constraints, liquidity, downgrades and the financial crisis affect our findings in Section 6. We
conclude in Section 7.
2
Data
We analyze the interconnectedness of insurance companies from 2002 to 2014 using infor-
mation from their statutory filings with the NAIC as distributed by A.M. Best. Specifically,
Parts 1 and 2 of Schedule D of these filings report for each insurer the par value and book
value of each security held at calendar year-end. Our holdings data includes all non-negative
reported positions and is constructed at an annual frequency. Disposals of securities are
reported in Parts 4 and 5 of Schedule D. The data includes, for each insurer, every security
disposed of during the year along with its par value, disposal value, and date of disposal. We
exclude security disposals due to maturity, repayment, calls, or other non-trading activity,
and construct selling transactions at a quarterly frequency.
Both portfolio holdings and sale transactions are reported at the individual security (9digit CUSIP) level. For each insurer, we aggregate this information to the asset class or to the
security issuer level.6 In order aggregate at the asset class level, we categorize each security
6
When calculating portfolio weights, we use par value for fixed-income security holdings. Since no
8
into one of the following asset classes: (1) U.S. government debt, (2) GSE debt (including
mortgage-backed securities), (3) municipal debt, (4) sovereign debt, (5) corporate debt, (6)
private-label RMBS, (7) private-label CMBS, (8) private-label ABS, (9) equity (common
and preferred stock), and (10) mutual fund shares. We do so using information from several
data sources. We identify RMBS and CMBS using the NAIC-provided list of PIMCO- and
BlackRock-modeled securities.7 We classify all remaining fixed-income securities using (1)
the sector and subsector codes in S&P RatingXpress, then (2) the type and subtype codes in
DataScope, then (3) the issue description and issuer name in NAIC Schedule D, and finally
(4) the issuer name and collateral asset type in SDC Platinum’s New Issues Module. If we
are unable to categorize a security using this algorithm, we assign it to the asset class its
holder reports in Schedule D.
We also aggregate holdings and sales at the security issuer level. We use the first 6 digits
of each CUSIP to identify securities issued by the same entity.8 We classify every security
with the same 6 digit CUSIP as having the same issuer.
We conduct most of our analysis at the holding company rather than the group or individual insurer level. The predominant organizational model in the insurance industry is the
insurance group, with a single holding company owning several insurance groups. Although
in many ways individual companies operate independently, some aspects of their operations
are centrally managed thus creating strong connections among the members of a group and
the subsidiaries of a holding company. By focusing on holding companies rather than on
individual companies or insurer groups, we are also able to compute equity market measures
of interconnectedness used in the existing literature that are available only at the publicly
traded holding company level.
comparable number exists for equity securities, we aggregate equity using book value.
7
The NAIC changed its capital assessment methodology for certain asset classes by replacing credit
ratings as the measure of expected loss with expected loss estimates from PIMCO for RMBS and BlackRock
for CMBS. The NAIC publishes the list of PIMCO- and BlackRock-modeled securities annually. For more
information on this change, see Hanley and Nikolova (2015).
8
The use of the 6-digit CUSIP only approximates the ultimate issuer of the securities as a parent company
may have different 6-digit subsidiary CUSIPs.
9
The number of individual insurance companies in the A.M.Best Schedule D data ranges
from 3,568 to 4,270 depending on the calendar year. They are organized in 1,194 to 1,727
insurance groups with an average of 2.8 individual insurers in a group. To aggregate Schedule
D data to the holding company level, we match insurer groups to firms in CRSP/Compustat
Merged manually by name. Of the groups for which we have holdings/trade data, between
73 and 99 (depending on the calendar year) have CRSP and Compustat data. For all of the
analysis except the determination of portfolio strategies using cluster analysis, we restrict
our sample to only these publicly traded insurance groups. Nonetheless, this subset accounts
for 68-76% of the book value of Schedule D holdings reported by the insurance industry. For
each of these insurers, we collect daily holding period returns from CRSP, and financial
statement information from Compustat. Appendix A provides detailed definitions of the
variables used in our analysis.
In addition, we perform some of our analysis separately for subsets of insurers. We classify
insurers as P&C, life, or other (e.g. health, fraternal, and title) by whether at least half of
the insurers’ portfolio assets are held in a given year by companies in the group that are in
that line of business. In order to examine whether systemically important insurers are more
likely to have similar portfolios, we classify insurers as Potentially Systemically Important
Institutions (or PSIFIs) if they have more than $50 billion in total assets in at least one year
of the sample. Based on this size threshold, we identify 34 insurers as potential candidates
for SIFI designation by the FSOC.
2.1
Sample Characteristics
Table 1 presents descriptive statistics on insurer portfolio composition by issuer and asset
class for our sample of publicly traded insurers. The table shows that the average insurer
in our sample holds 2,937 different securities issued by 1,475 different issuers. The median
number of securities (issuers) held is less than half of the sample average, implying that
some insurers invest in significantly more securities than others. The average total assets
10
for a given insurer is $138.6 billion. By construction, PSIFIs hold significantly more assets
($398.3 billion) compared to non-PSIFIs ($11.5 billion). We also find that life insurers are
approximately twice as large as P&C insurers.
The table also presents insurers’ portfolio composition by asset class. We average each
insurer’s holdings through time and report the cross-sectional mean, median, and standard
deviation of the time-series averages. Consistent with the common perception that insurers
are important investors in fixed-income markets, we find that fixed-income securities make
up the majority of their holdings with more than 79% on average. Corporate bonds (41%),
GSE securities (12%) and municipal bonds (9%) represent the largest proportion of insurers’
fixed-income investments in our sample. Equity holdings of insurers are primarily in the
form of common and preferred stock, and these securities account for 17% of the portfolio
on average. The median insurer tends not to hold any mutual fund shares in its portfolio.
Finally, there appears to be significant cross-sectional variation in asset class holdings as
indicated by the standard deviation of each asset class’ proportion across insurers.
Figure 1 summarizes the time-series variation of our sample of insurers’ cross-sectional
average holdings and indicates that shifts in and out of asset classes occur through time.
Over our sample period, the portfolio allocated to equity decreases while holdings of US
government securities and mutual fund shares increase. The figure also shows that insurers’
holdings of RMBS and CMBS increase in the period leading up to the financial crisis and then
gradually decrease consistent with the evidence presented in Hanley and Nikolova (2015).
Table 1 also shows that some of this variation in holdings by asset class is related to
insurer type. Life insurers invest in more securities and issuers than P&C and their portfolios
are more heavily weighted toward corporate bonds, and less toward municipal bonds, GSE
securities and equity.
We also separately examine the portfolio composition of PSIFI and non-PSIFI insurers.
The statistics in Table 1 indicate that PSIFIs hold an average of 6,678 different securities
issued by 3,183 issuers compared to the non-PSIFI mean of 1,466 securities issued by 820
11
issuers. The size of PSIFIs’ portfolios compared to non-PSIFIs is related to the overall size
of the two groups of insurers. Both PSIFIs’ and non-PSIFI portfolios invest most heavily in
corporate bonds, PSIFIs are much more heavily weighted toward equity and mutual funds
(mainly affiliated funds).
We measure the level of portfolio concentration at both the asset class level and the issuer
level using a Herfindahl index. Specifically, asset-class portfolio concentration is:
Concentration ACi,t =
N
X
2
wi,n,t
(1)
n=1
where wi,n,t is asset-class n weight for an insurer i and is calculated as the dollar amount
invested in asset class n relative to the total value of the insurer i portfolio at the end of
year t. Similarly, issuer-level concentration is:
Concentration Ii,t =
N
X
2
wi,n,t
(2)
n=1
where wi,n,t is issuer n weight for an insurer i and is calculated as the dollar amount invested
in issuer n relative to the total value of the insurer i portfolio at the end of year t.
Table 1 reports the mean, median and standard deviation of insurers’ time-series averages
of the two concentration measures. The average issuer concentration in our sample is 0.06
and the average asset class concentration is 0.37. Issuer (asset class) concentration appears
to differ across types of insurers with an average of 0.06 (0.39) for Life and 0.03 (0.30)
for P&C insurers. Although life insurers invest in more issuers and securities than P&C
insurers, they hold more concentrated portfolios, as measured by either asset class or issuer
composition. Finally, PSIFIs’ portfolios do not seem more concentrated or diverse than those
of non-PSIFIs.
12
3
Portfolio Composition Using Cluster Analysis
In this section, we examine the portfolio strategies of insurers at the asset class level using
cluster analysis. We are interested in whether insurance companies differ in their portfolio
allocation strategies and whether their strategies change over time. In order to examine the
portfolio choices of our sample of insurers more carefully, we first determine whether there
exist different portfolio strategies using the entire sample of insurer groups (approximately
1,750 groups in 2014). Using all insurer groups, rather than our smaller sample of publicly
traded insurers lets us gain power in detecting different strategies.
Cluster analysis allows us to separate insurers into subgroups (clusters) that are likely
to have closer connections with each other than with those outside the cluster. As shown in
Appendix B, the cluster validation process indicates that the optimal number of clusters in
our data is three. Therefore, only a small number of portfolio strategies are being employed
by firms in the insurance industry. The average structure of the three clusters, namely the
coordinates of their centroids, is displayed in Figure 2. The figure allows one to evaluate
both the composition of the clusters (color code) and the general asset class allocation of
the sample (y-axis). Independently from the cluster and consistent with Figure 1, the most
represented asset classes are corporate bonds, equity, and U.S. government debt. ABS,
CMBS, GSE securities, mutual fund shares, sovereign bonds, and RMBS account for a
smaller portion of insurers’ portfolios. Cluster 1 of the sample is heavily tilted towards
municipal bonds and equity. Cluster 2 is mainly invested in corporate and GSE debt. Cluster
3 is dominated by U.S. government debt. In terms of the number of insurance groups in each
cluster, Cluster 1 and Cluster 2 are evenly split with approximately 45% of the sample of
all insurers in each cluster. The remaining 10% of insurance groups are in Cluster 3. If we
apply the cluster analysis separately in each year, the optimal number of clusters remains at
three and the composition of each cluster is relatively stable.
Figure 3 shows the distribution of our sample of publicly traded insurers in each cluster.
Most notably, few insurers in our sample are in Cluster 3 and the majority of insurers are in
13
Cluster 2. (In unreported results, we find that insurers infrequently move between clusters.)
A similar pattern is found for the portfolio allocation strategies of PSIFIs. Most PSIFIs are
in Cluster 2 and very few in Cluster 3 and there appears to be little difference between the
portfolio allocation strategy of PSIFIs and non-PSIFIs.
Using cluster analysis to determine portfolio strategy suggests that insurance companies
are likely to be very similar in their asset composition and therefore, potentially in their
trading strategy. Unlike the mutual fund literature in which there are a large number of
portfolio strategies, insurers appear to follow only a few. In the next section, we discuss a
methodology that captures the granularity of portfolio choices among insurers.
4
Portfolio Similarity
In order to determine whether insurers with similar portfolios are likely to trade in a
similar fashion, we construct a measure of portfolio similarity using cosine similarity. Cosine
similarity is well-suited to comparing the “distance” between two vectors and in economics,
has been more recently used in text analytics (Hanley and Hoberg (2010) and Hanley and
Hoberg (2012)) and analyzing hedge fund portfolios (Sias, Turtle, and Zykaj (2015)). An
important benefit to this methodology is that cosine similarity does not rely on self-reported
or empirically determined investment strategies in the determination of similarity and thus,
allows an objective measure of interconnectedness in portfolio allocation and security sales.
To construct this measure, we first calculate the proportional dollar value of each asset
class or issuer in an insurer’s portfolio at the end of each year. For each insurer, we create a
vector of portfolio weights using all asset classes or issuers that appear in insurers’ portfolios
in a given year. For example, the maximum vector in a given year is approximately 32,000
unique issuers and therefore, each insurer’s portfolio has a vector length of 32,000. If an
insurer does not invest in a particular issuer in a given year, the weight is set to 0.
To measure the degree of similarity between insurers i and j in year t, we calculate the
14
cosine similarity as the dot product of the pair’s portfolio weight vectors normalized by the
vectors’ lengths. We refer to this quantity as Similarityi,j,t where the portfolio weights are
measured at either asset class or issuer level.
Similarityi,j,t =
wi,t · wj,t
kwi,t k kwj,t k
(3)
Because all portfolio weight vectors have elements that are non-negative, this measure of
portfolio similarity has the property of being bounded in the interval (0,1). Intuitively, the
portfolio similarity between two insurers is closer to one when they are more similar and can
never be less than zero if they are entirely different.
Table 2 presents descriptive statistics for Similarity for the whole sample period and
three different years, 2002, 2008 and 2014. Average similarity based on asset class is high
and close to one. This finding is consistent with our cluster analysis, which finds only a
small number of portfolio strategies employed by insurers. Asset class similarity decreases
from 2002 to 2014 from a high of 0.74 in 2002 to 0.62 in 2014.
Although insurers appear to follow similar investment strategies at the asset class level,
there appears to be little overlap in the investment in individual issuers. The average cosine
similarity at the issuer level is only 0.10 in 2002 and this is only slightly lower in 2014. Overall,
issuer level similarity is fairly constant with only a small decrease during the financial crisis
of 2008. These findings point to a potential difference in interpretation of how interconnected
insurers’ portfolios are when comparing similarity at the asset class and issuer levels.
The table presents portfolio similarity by whether the insurer pairs are both PSIFI or
non-PSIFI. We find little difference between the portfolio similarity of pairs of PSIFIs and
non-PSIFIs. In each year examined, the similarity between a pair of PSIFIs is lower than
the similarity between a pair of non-PSIFIs. Although PSIFIs are assumed by the FSOC
to hold more similar securities, we do not find that this is the case. The average asset class
similarity ranges between 0.58 to 0.74 for PSIFIs, and 0.66 and 0.77 for non-PSIFI insurers
from 2002 to 2014. PSIFIs are also less similar to other PSIFIs at the issuer level.
15
Figure 5 shows the time series of the average pairwise similarity at the asset class and
issuer level for the sample of all publicly-traded insurers and for the subsamples of non-PSIFI
and PSIFI pairs. At both the asset class and issuer level, PSIFI pairs have lower portfolio
holdings similarity than non-PSIFI pairs and the level of similarity fluctuates over time.
At the asset class level, there is little difference between pairs of PSIFIs and pairs of nonPSIFI insurers until just after the financial crisis in 2010 when the similarity between PSIFIs
begins to decline. Whether this is due to the increased regulation and scrutiny is unclear
but suggests that PSIFIs are not holding the same assets to the same degree as before the
financial crisis. At the issuer level, PSIFI pairs are also less similar than non-PSIFI pairs
but we do not see the same time trend. In fact, the level of similarity in portfolio holdings
measured at the issuer level rises until 2012 when both PSIFI and non-PSIFI pairs have the
same issuer similarity, and then begins to decline afterward.
Table 2 also presents the return correlation between issuer pairs. We measure the annual return correlation for a pair of insurers RetCorr Pair at year-end using daily returns
throughout the year. This measure is commonly used in the systemic risk literature to
capture the potential for two entities to be interconnected. The average return correlation
between insurer pairs is 0.32 and follows a concave pattern over the three years in the table
with the highest return correlation in 2008 during the financial crisis. However, even though
PSIFIs are less similar in their portfolio construction compared to non-PSIFIs, the average
return correlation between pairs of PSIFIs is higher than the average return correlation between non-PSIFIs, and both subsamples follow a similar pattern as the sample as a whole.
PSIFIs, being large financial institutions by construction, are more likely to be affected by
common market factors, which is consistent with their return correlation being larger than
that of other insurers. This result indicates that portfolio similarity and return correlation
measures may capture different characteristics of insurers interconnectedness.
16
4.1
Determinants of Portfolio Similarity
To gain a better understanding of the determinants of pairwise portfolio similarity, we
examine the relationship between pairwise similarity and insurer pair characteristics. Because our dependent variable Similarity is a pairwise variable, we construct our independent
variables in a similar fashion. In order to control for firm characteristics, we use indicator
variables equal to one if both insurers are life (Life Pair ) or P&C (PC Pair ) and if both
insurers are PSIFI (PSIFI Pair ) or non-PSIFI (Non-PSIFI Pair ). We define a large insurer
as one with portfolio assets above the median and create indicator variables equal to one if
both insurers are large (Big Pair ) or small (Small Pair ). We also consider the asset class or
issuer concentration of an insurer’s portfolio and define it as high (low) if it is above (below)
the median for the sample. We then construct two pairwise indicator variables at both the
asset class (AC) and issuer (I) level that equal one if both insurers in the pair have portfolios
with high (Conc Pair ) or low (NonConc Pair ) concentration.
We estimate OLS regressions where the dependent variable, Similarityi,j,t , is the pairwise
similarity measure in a given year defined at either the asset class (Models (1) - (3)) or
issuer level (Models (4) - (6)). Table 3 presents the regression results. When we use asset
class similarity, we find that the similarity between two insurers is greater if they are in the
same line of business. When both insurers are either PSIFIs or non-PSIFIs, they have greater
asset class and issuer similarity but the relationship between being a PSIFI pair and portfolio
similarity is only marginally significant. We find similar relationships when examining the
determinants of portfolio similarity at the issuer level. However, insurers pairs who both are
life insurers and insurers pairs who are both PSIFIs are not more likely to be similar to each
other.
We split the sample of insurers into PSIFI pairs and non-PSIFI pairs (all pairs of firms
that include one PSIFI and one non-PSIFI are not included) to examine whether the determinants of greater similarity within these categories differ. At the asset class level, non-PSIFI
pairs in the same line of business are more similar but only PSIFIs who are life insurers
17
are more similar to each other when they are in the same line of business. Non-PSIFIs are
more similar if they are both large insurers and less similar if they are both small. When we
consider the issuer level of portfolio similarity, we find that non-PSIFIs are not more similar
to each other if they are in the same line of business but PSIFIs are. The relationship on
size for non-PSIFIs flips for issuer level similarity.
When examining both asset class and issuer level similarity as well as subsamples of
insurers, we consistently find that non-concentrated (more diversified) insurers are more
similar than concentrated insurer pairs. Diversification is an important risk management
tool to reduce individual insurance risk. However, recent theoretical works have shown that
full diversification is not optimal because it can increase systemic risk through various forms
of financial contagion (Allen, Babus, and Carletti (2012), Castiglionesi and Navarro (2008),
Ibragimov, Jaffee, and Walden (2011), Wagner (2010)). Ibragimov, Jaffee, and Walden
(2011), Wagner (2010), and Beale, Rand, Battey, Croxson, May, and Nowak (2011) show
that even though diversification of assets reduces each institution’s individual probability of
failure, it makes systemic crises more likely.
4.2
Return Correlation and Portfolio Similarity
The previous section documents that portfolio similarity is generally related to insurer
characteristics. In this section, we examine whether portfolio similarity is captured by public market information such as return correlation. For example, Billio, Getmansky, Lo, and
Pelizzon (2012), Neale, Drake, Schorno, and Semann (2012), and Brunetti, Harris, Mankad,
and Michailidis (2015) use return correlation as a measure of interconnectedness for insurance and interbank markets. This analysis is important because market-based measures of
interconnectedness for other types of nonbank financial companies that may contribute to
systemic risk, such as hedge funds, is not readily available but portfolio similarity may be
(Sias, Turtle, and Zykaj (2015)).9
9
See Ben Bernanke, in his speech to the Federal Reserve Bank of Atlanta in 2006 discussing the systemic
risk of hedge funds http://www.federalreserve.gov/newsevents/speech/bernanke20060516a.htm.
18
The advantage of using a market-based measure of interconnectedness like return correlation is that it is easy to compute. However, return correlation may not capture the full
extent to which insurers are connected through their portfolio strategies. Table 2 presents the
average return correlation between a pair of insurers for the sample as a whole and by PSIFI
category in 2002, 2008 and 2014. Return correlation increases during 2008 at the height of
the financial crisis from 0.26 to 0.45 and then declines during the last year of sample period
to 0.36. As can be seen in each year, PSIFI pairs have higher average return correlation
when compared to non-PSIFI pairs although the difference between the two narrows during
the financial crisis.
In order to determine whether the pairwise correlation of stock returns (RetCorr Pair )
is a good proxy for a pair of insurers’ portfolio similarity, we re-estimate the specifications
in Table 3 and include the return correlation between two pairs of insurers as an independent variable. Table 4 presents the estimation results. We generally find no relationship
between RetCorr Pair and asset class portfolio similarity after adjusting for the pair’s portfolio concentration, size, and line of business. The exception are PSIFI pairs, where we
find a positive relationship between asset class similarity and return correlation. Looking at
the issuer similarity, the relationship between return correlation and similarity for PSIFIs is
negative.
Overall, the findings in Table 4 indicate that a market-based measure of interconnectedness such as the correlation of insurers’ equity returns does not capture portfolio similarity
well. This is likely due to the fact that equity returns reflect many different aspects of
an insurer’s a balance sheet and operations of which portfolio allocation is one characteristic. Using return correlation alone to predict asset liquidation for insurers and possibly
other portfolio managers may be problematic. Below we investigate whether our portfolio
similarity measures contain information about future asset liquidation.
19
5
Portfolio Similarity and Asset Liquidation
We begin our analysis of the link between portfolio similarity and asset liquidation with
the descriptive statics presented in Table 5. The table reports insurers’ sales of each asset
class as a percentage of all sales from 2002 to 2014. The table shows that the majority of
sales are of corporate bonds, equity, GSE securities, mutual fund shares, and U.S. government
bonds. The proportion of corporate bonds and equity sold has decreased over time while the
proportion of mutual fund shares sold has increased. During the financial crisis of 2008 the
sale of equity decreased while the sale of GSE securities increased.
In order to examine whether insurers that are more similar are more likely to trade either
the same asset class or issuer, we construct a measure of sales similarity, Sales Similarityi,j,q
between insurers i and j in quarter q using the cosine similarity methodology described
earlier. For each insurer, we create a vector of sales weights for every possible asset class or
security issuer in a given quarter. If an insurer does not sale assets in a particular asset class
or issuer, the weight is set to 0.10
Figure 6 presents the times series of quarterly sales similarity for the sample of insurers
as well as for PSIFIs and non-PSIFIs. There is a great deal of variation in the similarity
in portfolio sales over the sample time period. At the asset class level, PSIFIs have greater
sales similarity, in general, with a slight decline over the sample period from around 0.6 to
0.45. When we define sales at the issuer level, PSIFIs always have greater sales similarity
than non-PSIFIs even though their portfolio similarity is lower. We do not see any evidence
that the sales similarity of insurers increased during the financial crisis. The figure provides
preliminary evidence to support regulators’ concerns about correlated trading behavior.
In Table 6, we investigate the determinants of sales similarity. At the asset class level, life
insurer pairs but not P&C insurer pairs tend to have greater sales similarity. Both more and
less concentrated pairs of insurers tend to sell the same asset classes. The results are different
10
If an insurer does not sell anything during the quarter, we cannot compute the cosine similarity with
other insurers and therefore, these pairs are not included in our tests.
20
at the issuer level. At the issuer level, insurers in the same business line are less likely to
sell the same issuers. If both insurers have high (low)concentration of issuers, they have
greater (lower) sales similarity. Therefore, diversification, while important for the similarity
in portfolio holdings, is less so when determining asset liquidation.
When examining whether PSIFI pairs have greater similarity, we find that regardless of
the type of sales similarity examined, PSIFI Pair is positively significant and non-PSIFI Pair
is negative. However, we find that if both non-PSIFIs are large, they have greater sales similarity. This means that both potentially systemic insurers and relatively large insurers tend
to sell similar issuers but smaller non-systemic insurers do not.
Next we investigate whether our measure of portfolio similarity is able to predict sales
similarity as suggested by the theoretical work of Allen, Babus, and Carletti (2012). Specifically, we estimate OLS regressions of quarterly sales similarity on the prior year portfolio
holdings similarity constructed at the issuer or asset class levels. We control for other pair
characteristics and include year-quarter fixed effects. Table 7 presents the results from these
estimations.
At both the asset class and issuer level, we find a significant relationship between the
similarity in portfolio holdings and the similarity in asset sales. Insurers with more similar
holdings have more similar sales and this relationship is stronger at the issuer level than at
the asset class level. The R2 for the models using asset class is around 10% compared to
the R2 of 15% when using issuer similarity. With the exception of life insurer pairs at the
asset class level, insurers in the same line of business have lower, not higher, sales similarity.
When we split the sample into PSIFI and non-PSIFI pairs, we continue to find a significant
relationship between portfolio and sales similarity.
We also document that two insurers that both hold concentrated portfolios have greater
sales similarity than two insurers that both hold non-concentrated portfolios. Although
concentration and size have been proposed as potential metrics to identify systemically important financial institutions (Haldane and May (2011) Gai, Haldane, and Kapadia (2011)
21
and Allen, Babus, and Carletti (2012)), our results suggest that portfolio similarity adds
significantly in determining the potential for similar asset liquidation.
All of the relationships for the full sample hold generally for the subsamples of PSIFI
pairs and non-PSIFI pairs. It should be noted that for non-PSIFIs, the similarity in sales is
greater for larger insurers and smaller for smaller insurers. Overall our findings confirm that
size and concentration are important determinants of sales similarity and that our measure
of portfolio similarity is significantly related to the decision as to whether or not to sell
the same asset classes and issuers. Thus, our portfolio interconnectedness measure appears
to capture information about future sales that could be used to monitor insurers who may
contribute more to the transmission of systemic risk through the asset liquidation channel.
5.1
Return Correlation and Sales Similarity
As noted previously, a number of studies, most notable Billio, Getmansky, Lo, and Pelizzon (2012) use return correlations as a proxy for interconnectedness among financial institutions. In this section, we examine whether our findings on the sales similarity between two
issuers can be characterized by the pairwise return correlation.
In Table 8, we include the return correlation between a pair of issuers in the quarter
before sales similarity is measured. At the asset class level, the return correlation is positively related to sales similarity. Including portfolio similarity in the set of independent
variables results in the coefficients of both return correlation and similarity being positive
and significant. We find that the the return correlation is only significant for PSIFI pairs.
For non-PSIFI pairs, return correlation does not add additional information.
When we measure portfolio similarity at the issuer level, we document that by itself, the
return correlation does not have any marginal contribution in predicting issuer level sales
similarity for the sample as a whole. However, when coupled with portfolio similarity, the
return correlation is significantly and positively related to sales similarity but is a significant
predictor of sales similarity only for the subsample of non-PSIFI pairs. For PSIFI pairs,
22
adding return correlation to the specification does not add any additional power. Comparing
models (5) and (6) in Table 6 to models (7) and (8) in Table 8, we see that adding return
correlation does not increase the R2 of the regressions for either subsample.
6
Effect of Capital Regulation and the Financial Crisis
on Sales Similarity
Entities subject to capital regulation have an incentive to engage in asset sales when
capital is depleted. The literature has documented that insurers replenish capital by selling
downgraded assets (Ellul, Jotikasthira, and Lundblad (2011)) and/or selling liquid assets
(Ellul, Jotikasthira, Lundblad, and Wang (2015)). In this section, we examine whether
our findings on the relationship between portfolio similarity and sales similarity are due to
capital constrained insurers. In addition, we analyze whether the relationship is due to the
similarity in downgraded securities and liquid securities. Since the need to replenish capital
is greatest in the period surrounding the financial crisis, we explore whether our results only
hold during the financial crisis period.
6.1
Risk-Based Capital Ratios
We assess the extent to which an insurer is regulatory-capital constrained through its ratio
of statutory to risk-based capital (RBC ratio). A low RBC ratio may provide an incentive for
an insurer to engage in asset sales to rebalance their portfolio in order to replenish capital.
Therefore, pairs of insurers who both have low RBC ratios and who hold similar assets, may
be more likely to re-balance their portfolios in similar ways.
We consider an insurer under-capitalized if its RBC ratio is below the median and wellcapitalized if the RBC ratio is above the median for the sample. We then construct two
pairwise indicator variables, (RBC High Pair ) and (RBC Low Pair ), equal to one if both
insurers are well-capitalized or under-capitalized, respectively. In our specifications in Table
23
9, we include RBC High Pair and RBC Low Pair and the interaction terms between these
indicator variables and portfolio similarity at both the asset class and issuer level. As in
prior tables, we also examine PSIFI pairs and non-PSIFI pairs separately.
The results in Table 9 captures the interaction effect of regulatory capital constraints
and portfolio similarity on sales similarity. At the asset-class level, we find that pairs of
insurers with greater portfolio similarity and who both have low RBC ratios, have greater
sales similarity. This finding is consistent with the literature on the effects of regulation on
asset liquidation. Capital constrained insurers who have similar portfolios are more likely to
sell the same types of assets. The interaction term of similarity and low RBC ratio remains
significant and positive for both the samples of PSIFI pairs and non-PSIFI pairs. There is
no relationship between sales similarity of insurer pairs who have high asset class similarity
and high RBC ratios.
When considering the interaction between portfolio similarity at the issuer level and low
RBC ratio, we do not find any evidence that insurer pairs that are more similar in their
portfolio holdings but capital constrained have greater sales similarity. This suggests that
any asset liquidation that occurs when insurers are capital constrained occur within the
same asset class but not necessarily the same issuer. Interestingly, when we examine the
subsample of PSIFI pairs, capital constrained PSIFIs are less likely to make the same selling
decision at the issuer level.
Overall, this subsection suggests a stronger relationship between portfolio similarity and
sales similarity when both pairs of insurers are constrained and is only apparent for similarity at the asset class level. Next we examine whether our findings are driven by either
downgraded securities or the liquidity of the assets.
6.2
Liquidity and Downgrades
In examining the impact of liquidity, we are trying to determine whether the similarity
in portfolio holdings has the potential to be disruptive to markets. If the similarity of the
24
sales among insurers is due to the similarity in the portfolio of liquid assets, regulators may
be less concern about forced selling and price impact. On the other hand, if the similarity
of sales is a function of the similarity in the portfolio of illiquid assets, then systemic risk
regulators might be worried that having interconnected insurers will have a larger impact
on prices and increase the probability of a downward spiral in valuation (Brunnermeier and
Pedersen (2009) and Cont and Wagalath (2015)). We classify an issuer as liquid if it falls
within the following asset classes: equity, mutual fund shares, US government securities,
GSE securities, and sovereign bonds. Issuers in the remaining asset classes are considered
illiquid: corporate bonds, municipal bonds, RMBS, CMBS and ABS.
A similar logic to the one discussed above regarding liquidity prevails with respect to the
effect of downgraded assets. A number of studies document that asset sales of distressed
securities by capital constrained insurers tend to depress prices (Ellul, Jotikasthira, and
Lundblad (2011) and Merrill, Nadauld, Stulz, and Sherlund (2013)) and is one of the reasons
why some insurers have been designated as systemically important. Thus, it could be the
case that the relationship between portfolio similarity and sales similarity is simply due to
the similarity in downgraded assets. Using credit ratings information from DataScope, we
identify the year in which a security is first downgraded from investment grade (IG) to noninvestment grade (NIG) by S&P, Moodys or Fitch. This information is aggregated to the
security’s issuer level and we define a downgraded issuer as one that has at least one of its
securities downgraded from IG to NIG in a given year. For portfolio holdings, we define a
variable Downgraded as an indicator variable equal to 1 if an issuer, held by a given insurer
at the end of a given year is downgraded in the following year, and 0 otherwise. All holdings
that are not downgraded are classified as NotDowngraded.
Figure 7 shows the proportion of portfolio holdings and sales that are comprised of illiquid
issuers and downgraded issuers. In Panel (a), approximately 50% of insurers holdings are
composed of issuers in asset classes that we consider illiquid and this is relatively constant
over the time period. In contrast, insurers are selling fewer illiquid assets during and after
25
the financial crisis than before. Panel (b) shows a significant time trend in the percentage
of insurers holdings that are downgraded and a sharp increase in 2009 in the sale of these
securities. Thus, it could be the case that the liquidity or credit quality of the securities
affects the relationship between portfolio similarity and sales similarity.
We decompose the portfolio of an insurer into two pares: one that is composed of issuers
that are classified as liquid and another that is composed of issuers classified as illiquid.
In the same fashion, we decompose the portfolio of an insurer into issuers that are downgraded in a year and issuers that are not downgraded. We compute four different portfolio
similarities between firms i and firms j in year t: Similarity Illiquid, Similarity Liquid, Similarity Downgraded, and Similarity NotDowngraded. Each of these portfolio similarities measures the extent to which a pair of insurers are exposed to illiquid (liquid) and downgraded
(not downgraded) issuers.11
Table 10 presents the analysis of the relationship of the decomposed portfolio similarities
and the sales similarities between a pair of insurers. As can be seen from the table, our
results are not driven specifically by the liquidity of the asset or by downgraded issuers.
The coefficient on each of the portfolio similarity interactions with liquidity and downgrade
status is highly significant and positive. The sales similarity of a pair of issuers is affected
by the portfolio similarity of all securities in the portfolio and not just those that have lower
liquidity or lower credit ratings. Next, we examine whether the financial crisis exacerbates
any of these relationships.
6.3
The Financial Crisis
Given the concern about the selling behavior of regulated entities during the financial
crisis, it is natural to examine whether our findings are due only to the time period of the
crisis. We consider three different time periods and create indicator variables equal to 1 if
the time period is: (1) Pre-Crisis from 2002-2006, (2) Crisis from 2007 to 2009, and (3)
11
Note that we do not include asset class in this table because the downgrade variable requires issuer
information.
26
Post-Crisis from 2010 to 2014. We interact two of these indicators, Crisis and Post-Crisis
with portfolio similarity using liquid/illiquid assets and issuers that are downgraded/not
downgraded.
In Table 10, for the sample as a whole, we find that the relationship between the portfolio
similarity of both illiquid and downgraded portfolios with sales similarity is stronger during
the crisis. We do not find any change in the relationship for the portfolios composed of
liquid or non-downgraded issuers. In other words, during the financial crisis, correlated
selling decisions of pairs of insurers are more greatly influenced by the holdings of similar
securities that are either hard to trade or have experience credit deterioration.
For downgraded issuer, PSIFI pairs have a stronger relationship between the portfolio
similarity of these issuers and sales similarity but a decline in the relationship for the portfolio
of non-downgraded issuers. We interpret these findings to mean that, during the financial
crisis, PSIFIs’ selling decisions were more heavily weighted toward their joint holdings of
downgraded assets at the expense of their joint holdings of non-downgraded issuers. The
greater relationship between the portfolio holdings of downgraded issuers and sales similarity
continues to hold after the financial crisis as well indicating that investment in downgraded
issuers has a greater impact on the re-balancing decisions of insurers during and after the
financial crisis.
The greater influence of the portfolio similarity of downgraded issuers in predicting sales
similarity is likely due to the fact that a greater proportion of both the portfolio holdings and
sales of insurers are now composed of downgraded assets. Thus, the influence of downgraded
issuers in predicting selling behavior may be attributed to its increase importance of these
securities in insurers’ portfolios. It seems worth also pointing out that, even though the crisis
dummy is highly statistically significant when interacted with the downgraded variable, the
beta is still quite small in magnitude (0.076). As a consequence, the overall beta during the
financial crisis for the downgraded holdings similarities (0.157+0.076) it is still smaller than
the beta during the financial crisis for the non-downgraded holdings similarities (0.464-0.049).
27
This result points to one of the main finding of our paper, i.e. that the co-investment decision
of insurers are an important determinant of selling behavior in all economic condition and
across all type of securities.
7
Conclusion
In this paper, we develop a novel measure of pairwise insurance interconnectedness that
focuses on insurance portfolio similarities. The Financial Stability Oversight Council (FSOC)
has an authority to designate nonbank Systemically Important Financial Institutions (SIFIs).
Size and interconnectedness between financial companies are two factors FSOC currently
considers in the designation. However, there are no specific guidelines about calculating
interconnectedness. We attempt to add to the academic literature and to the important
policy discussion by explicitly measuring pairwise insurance interconnectedness using the
cosine similarity of insurers’ portfolio holdings and measuring its effect on selling behavior.
Our findings show that pairs of insurers that have greater portfolio similarity will have
greater sales similarity at both the asset class and issuer level and this result holds when
both pairs are potential SIFIs (PSIFIs) and non-PSIFIs. Portfolio re-balancing at the asset
class but not issuer level is more likely when both insurers are capital constrained indicating
that insurers sell the same type of security but not the same issuer of the security. We find
that both the portfolio similarity of liquid and illiquid securities as well as downgraded and
non-downgraded issuers predict liquidation decisions. However, the relationship between
the portfolio similarity of illiquid and downgraded issuers and sales similarity is stronger
during the financial crisis meaning that the co-investment decisions of these assets are more
likely to affect selling behavior in times of market stress. Finally, we show that our measure
of interconnectedness and its relationship to selling behavior is not captured by market
observable characteristics such as stock return correlation.
Overall, our results support the use of our measure to capture the important mechanics
28
of the asset liquidation channel of systemic risk transmission in the insurance industry.
Specifically, it can predict the sales of similar assets and issuers. The measure can be used by
insurance companies and regulators to assess the level of interconnectedness in the industry
and the possible impact of interconnectedness on asset liquidation.
29
References
Allen, Franklin, Ana Babus, and Elena Carletti, 2012, Asset commonality, debt maturity
and systemic risk, Journal of Financial Economics 104, 519 – 534.
Ambrose, Brent W., Nianyun (Kelly) Cai, and Jean Helwege, 2008, Forced selling of fallen
angels, The Journal of Fixed Income 18, 72–85.
Beale, N., D.G. Rand, H. Battey, K. Croxson, R.M. May, and M. A. Nowak, 2011, Individual
versus systemic risk and the regulators dilemma, Proceedings of the National Academy of
Sciences 108, 12647–12652.
Billio, Monica, Mila Getmansky, Andrew W. Lo, and Loriana Pelizzon, 2012, Econometric
measures of connectedness and systemic risk in the finance and insurance sectors, Journal
of Financial Economics 104, 535 – 559.
BIS, 2014, The G-SIB Assessment Methodology - Score Calculation, Working paper, .
Blei, Sharon K., and Bakhodir Ergashev, 2014, Asset commonality and systemic risk
among large banks in the United States, Working paper available at http://ssrn.com/
abstract=2503046.
Brunetti, Celso, Jeffrey H. Harris, Shawn Mankad, and George Michailidis, 2015, Interconnectedness in the Interbank Market, Federal Reserve Board of Governors working paper.
Brunnermeier, Markus K., and Lasse Heje Pedersen, 2005, Predatory Trading, Journal of
Finance 60, 1825–1863.
Brunnermeier, Markus K., and Lasse Heje Pedersen, 2009, Market Liquidity and Funding
Liquidity, The Review of Financial Studies 22, 2201–2238.
Castiglionesi, Fabio, and Noemı́ Navarro, 2008, Optimal fragile financial networks, Working
paper available at http://ssrn.com/abstract=1089357.
Chiang, Chia-Chun, and Greg Niehaus, 2016, Investment herding by life insurers and its
impact on bond prices, University of South Carolina working paper.
Cont, Rama, and Lakshithe Wagalath, 2015, Fire Sales Forensics: Measuriing Endogenous
Risk, Wharton working paper.
Cummins, J David, and Mary A Weiss, 2014, Systemic risk and the US insurance sector,
Journal of Risk and Insurance 81, 489–528.
Danielsson, Jon, Hyun Song Shin, and Jean-Pierre Zigrand, 2004, The Impact of Risk Regulation on Price Dynamics, Journal of Banking and Finance 28, 1069–1087.
Dunn, Joseph C, 1974, Well-separated clusters and optimal fuzzy partitions, Journal of
cybernetics 4, 95–104.
30
Ellul, Andrew, Chotibhak Jotikasthira, and Christian T. Lundblad, 2011, Regulatory pressure and fire sales in the corporate bond market, Journal of Financial Economics 101, 596
– 620.
Ellul, Andrew, Chotibhak Jotikasthira, Christian T. Lundblad, and Yihui Wang, 2015, Is historical cost accounting a panacea? Market stress, incentive distortions, and gains trading,
Journal of Finance 70, 2489–2538.
Fang Cai, Song Han, Dan Li, and Yi Li, 2016, Institutional herding in the corporate bond
market, Federal Reserve Board of Governors working paper.
FSOC, 2013, Basis for the Financial Stability Oversight Council’s Final Determination Regarding Prudential Financial, Inc., Working paper, .
Gai, Prasanna, Andrew Haldane, and Sujit Kapadia, 2011, Complexity, concentration and
contagion, Journal of Monetary Economics 58, 453–470.
Geneva Association, The, 2010, Systemic risk in insurance: an analysis of insurance and
financial stability, Special Report of the Geneva Association Systemic Risk Working Group.
Grace, Martin F, 2010, The insurance industry and systemic risk: evidence and discussion,
Networks Financial Institute Policy Brief p. 02.
Haldane, Andrew G., and Robert M. May, 2011, Systemic risk in banking ecosystems, Nature
469, 351–355.
Handl, Julia, Joshua Knowles, and Douglas B Kell, 2005, Computational cluster validation
in post-genomic data analysis, Bioinformatics 21, 3201–3212.
Hanley, Kathleen Weiss, and Gerard Hoberg, 2010, The information content of IPO prospectuses, Review of Financial Studies pp. 2821–2864.
Hanley, Kathleen Weiss, and Gerard Hoberg, 2012, Litigation risk, strategic disclosure and
the underpricing of initial public offerings, Journal of Financial Economics pp. 235–254.
Hanley, K. W., and S. Nikolova, 2015, Rethinking the use of credit ratings in capital regulation, Working paper available at http://ssrn.com/abstract=2357145.
Harrington, Scott E, 2009, The financial crisis, systemic risk, and the future of insurance
regulation, Journal of Risk and Insurance 76, 785–819.
IAIS, 2013, Global Systemically Important Insurers: Initial Assessment Methodology, Working paper, .
IAIS, 2015, Global Systemically Important Insurers: Proposed Updated Assessment Methodology, Working paper, .
Ibragimov, Rustam, Dwight Jaffee, and Johan Walden, 2011, Diversification disasters, Journal of Financial Economics 99, 333 – 348.
31
Kartasheva, Anastasia, 2014, Insurance in the Financial System, Bank of International Settlements Working Paper.
MacQueen, James, 1967, Some methods for classification and analysis of multivariate observations, Proceedings of the fifth Berkeley symposium on mathematical statistics and
probability 1, 281–297.
Manconi, Alberto, Massimo Massa, and Ayako Yasuda, 2012, The role of institutional investors in propagating the crisis of 2007172008, Journal of Financial Economics 104, 491
– 518.
Merrill, Craig B., Taylor Nadauld, Rene M. Stulz, and Shane M. Sherlund, 2013, Why were
there fire sales of mortgage-backed securities by financial institutions during the financial
crisis?, Working paper available at http://ssrn.com/abstract=2212684.
Neale, Faith R., Pamela P. Drake, Patrick Schorno, and Elias Semann, 2012, Insurance
and Interconnectedness in the Financial Services Industry, University of North CarolinaCharlotte working paper.
Park, Sojung Carol, and Xiaoying Xie, 2014, Reinsurance and Systemic Risk: The Impact
of Reinsurer Downgrading on Property–Casualty Insurers, Journal of Risk and Insurance
81, 587–622.
Rousseeuw, Peter J, 1987, Silhouettes: a graphical aid to the interpretation and validation
of cluster analysis, Journal of computational and applied mathematics 20, 53–65.
Shleifer, Andrei, and Robert Vishny, 2011, Fire Sales in Finance and Macroeconomics, Journal of Economic Perspectives 25, 29–48.
Sias, Richard, J.J. Turtle, and Bierina Zykaj, 2015, Hedge fund crowds and mispricing,
Management Science p. forthcoming.
Wagner, Wolf, 2009, Efficient asset allocations in the banking sector and financial regulation,
International Journal of Central Banking 5, 751797.
Wagner, Wolf, 2010, Diversification at financial institutions and systemic crises, Journal of
Financial Intermediation 19, 373 – 386.
32
Appendix A: Variable Definitions
Variable
Definition
Big
An indicator variable equal to 1 if a non-PSIFI insurer’s portfolio assets are above the sample
median at calendar year end, 0 otherwise.
Big Pair
An indicator variable equal to 1 if Big=1 for both insurers in a pair, 0 otherwise.
Concentration
Issuer/asset class Herfindahl index of an insurer’s portfolio at calendar year end:
N
P
2
Concentrationi,t =
wi,n,t
where wi,n,t is issuer/asset class n’s proportion in insurer
n=1
i’s portfolio at the end of year t. Issuer/asset class proportions are calculated as the dollar
amount invested in each issuer/asset class relative to the total value of the insurer portfolio
Conc AC
An indicator variable equal to 1 if an insurer’s asset class concentration is above the sample
median at calendar year end, 0 otherwise.
Conc Pair AC
An indicator variable equal to 1 if Conc AC=1 for both insurers in a pair, 0 otherwise.
Conc I
An indicator variable equal to 1 if an insurer’s issuer concentration is above the sample
median at calendar year end, 0 otherwise.
Conc Pair I
An indicator variable equal to 1 if Conc I=1 for both insurers in a pair, 0 otherwise.
CRISIS
An indicator variable equal to 1 for the years 2007, 2008, and 2009 and 0 otherwise.
Life
An indicator variable equal to 1 if more than 50% of portfolio assets are held by insurance
companies in the group that are categorized by A.M. Best as providing life insurance, 0
otherwise.
Life Pair
An indicator variable equal to 1 if Life=1 for both insurers in a pair, 0 otherwise.
Non-PSIFI
An indicator variable equal to 1 if an insurance company does not meet the $50 billion in
assets SIFI designation threshold in any year during our sample period.
NonConc Pair
An indicator variable equal to 1 if Conc=0 (either Conc AC or Conc I) for both insurers in
a pair, 0 otherwise.
Other Pair
An indicator variable equal to 1 if Life=0 and P&C=0 for both insurers in a pair, 0 otherwise.
P&C
An indicator variable equal to 1 if more than 50% of portfolio assets are held by insurance
companies in the group that are categorized by A.M. Best as providing property and casualty
insurance, 0 otherwise.
P&C Pair
An indicator variable equal to 1 if P&C=1 for both insurers in a pair, 0 otherwise.
PSIFI
An indicator variable equal to 1 if an insurer could potentially be designated as a SIFI because
it meets the $50 billion in assets threshold in at least one year during our sample period.
RBC
A measure of capital adequacy calculated as the ratio of total adjusted capital to authorized
control level risk-based capital.
RBC Pair High
An indicator variable equal to 1 if RBC is above the sample median in a given year for both
insurers in a pair, 0 otherwise.
RBC Pair Low
An indicator variable equal to 1 if RBC is below the sample median in a given year for both
insurers in a pair, 0 otherwise.
RetCorr Pair
For a pair of insurers, the annual correlation of daily holding-period returns calculated at
year end for each pair of insurers.
RetCorr Qtr Pair
For a pair of insurers, the annual correlation of daily holding-period returns calculated at a
three month lag for each pair of insurers.
Similarity
Similarity measure based on the cosign similarity between a pair of insurers’ issuer portfolio
weights (Similarity I) or between a pair of insurers’ asset class portfolio weights (Similarity AC).
Small
An indicator variable equal to 1 if a non-PSIFI insurer’s portfolio assets are below the sample
median at calendar year end, 0 otherwise.
Small Pair
An indicator variable equal to 1 if Small=1 for both insurers in a pair, 0 otherwise.
TA
Total assets at calendar year end (AT Y in Compustat).
ln(TA Pair)
Natural logarithm of the sum of total assets for a pair of insurers at a calendar year end.
33
Appendix B: Cluster Analysis
Cluster Algorithm
Cluster analysis could be performed using several algorithms that differ significantly in
their notion of what constitutes a cluster and how to efficiently find clusters. The approach
used in our paper is largely based on the concept that clusters include groups with small
distances among the cluster members with particular statistical distributions. As described
in more detail below, we apply internal validation measures, namely Dunn Index (Dunn,
1974), Silhouette Width (Rousseeuw, 1987) and Connectivity (Handl, Knowles, and Kell,
2005), on the most utilized unsupervised clustering algorithms (Self Organizing Maps, Self
Organizing Tree Maps, K-means, hierarchical).
The optimal number of clusters (Nopt ) is finally obtained computing the mode of the
optimal number of clusters in each of the 12 years (Nt ).
Nopt = M o(Nt )
(4)
Coherently, the optimal algorithm (Copt ) is derived by counting the number of times an
algorithm appears as locally optimal over the 12 years (Ct ) and selecting the maximum
value.
Copt = M ax(
12
X
Ct )
(5)
i=1
We run the unsupervised K-means algorithm (MacQueen, 1967), yearly, with the following setting:12
i) for the first year (Yt with t = 1) the number of clusters (3);
ii) for the following year (Yt with t = [2 : 12]) the centroids obtained by the cluster of the
12
The algorithm is based a finite number of cycle aimed at defining the optimal cluster centroids according
to the minimization of the distance of the
Pk n data
Pn points from their respective cluster centers, represented by
the following objective function: J = j=1 i=1 kxji − cj k2 where xji is a data point and cj is the cluster
center.
34
previous year (Yt−1 ).
The constraint for the cluster number in the first year comes from the outcome of the
validation step. The constraint for the centroids’ structure of the other years is set to
introduce a short-time memory effect in the evolution of the clusters over time. The link of
the cluster structures over time allows us to observe the transitions of the insurers among
clusters year by year.
We then analyze the clusters looking at:
i) size both in term of number of companies and volumes (amount of assets);
ii) centroids’ structure;
iii) transitions of companies among clusters over time.
The average structure of the 3 cluster’s centroids (xi ) is computed as the average over
time of the centroids’ components (xit ).
12
xi
1 X i
=
(x )
12 t=1 t
(6)
Finally the yearly net flow (N etF lowi ) for cluster i is computed as follows:
F lowi,t =
X
Ij,t In −
X
Ij,t Out
(7)
j=i
j6=i
The cluster validation process applied to the yearly dataset provides the best fitting
algorithm for the number of clusters. Each validation methodology is applied yearly kmeans.
A clear indication emerges for the optimal number of clusters being 3 clusters as this appears
21 times over 33 possible outcomes.
13
13
Details on the validation are provided upon request.
35
Cluster Validation
To validate the cluster approach we selected a set of measures that reflect the degree of
compactness, connectedness, and separation of the cluster partitions,tested respectively with
Connectivity, Dunn index and Silhouette width.
Connectivity (Handl, Knowles, and Kell, 2005) Connectivity measures estimates to
what extent the nearest observations (in our case insurers) are placed in the same cluster.
We define N as the number of observations in the sample, M the number of attributes of
each observation (namely the coordinate of the observation in an M-dimensional space) and
nni(j) the j th nearest neighbor of observation i. Let xi,nni(j) be
xi,nni(j)




0, if i and i are in the same cluster




1
=
, otherwise

j






(8)
Stated that, for a specific cluster partition C = {C1 , ...Ck } of the N observations, connectivity
is defined as:
Conn(C) =
N X
L
X
xi,nni(j)
(9)
i=1 j=1
where L defines the number of neighbor to use.
The connectivity has values between 0 and ∞ and should be minimized.
Silhouette Width (Rousseeuw, 1987) The Silhouette Width is the average of each observation’s Silhouette Value. The Silhouette Value is defined as:
S(i) =
b i − ai
,
max(bi , ai )
(10)
where ai is the average distance between observation i and the other observations belonging
to the same cluster and bi is the average distance between i and the observations in the
36
”nearest neighboring” cluster defined as:
bi =
X dist(i, j)
,
C(i)
n(C
k)
j∈C
min
Ck ∈C
(11)
k
where C(i) is the cluster containing observation i, dist(i, j) is the distance between observation i and j and n(C) is the cardinality of cluster C.
Silhouette Width values lies in [−1, 1] and it should be maximized.
Dunn Index (Dunn, 1974) Dunn Index is the ratio of the smallest distance between
observations not in the same cluster and the largest intra-cluster distance
D(C) =
minCk ,Cl ∈C,C(k)6=Cl (mini∈Ck ,j∈Cj dist(i, j))
,
maxCm ∈C diam(Cm )
where diam(Cm ) is the maximum distance between observations in cluster Cm .
Dunn Index lies between [0, ∞] and should be maximized.
37
(12)
Figure 1: Portfolio Composition Through Time
This figure presents the cross-sectional average composition of insurer portfolios over 2002-2014 by asset
class.
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
2002
2003
ABS
2004
2005
CMBS
Debt
2006
Equity
2007
GSE
2008
Muni
38
2009
Mutual Funds
2010
RMBS
2011
Sovereign
2012
US Govt
2013
2014
Figure 2: Clusters Structure for Holdings by Asset Class
This figure presents the average asset class dollar composition for the three clusters based on the holdings
for all insurer groups over 2002-2014.
100%
75%
50%
25%
0%
Cluster(1
ABS
CMBS
Debt
Cluster(2
Equity
GSE
Municipal;bonds
39
Mutual;funds
Cluster(3
RMBS
Sovereign;bonds
US;Govt
Figure 3: Cluster Distribution of Publicly Traded Insurers Using Asset Class Holdings
The figure shows the distribution among the three clusters from 2002 to 2014 for the sample of publicly
traded insurance companies used in this analysis.
100
90
80
70
60
50
40
30
20
10
0
2002
2003
2004
2005
2006
2007
Cluster21
2008
Cluster22
40
2009
2010
Cluster23
2011
2012
2013
2014
Figure 4: Cluster Distribution of Asset Classes for PSIFIs and Non-PSIFIs
The figure shows the distribution among the three clusters from 2002 to 2014 for insurers in the sample
classified as PSIFIs and non-PSIFIs We identify PSIFIs as any insurer who has assets of at least $50 billion
at any time during the sample period.
30
25
20
15
10
5
0
2002
2003
2004
2005
2006
2007
Cluster21
2008
2009
Cluster22
2010
2011
2012
2013
2014
2011
2012
2013
2014
Cluster23
(a) PSIFI
60
50
40
30
20
10
0
2002
2003
2004
2005
2006
2007
Cluster21
2008
2009
Cluster22
(b) Non-PSIFI
41
2010
Cluster23
Figure 5: Pairwise Portfolio Similarity
The figure shows average pairwise portfolio similarity computed at the (a) asset class and (b) issuer level
over 2002-2014. The red line represents the average for the sample of all publicly traded insurers. The violet
line represents the average for PSIFI insurers. The green line represents the average for non-PSIFI insurers.
0.80
0.75
0.70
0.65
0.60
0.55
0.50
2002
2003
2004
2005
2006
All
2007
2008
2009
Non-­‐‑PSIFI_Pair
2010
2011
2012
2013
2014
2012
2013
2014
PSIFI_Pair
(a) Asset Class Similarity
0.14
0.13
0.12
0.11
0.10
0.09
0.08
0.07
0.06
2002
2003
2004
2005
All
2006
2007
2008
2009
Non-­‐‑PSIFI_Pair
(b) Issuer Similarity
42
2010
2011
PSIFI_Pair
Figure 6: Quarterly Pairwise Sales Similarity
The figure shows average quarterly pairwise sales similarity computed at the (a) asset class and (b) issuer
level from 2002-2014. The red line represents the average for the sample of all publicly traded insurers. The
violet line represents the average for the subsample of PSIFI insurers. The green line represents the average
for the subsample of non-PSIFI insurers.
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
2002
2003
2004
2005
2006
All
2007
2008
2009
2010
Non-­‐‑PSIFI_Pair
2011
2012
2013
2014
2012
2013
2014
PSIFI_Pair
(a) Asset Class Similarity
0.3
0.25
0.2
0.15
0.1
0.05
0
2002
2003
2004
2005
2006
All
2007
2008
2009
2010
Non-­‐‑PSIFI_Pair
(b) Issuer Similarity
43
2011
PSIFI_Pair
Figure 7: Proportion of Holdings and Sales of Illiquid and Downgraded Securities
This figure presents the proportion of holdings and sales that are composed of illiquid and downgraded
issuers.
60.0%
50.0%
40.0%
30.0%
20.0%
10.0%
0.0%
2002
2003
2004
2005
2006
2007
2008
% Illiquid Holdings
2009
2010
2011
2012
2013
2014
% Illiquid Sales
(a) Illiquid Issuers
14.0%
12.0%
10.0%
8.0%
6.0%
4.0%
2.0%
0.0%
2002
2003
2004
2005
2006
2007
2008
% Downgraded Holdings
2009
(b) Downgraded Issuers
44
2010
% Downgraded Sales
2011
2012
2013
2014
45
Insurer Characteristics
TA ($B)
Asset Class Composition
Corporate bonds
GSE securities
Municipal bonds
US govt securities
RMBS
CMBS
ABS
Sovereign bonds
Equity
Mutual fund shares
Asset class concentration
Issue/Issuer Composition
Number of issues
Number of issuers
Issuer concentration
13.5
39.5
10.1
2.8
3.7
2.1
2.0
2.4
0.2
11.6
0.0
0.36
1,199
790
0.02
41.2
12.0
9.1
7.1
3.3
2.7
2.7
1.0
17.4
3.4
0.37
2,937
1,475
0.06
Median
138.6
Mean
All
4,995
2,065
0.15
17.8
9.8
12.1
9.9
4.0
2.5
2.2
3.6
17.4
9.0
0.14
316.7
SD
2,819
1,479
0.06
48.4
12.3
5.3
7.8
3.8
2.9
2.8
1.1
12.9
2.7
0.39
160.6
Mean
1,252
822
0.02
48.4
9.6
2.2
3.7
2.2
2.3
2.6
0.2
9.7
0.0
0.37
19.2
Median
Life
4,731
2,067
0.17
17.5
10.9
8.0
13.7
5.0
2.6
2.3
4.0
11.8
7.5
0.16
343.2
SD
2,119
995
0.03
32.7
11.2
16.0
6.5
2.9
2.8
4.0
2.0
21.1
0.8
0.30
85.6
Mean
816
646
0.02
32.9
10.9
13.8
5.0
2.3
1.9
3.1
0.2
17.7
0.0
0.27
8.8
Median
P&C
3,644
1,079
0.03
13.9
6.7
13.7
5.7
2.8
3.3
4.5
6.2
17.3
2.0
0.10
268.1
SD
6,678
3,183
0.06
36.4
7.9
5.1
7.0
3.1
4.1
3.3
2.0
25.5
5.7
0.37
398.3
Mean
4,575
2,309
0.02
33.5
6.7
2.0
3.1
2.3
4.0
3.2
0.8
18.0
0.1
0.34
175.1
Median
PSIFI
7,055
2,934
0.13
17.0
6.0
7.9
12.3
3.0
3.0
2.2
4.8
21.1
10.0
0.16
446.4
SD
1,466
820
0.05
44.3
13.3
11.3
6.5
3.5
2.5
2.7
0.9
13.2
1.7
0.36
11.5
Mean
727
526
0.02
46.4
10.9
5.0
4.6
2.1
1.6
2.2
0.1
9.1
0.0
0.32
4.4
Median
Non-PSIFI
2,540
876
0.14
17.3
10.8
14.0
7.4
4.4
2.6
2.6
3.8
14.1
5.8
0.14
13.6
SD
The table presents portfolio composition statistics for the sample of publicly traded insurers from 2002 to 2014. Life insurers operate primarily in life
and health lines of business. P&C insurers operate predominantly in property and casualty lines of business. PSIFI indicates potentially systemically
important insurers defined as those with total assets above $50B in at least one year during out sample period. TA is total assets at calendar year
end (AT Y in Compustat). Corporate bonds, GSE securities, Municipal bonds, US govt securities, RMBS, CMBS, ABS, Sovereign bonds, Equity,
and Mutual fund shares are the dollar-value percentages of an insurer’s portfolio invested in these asset classes.Number of securities is the number of
unique 9-digit CUSIPs in an insurer’s portfolio in a given year. Number of issuers is the number of unique issuers, identified using 6-digit CUSIPs, in
an insurer’s portfolio in a given year. Issuer concentration is a Herfindahl index constructed for each insurer in each year as the sum of the squared
weights of issuers in its portfolio. Asset class concentration is a Herfindahl index constructed for each insurer in each year as the sum of the squared
weights of asset classes in its portfolio. Issuer/asset class weights are calculated as the dollar amount invested in each issuer/asset class relative to the
total value of an insurer’s portfolio. Mean, medians and standard deviations are based on the cross-sectional variation of insurers’ time series average.
Table 1: Portfolio Composition Statistics
Table 2: Insurer Pair Statistics
The table presents descriptive statistics for our sample of insurer pairs during 2002 to 2014. PSIFI indicates
potentially systemically important insurers defined as those with total assets above $50B in at least one year
during our sample period. Similarity AC is a similarity measure based on the cosign similarity between a
pair of insurers’ asset class portfolio weights. Similarity AC (Similarity I) is a similarity measure based on
the cosign similarity between a pair of insurers’ asset class (issuer) portfolio weights. An issuer/asset class
weight is calculated as the dollar amount invested in that issuer/asset class relative to the total value of an
insurer’s portfolio. Ret Corr Pair is the annual correlation of daily holding-period returns calculated at year
end for each pair of insurers. TA Pair is the sum of total assets for a pair of insurers at a calendar year end.
YEARS 2002 - 2014
Whole Sample
Similarity AC
Similarity I
Ret Corr Pair
ln(TA Pair)
Non-PSIFI
PSIFI
Mean
Median
SD
Mean
Median
SD
Mean
Median
SD
0.68
0.09
0.32
11.37
0.74
0.05
0.31
11.36
0.23
0.13
0.22
1.90
0.70
0.11
0.25
9.34
0.76
0.06
0.22
9.60
0.22
0.14
0.21
1.07
0.67
0.09
0.49
13.21
0.73
0.05
0.50
13.3
0.24
0.11
0.18
0.95
YEAR 2002
Whole Sample
Similarity AC
Similarity I
Ret Corr Pair
ln(TA Pair)
Non-PSIFI
PSIFI
Mean
Median
SD
Mean
Median
SD
Mean
Median
SD
0.74
0.10
0.26
10.99
0.79
0.06
0.23
11.03
0.19
0.12
0.21
1.78
0.77
0.11
0.17
9.19
0.82
0.07
0.14
9.31
0.17
0.13
0.18
1.02
0.71
0.10
0.47
12.81
0.76
0.05
0.51
12.81
0.22
0.12
0.17
0.86
YEAR 2008
Whole Sample
Similarity AC
Similarity I
Ret Corr Pair
ln(TA Pair)
Non-PSIFI
PSIFI
Mean
Median
SD
Mean
Median
SD
Mean
Median
SD
0.67
0.08
0.45
11.51
0.74
0.05
0.46
11.44
0.24
0.10
0.17
1.86
0.69
0.09
0.39
9.35
0.75
0.05
0.42
9.60
0.23
0.11
0.19
0.99
0.67
0.08
0.54
13.19
0.74
0.05
0.55
13.21
0.24
0.09
0.12
0.99
YEAR 2014
Whole Sample
Similarity AC
Similarity I
Ret Corr Pair
ln(TA Pair)
Non-PSIFI
PSIFI
Mean
Median
SD
Mean
Median
SD
Mean
Median
SD
0.62
0.09
0.36
11.81
0.68
0.05
0.36
11.71
0.27
0.13
0.19
1.75
0.66
0.10
0.30
9.77
0.73
0.06
0.34
10.05
0.26
0.13
0.19
1.01
0.58
0.08
0.48
13.40
0.58
0.05
0.49
13.56
0.27
0.12
0.14
0.86
46
Table 3: Determinants of Portfolio Holdings Similarity
The table presents OLS estimation results for the sample of insurer pairs from 2002 to 2014. The dependent
variable is the portfolio similarity of a pair of insurers measured at the asset class or issuer level. All independent variables are defined in Appendix A. Robust t-statistics are in parentheses. Statistical significance
is denoted by ***, **, and * at the 1%, 5%, and 10% level respectively.
Asset Class Portfolio Similarity
All
Non-PSIFI
PSIFI
(1)
(2)
(3)
Life Pair
PC Pair
PSIFI Pair
Non-PSIFI Pair
0.180***
(17.36)
0.069***
(11.48)
0.013*
(1.91)
0.026***
(5.26)
Large Pair
Small Pair
Conc Pair AC
NonConc Pair AC
-0.035***
(-7.05)
0.092***
(20.88)
0.149***
(10.54)
0.077***
(10.70)
0.051***
(9.15)
-0.034***
(-4.54)
0.005
(0.41)
0.080***
(13.41)
0.171***
(16.01)
-0.023
(-1.15)
N
R2
-0.016*
(-1.96)
0.000
(0.07)
0.035***
(3.98)
0.059***
(4.67)
-0.061***
(-5.10)
0.119***
(11.14)
0.644***
(199.55)
Y
0.676***
(116.73)
Y
0.657***
(135.82)
Y
-0.001
(-0.23)
0.032***
(15.28)
0.085***
(58.64)
Y
41,011
0.171
16,046
0.156
5,579
0.200
41,011
0.030
NonConc I Pair
Year FE
0.004
(1.17)
0.019***
(6.33)
-0.000
(-0.22)
0.016***
(9.17)
-0.018***
(-8.76)
0.041***
(8.23)
Conc I Pair
Constant
Issuer Level Portfolio Similarity
All
Non-PSIFI
PSIFI
(4)
(5)
(6)
47
0.004
(0.70)
0.029***
(8.47)
0.106***
(35.87)
Y
-0.013***
(-3.10)
0.046***
(10.21)
0.076***
(25.97)
Y
16,046
0.044
5,579
0.093
Table 4: Pairwise Return Correlation and Portfolio Holdings Similarity
The table presents OLS estimation results for the sample of insurer pairs from 2002 to 2014. The dependent variable is the portfolio similarity of a pair of insurers measured at the asset class or issuer level.
Ret Corr Pair is the annual correlation between a pair of insurers’ daily holding-period returns during the
year. All independent variables are defined in Appendix A. Robust t-statistics are in parentheses. Statistical
significance is denoted by ***, **, and * at the 1%, 5%, and 10% level respectively.
All
(1)
Ret Corr Pair
Asset Class Portfolio Similarity
All
Non-PSIFI
PSIFI
(2)
(3)
(4)
-0.021
(-0.98)
Life Pair
PC Pair
PSIFI Pair
Non-PSIFI Pair
0.031
(1.51)
0.172***
(16.88)
0.055***
(9.56)
0.002
(0.23)
0.051***
(8.13)
Large Pair
Small Pair
Conc Pair AC
-0.021***
(-5.44)
0.083***
(18.49)
NonConc Pair AC
0.006
(0.25)
0.143***
(9.65)
0.072***
(8.30)
0.018**
(2.74)
-0.000
(-0.05)
0.042***
(3.88)
0.059***
(11.04)
0.095**
(2.74)
0.144***
(13.20)
-0.025
(-1.31)
All
(5)
Issuer Level Portfolio Similarity
All
Non-PSIFI
PSIFI
(6)
(7)
(8)
-0.050***
(-4.09)
N
R2
-0.121***
(-3.84)
0.033***
(3.91)
0.052***
(4.09)
-0.052**
(-2.94)
0.131***
(10.30)
0.701***
(57.88)
Y
0.639***
(105.64)
Y
0.688***
(117.68)
Y
0.619***
(33.24)
Y
0.110***
(17.05)
Y
-0.003
(-0.89)
0.029***
(16.92)
0.110***
(34.05)
Y
28,491
0.000
28,491
0.172
9,950
0.156
4,633
0.182
28,491
0.008
28,491
0.041
NonConc Pair I
Year FE
-0.027**
(-2.29)
-0.009
(-1.22)
-0.021***
(-6.08)
-0.020***
(-6.11)
0.057***
(10.24)
Conc Pair I
Constant
-0.079***
(-7.32)
0.006*
(1.83)
0.004
(1.42)
0.014***
(5.87)
0.008***
(3.74)
48
0.007
(1.42)
0.024***
(7.47)
0.110***
(38.07)
Y
-0.014**
(-2.65)
0.042***
(7.86)
0.141***
(11.13)
Y
9,950
0.070
4,633
0.104
Table 5: Sales Statistics
The table presents the composition of insurers’ sales as percentage of total sales from 2002 to 2014.
Asset Class
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
Corporate Bonds
ABS
CMBS
RMBS
US Govt Securities
GSE Securities
Sovereign bonds
Equity
Municipal Bonds
Mutual Funds
22.3
1.3
1.1
1.5
13.8
26.7
1.5
28.1
3.8
0.0
22.2
1.4
1.0
1.4
15.3
31.4
2.6
20.5
4.3
0.0
20.0
1.8
1.6
1.5
14.0
28.3
1.9
22.5
5.6
2.8
21.3
1.9
1.7
1.7
18.1
23.7
2.6
20.9
5.1
3.1
21.6
2.0
2.1
1.8
16.8
21.9
1.4
22.4
5.1
4.9
17.0
1.6
1.9
1.6
16.6
24.4
1.5
24.2
5.7
5.4
16.5
0.9
1.4
1.2
13.8
30.9
1.0
19.4
5.1
9.7
19.1
1.1
1.8
1.9
17.1
27.1
1.1
19.0
3.8
8.0
16.7
1.3
2.0
1.3
17.9
28.0
2.1
18.0
3.8
8.8
16.8
1.1
1.2
0.8
16.7
30.9
1.1
17.0
3.1
11.3
19.2
1.4
1.7
1.8
14.9
32.2
2.5
13.0
1.9
11.4
18.5
1.3
1.3
1.1
17.4
31.2
2.2
13.6
2.4
11.0
17.2
1.8
1.1
0.8
19.0
28.2
1.7
14.1
2.6
13.5
49
Table 6: Determinants of Sales Similarity
The table presents OLS estimation results for the sample of insurer pairs from 2002 to 2014. The dependent
variable sales similarity of a pair of insurers measured at the asset class or issuer level. All independent
variables are defined in Appendix A. Robust t-statistics are in parentheses. Statistical significance is denoted
by ***, **, and * at the 1%, 5%, and 10% level respectively.
Asset Class Sales Similarity
All
Non-PSIFI
PSIFI
(1)
(2)
(3)
Life Pair
PC Pair
PSIFI Pair
Non-PSIFI Pair
0.118***
(19.12)
0.015
(1.67)
0.063***
(13.21)
-0.009
(-1.67)
Big Pair
Small Pair
Conc Pair AC
NonConc Pair AC
0.022***
(3.04)
0.022***
(4.75)
0.127***
(13.02)
0.042***
(3.54)
0.047***
(6.25)
-0.033***
(-5.31)
0.017*
(1.79)
0.016**
(2.56)
0.101***
(10.17)
-0.058***
(-3.42)
Issuer Level Sales Similarity
All
Non-PSIFI
PSIFI
(4)
(5)
(6)
-0.016***
(-4.76)
-0.010***
(-3.07)
0.083***
(15.77)
-0.029***
(-20.09)
-0.011*
(-1.73)
-0.010
(-0.71)
0.004**
(2.65)
-0.004**
(-2.11)
0.010
(1.23)
0.057***
(6.31)
Conc Pair I
Year-Quarter FE
0.532***
(160.64)
Y
0.564***
(141.84)
Y
0.535***
(121.56)
Y
0.017***
(5.01)
-0.007***
(-4.36)
0.065***
(57.05)
Y
N
R2
148,951
0.067
58,035
0.086
20,403
0.074
149,098
0.059
NonConc Pair I
Constant
-0.012***
(-4.01)
-0.014***
(-5.37)
50
0.012***
(4.51)
-0.009***
(-5.04)
0.052***
(40.33)
Y
0.040***
(5.19)
-0.013**
(-2.31)
0.094***
(42.95)
Y
58,075
0.015
20,432
0.046
Table 7: Predicting Sales Similarity with Portfolio Similarity
The table presents OLS estimation results for the sample of insurer pairs from 2002 to 2014. The dependent
variable is insurers’ quarterly pairwise sales similarity at the asset class and issuer level. All independent
variables are defined in Appendix A and are measured as of the year-end prior to the sales quarter. Robust
t-statistics are in parentheses. Statistical significance is denoted by ***, **, and * at the 1%, 5%, and 10%
level respectively.
Asset Class Sales Similarity
All
Non-PSIFI
PSIFI
(1)
(2)
(3)
Similarity AC
0.296***
(17.81)
0.275***
(13.81)
0.299***
(10.12)
Similarity I
Life Pair
PC Pair
PSIFI Pair
Non-PSIFI Pair
0.061***
(7.51)
0.003
(0.34)
0.057***
(11.94)
-0.018***
(-3.51)
Big Pair
Small Pair
Conc Pair AC
NonConc Pair AC
0.042***
(6.44)
-0.010**
(-2.63)
0.077***
(6.50)
0.022*
(1.85)
0.036***
(4.67)
-0.023***
(-3.33)
0.036***
(3.73)
-0.018***
(-3.54)
Issuer Level Sales Similarity
All
Non-PSIFI
PSIFI
(4)
(5)
(6)
0.043***
(3.84)
-0.050***
(-2.72)
0.438***
(23.20)
-0.016***
(-5.40)
-0.018***
(-4.72)
0.085***
(15.12)
-0.037***
(-24.13)
0.686***
(18.14)
-0.019**
(-2.46)
-0.060***
(-4.26)
0.010***
(5.06)
-0.015***
(-7.05)
0.026***
(3.26)
0.023**
(2.29)
Conc Pair I
Year-Quarter FE
0.280***
(26.93)
Y
0.312***
(21.15)
Y
0.315***
(15.22)
Y
0.019***
(5.49)
-0.021***
(-7.80)
0.041***
(21.23)
Y
N
R2
129,944
0.097
49,635
0.104
18,352
0.110
130,008
0.156
NonConc Pair I
Constant
0.340***
(14.58)
-0.008***
(-4.03)
-0.011***
(-4.28)
51
0.009***
(3.49)
-0.016***
(-7.04)
0.024***
(8.39)
Y
0.055***
(7.58)
-0.054***
(-8.47)
0.061***
(18.95)
Y
49,674
0.127
18,352
0.134
Table 8: Sales Similarity and Return Correlation
The table presents OLS estimation results for the sample of insurer pairs from 2002 to 2014. The dependent
variable is insurers’ sales similarity at asset class and issuer level for quarters Q1 to Q4 in year t + 1. Return
correlation is measured in the quarter prior to the sales. All independent variables are defined in Appendix
A and are measured as of the year-end prior to the sales quarter. Robust t-statistics are in parentheses.
Statistical significance is denoted by ***, **, and * at the 1%, 5%, and 10% level respectively.
All
(1)
Similarity AC
Asset Class Sales Similarity
All
Non-PSIFI
PSIFI
(2)
(3)
(4)
0.309***
(14.81)
0.327***
(12.54)
All
(5)
Life Pair
PC Pair
PSIFI Pair
Non-PSIFI Pair
0.057***
(3.11)
0.116***
(16.04)
0.037***
(3.60)
0.061***
(10.33)
-0.002
(-0.43)
0.041**
(2.13)
0.063***
(7.21)
0.021**
(2.16)
0.058***
(9.68)
-0.014**
(-2.49)
Big Pair
Small Pair
Conc Pair AC
NonConc Pair AC
0.035***
(4.68)
0.018***
(4.25)
0.040***
(5.50)
-0.005
(-1.21)
-0.015
(-0.66)
0.086***
(6.48)
0.066***
(3.92)
0.037***
(3.78)
-0.020**
(-2.10)
0.022*
(2.01)
-0.016***
(-3.14)
0.087**
(2.42)
0.036***
(2.99)
-0.052***
(-2.99)
N
R2
0.325***
(10.25)
0.022***
(2.93)
0.002
(0.74)
-0.009***
(-4.18)
0.702***
(17.74)
-0.020
(-0.72)
-0.015*
(-1.70)
-0.070***
(-5.06)
0.016
(1.52)
0.031***
(2.72)
0.453***
(83.34)
Y
0.254***
(17.40)
Y
0.260***
(13.51)
Y
0.297***
(15.07)
Y
0.020***
(4.89)
-0.010***
(-4.33)
0.077***
(38.98)
Y
91,984
0.103
91,984
0.066
32,038
0.127
14,983
0.092
92,035
0.168
NonConc Pair I
Year-Quarter FE
-0.008
(-0.99)
-0.002
(-0.64)
-0.018***
(-4.42)
0.103***
(15.93)
-0.030***
(-21.76)
0.438***
(19.96)
0.022**
(2.62)
-0.005
(-1.63)
-0.020***
(-5.05)
0.098***
(14.57)
-0.034***
(-24.88)
0.005**
(2.03)
-0.015***
(-4.77)
Conc Pair I
Constant
PSIFI
(8)
0.261***
(7.81)
Similarity I
Ret Corr Pair
Issuer Level Sales Similarity
All
Non-PSIFI
(6)
(7)
52
0.023***
(5.74)
-0.022***
(-8.43)
0.032***
(11.15)
Y
0.011***
(3.60)
-0.015***
(-6.23)
0.020***
(4.48)
Y
0.052***
(6.85)
-0.052***
(-8.03)
0.077***
(5.58)
Y
92,035
0.080
32,067
0.123
14,983
0.138
Table 9: Sales Similarity and RBC Ratios
The table presents OLS estimation results for the sample of insurer pairs from 2002 to 2014. The dependent
variable is insurers’ quarterly pairwise sales similarity at the asset class and issuer level. The two pairwise
indicator variables, (RBC High Pair ) and (RBC Low Pair ) are equal to one if both insurers are above the
median RBC ratio for the sample or under the median RBC ratio, respectively. All independent variables
are defined in Appendix A and are measured as of the year-end prior to the sales quarter.Robust t-statistics
are in parentheses. Statistical significance is denoted by ***, **, and * at the 1%, 5%, and 10% level
respectively.
Asset Class Sales Similarity
All
Non-PSIFI
PSIFI
(1)
(2)
(3)
Similarity AC
Similarity AC*RBC High Pair
Similarity AC*RBC Low Pair
0.233***
(12.44)
-0.012
(-0.78)
0.099***
(4.16)
0.214***
(9.06)
-0.033
(-1.43)
0.153***
(4.79)
0.218***
(6.69)
-0.027
(-0.84)
0.094***
(2.86)
Similarity I
Similarity I*RBC High Pair
Similarity I*RBC Low Pair
RBC High Pair
RBC Low Pair
Life Pair
PC Pair
PSIFI Pair
Non-PSIFI Pair
0.003
(0.21)
-0.065***
(-3.91)
0.065***
(6.83)
0.019*
(1.73)
0.053***
(10.74)
-0.018***
(-3.31)
Big Pair
Small Pair
Conc Pair AC
NonConc Pair AC
0.042***
(5.77)
-0.012***
(-2.88)
0.031*
(1.72)
-0.126***
(-5.80)
0.076***
(5.88)
0.026*
(1.96)
0.035***
(4.80)
-0.030***
(-4.29)
0.043***
(4.26)
-0.021***
(-3.62)
Issuer Level Sales Similarity
All
Non-PSIFI
PSIFI
(4)
(5)
(6)
-0.010
(-0.45)
-0.045*
(-1.72)
0.043***
(3.13)
-0.077***
(-3.64)
0.444***
(22.13)
0.055
(1.59)
-0.014
(-0.54)
-0.007***
(-3.30)
0.002
(0.74)
-0.019***
(-5.78)
-0.019***
(-5.17)
0.075***
(14.12)
-0.034***
(-20.91)
0.014
(1.49)
0.019
(1.64)
Year-Quarter FE
0.333***
(27.44)
Y
0.369***
(21.66)
Y
0.362***
(16.35)
Y
0.017***
(5.12)
-0.021***
(-7.58)
0.044***
(20.78)
Y
N
R2
100,115
0.094
39,271
0.107
13,496
0.107
100,167
0.151
NonConc Pair I
53
0.721***
(16.67)
0.063
(1.05)
-0.223***
(-4.66)
-0.001
(-0.18)
0.007
(0.82)
-0.020**
(-2.30)
-0.092***
(-4.25)
0.009***
(4.20)
-0.016***
(-6.67)
Conc Pair I
Constant
0.366***
(14.84)
0.000
(0.01)
0.005
(0.13)
0.000
(0.25)
0.002
(1.00)
-0.010***
(-3.77)
-0.008**
(-2.52)
0.008***
(2.89)
-0.017***
(-6.91)
0.023***
(7.03)
Y
0.048***
(6.21)
-0.051***
(-8.08)
0.071***
(17.55)
Y
39,303
0.136
13,496
0.126
Table 10: Sales Similarity, Liquidity, Downgrades and the Crisis
The table presents OLS estimation results for the sample of insurer pairs from 2002 to 2014. The dependent
variable is insurers’ quarterly pairwise sales similarity at the asset class and issuer level. Similarity I Liquid is
the portfolio similarity of issuers that fall within the following asset classes that are considered liquid: equity,
mutual fund shares, US government securities, GSE securities, and sovereign bonds. Similarity I Illiquidis the
portfolio similarity of issuers that fall within the following asset classes that are considered illiquid: corporate
bonds, municipal bonds, RMBS, CMBS and ABS. Similarity Downgraded is the portfolio similarity of issuers
that are downgraded in the following year. Similarity NotDowngraded is the portfolio similarity of issuers
that are not downgraded in the following year. Crisis is a dummy variable equal to 1 for the years 2007 to
2009 and Post-Crisis is a dummy variable equal to 1 for the years 2010 to 2014. All independent variables are
defined in Appendix A and are measured as of the year-end prior to the sales quarter. Robust t-statistics are
in parentheses. Statistical significance is denoted by ***, **, and * at the 1%, 5%, and 10% level respectively.
Similarity I Illiquid
Similarity I Illiquid*Crisis
Similarity I Illiquid*Post-Crisis
Similarity I Liquid
Similarity I Liquid*Crisis
Similarity I Liquid*Post-Crisis
All
(1)
Non-PSIFI
(2)
0.235***
(9.47)
0.064**
(4.97)
0.012
(0.34)
0.316***
(13.91)
-0.037
(-1.07)
-0.029
(-1.44)
0.077**
(5.14)
0.010
(0.40)
0.034
(1.19)
0.286***
(12.52)
-0.035
(-0.97)
-0.067*
(-2.62)
Issuer Level Sales Similarity
PSIFI
All
Non-PSIFI
(3)
(4)
(5)
0.443***
(7.60)
0.014
(0.79)
-0.114
(-1.60)
0.484***
(23.40)
-0.204**
(-5.62)
0.063
(0.83)
Similarity I Downgraded
Similarity I Downgraded*Crisis
Similarity I Downgraded*Post-Crisis
Similarity I NotDowngraded
Similarity I NotDowngraded*Crisis
Similarity I NotDowngraded*Post-Crisis
Crisis
Post-Crisis
Life Pair
PC Pair
PSIFI Pair
Non-PSIFI Pair
-0.005
(-1.51)
-0.004
(-0.62)
-0.018**
(-5.36)
-0.005
(-1.52)
0.070***
(6.51)
-0.032***
(-18.56)
Big Pair
0.023***
(6.46)
-0.024***
(-9.28)
0.014**
(4.59)
Y
0.008**
(4.89)
-0.016***
(-8.88)
0.012***
(7.96)
-0.017***
(-9.87)
0.008**
(4.19)
Y
129,771
0.165
54
0.113
Small Pair
Conc Pair I
NonConc Pair I
Constant
Quarter FE
N
R2
-0.002
(-0.59)
-0.004
(-0.61)
-0.008***
(-6.30)
-0.007*
(-2.75)
49,674
PSIFI
(6)
0.026*
(3.16)
0.011
(1.09)
-0.017
(-1.32)
-0.011
(-1.01)
0.157***
(11.42)
0.076***
(12.70)
0.082***
(6.93)
0.464***
(30.68)
-0.049
(-1.30)
-0.103**
(-4.46)
-0.006*
(-2.55)
-0.004
(-1.07)
-0.018***
(-6.19)
-0.007
(-2.11)
0.055**
(5.71)
-0.027***
(-20.19)
0.068***
(17.53)
0.010
(0.65)
0.018*
(2.78)
0.378***
(19.64)
-0.023
(-0.46)
-0.088*
(-3.17)
-0.002
(-0.45)
0.001
(0.35)
-0.009***
(-10.19)
-0.010**
(-4.48)
0.262***
(7.10)
0.076**
(3.60)
0.136***
(7.19)
0.643***
(26.99)
-0.232***
(-13.95)
-0.062
(-0.74)
0.014
(1.07)
-0.029**
(-5.13)
-0.013
(-1.09)
0.025**
(3.31)
0.045**
(4.18)
-0.043***
(-12.23)
-0.003
(-0.21)
Y
18,043
0.207
0.053**
(4.43)
-0.046***
(-8.24)
0.016
(1.58)
Y
0.022***
(6.80)
-0.025***
(-10.56)
0.006
(1.71)
Y
0.007**
(5.31)
-0.014***
(-8.57)
0.012***
(7.53)
-0.018***
(-8.99)
0.003
(1.21)
Y
18,267
0.156
125,179
0.195
47,254
0.131