Mutual Fund Holdings of Credit Default Swaps: Liquidity Management and Risk Taking1 Wei Jiang2 Columbia Business School Zhongyan Zhu3 CUHK Business School, The Chinese University of Hong Kong First draft: October, 2014 This draft: April, 2015 1 We thank Martin Oehmke for valuable comments and suggestions. Wei Jiang can be reached at wj2006@columbia.edu. 3 Corresponding author. Zhongyan Zhu can be reached at zhongyan@cuhk.edu.hk. 2 1 Mutual Fund Holdings of Credit Default Swaps: Liquidity Management and Risk Taking ABSTRACT Using a comprehensive dataset of mutual funds’ quarterly holdings of credit default swap (CDS) contracts in 2007-2009, we analyze the motivation for and consequences of mutual funds’ participation in the CDS market, especially during a financial crisis. Consistent with Oehmke and Zawadowski (2014), funds resort to CDS (especially selling) when they face unpredictable liquidity needs and when the CDS securities are themselves are liquid relative to the underlying bonds. Smaller funds followed leading funds in initiating the selling or speculative buying of CDS contracts on new reference entities, consistent with the “tail risk taking” in Rajan (2006). Moreover, the reference entities that attracted the most selling interest from the large mutual funds are disproportionately firms that were perceived to be “too large to fail” or “too systemic to fail,” thus indicating that the use of CDS by mutual funds heightened the correlation in financial health among the large and systemically important financial institutions. 2 By 2007, the CDS market had grown to over $60 trillion in total notional value, and more than 60 percent of all fixed income mutual funds in the U.S. had some CDS positions (see Jarrow (2011) for a review of the market). Despite the popularity of credit default swaps as a synthetic security to gain or hedge credit exposure and the numerous research studies examining the effect of CDS contracts on the issuers’ borrowing costs, there has been little empirical research on how and why individual funds initiate and hold CDS contracts on individual reference entities. Due to data availability, most empirical research based on mutual fund holdings has focused on their long positions in equity and bonds. A comprehensive study on their use of CDS contracts provides us a more complete understanding of mutual funds’ investment strategies, and more importantly, their impact on the functioning and stability of the financial markets. Since the financial crisis in the late 2000s, there have been renewed concerns over whether the asset management industry could be a destabilizing force in markets. In January 2014, the Financial Stability Board (FSB)—an international organization aimed at preventing financial crises—proposed that some large fund managers might need to be designated “systematically important financial institutions” (SIFIs), which would require them to be subject to heavier regulation. In the meanwhile, the SEC is also preparing rules to request more portfolio data from large asset managers and to conduct stress tests. The rules that have been discussed thus far are similar to the post-crisis requirements put in place for big banks and other large financial institutions that regulators believe could pose a risk to the financial system and broader economy if they were to collapse. In the context of the asset management industry, the worry is mainly two-fold. First, if there is a herding pattern in risk-taking among fund managers, this behavior might lead to a general sell-off, such as that which occurred with mortgage securities in 2008. This concern is amplified by the fact that the fund 3 management industry is becoming more and more concentrated.4 Second, although mostly unlevered, market participants like mutual funds may become the locus of potential financial instability because they are largely motivated by relative performance (Feroli, Kashyap, Schoenholtz, and Shin, 2014). The potential consequences of mutual funds’ increasing use of derivatives, especially CDS contracts, are at the center of this debate because of their unique features that allow fund managers to take “hidden tail risk” and to mimic the behavior of leading funds, both of which were succinctly and presciently summarized in Rajan (2006). Credit default swaps, like many other derivative contracts, allow funds to gain leveraged returns compared to transactions that occur directly in the underlying securities. Funds may utilize CDSs to seek incremental returns by buying or selling credit protection on a the bond of a reference entity, which will often require less capital to create the same return profile than buying or selling a comparable bond of the same issuer. The incremental returns from selling CDS come at the cost of a “hidden tail risk” that is similar to selling disaster insurance and that is usually not captured in the benchmark. CDS selling allows fixed income funds to have the appearance of producing high alphas (returns adjusted by the conventional notions of risk). Since the true performance can only be assessed over a period that is much longer than the typical horizon set for the average fund manager or implied by the latter’s expected tenure and incentive schemes, managers will have an incentive to take such risk. Every once in a while, such tail risk manifests itself in disastrous returns, which often coincide with a systemic meltdown. Such hidden tail risk could “spread” given the relative performance-based incentives in the mutual fund industry, which are explicit in most incentive contracts for portfolio managers and implicit in the strong and convex flow-to-performance responses (Chevalier and Ellison, 1997). After some entrepreneurial mutual funds venture into the relatively new CDS world and produced superior returns during the “good time,” the relative performance evaluation created the incentive for other funds to herd 4 According to Morningstar (a leading investment research firm), at the end of 2012, the top five mutual fund complexes managed 48 percent of total assets of equity funds and 53 percent of fixed income funds. The same numbers for the top 25 mutual fund complexes are both around 75 percent. 4 into the same market or even to the same reference entities. Such herding into the same tail risk space could create amplified and wide-spread losses for many mutual funds at the same time, destabilizing the financial system due to both “runs” (or run-like behavior) on the funds and price pressure on the securities market. We observe episodes of such nature in the “breaking the buck” period in 2008 (when a lot of the money-market mutual funds in the U.S. came close to having their net asset value break the $1 par, while investors had considered money market funds to be virtually riskless). We note that Chen, Goldstein, and Jiang (2010) documented a milder version of “runs” on the funds. Analyzing a comprehensive dataset of CDS quarterly holdings by all U.S. fixed income mutual funds from 2007 through 2009, we find evidence supporting both hypotheses. First, mutual funds as a whole were net sellers of CDS, where the total selling notional amount during the sample period ($141 billion) exceeded the buying notional amount by 41 percent. Nevertheless, the sell-buy skew was concentrated at the top— Pacific Investment Management Company, LLC (PIMCO), the largest fixedincome mutual fund complex, sold 76 percent more CDS contracts than it purchased, while the smallest 70% of mutual fund families were net buyers of CDS. This contrast indicates that tail risks tended to be concentrated among the largest funds which are also the most important to the stability of the financial markets. Second, we document a strong lead-follow pattern in CDS trading between PIMCO and the smaller funds. In particular, the probability that a smaller mutual fund initiated selling CDS on a new reference entity doubles after PIMCO disclosed a large selling position in the same reference entity in the previous period. A similar effect also exists in CDS buying, but the lead-follow connection is noticeably weaker. The tendency to herd seems to be stronger in taking on more risk rather than in buying protection. What potentially further aggravates the tail risk and herding effects is our finding that mutual funds systematically bet on institutions that were perceived as “too big to tail” or “too systematic to fail.” That is, mutual funds, and especially the large funds, demonstrated a strong preference for the sale, but not the purchase, of CDS positions in large reference entities. Among the large reference entities, the 5 effect was most notable in the “systematically important financial institutions.” Our finding thus suggests that mutual funds contributed to the fragility of the financial system by their heightened correlation with the financial health of the SIFIs due to their CDS positions. Stultz (2010) attributed the bail-out of financial institutions forced upon the taxpayers to a “web of linkage across financial institutions” with derivatives, especially credit default swaps. Our study also provides a concrete picture of such a web involving mutual funds. Indeed, thanks to the government bailout, the mutual funds in our sample ironically came out of the financial crisis mostly unscathed despite a period during which their CDS selling positions incurred colossal losses on paper. A safe landing in a time of rare opportunity when the tail risk was supposed to take its toll and to exert discipline no doubt makes it even more difficult to prevent asset managers from adopting the same strategy that generates benefits to the funds/managers but exerts negative externalities to the financial system. Our study thus offers a justification for the largest asset managers to be designated as SIFIs despite that they do not take risk directly with their own capital, the most commonly used defense by parties who oppose such a proposal.5 Our study also provides strong empirical support for the theoretical research on the difference between trading in CDS and in the underlying bonds (Oehmke and Zawadowski (2013a)). We find that mutual funds with more volatile fund flows, and hence more frequent trading needs in the fixed income market, are more likely to substitute long positions in underlying bonds with short positions in CDS to take advantage of the relatively higher trading liquidity. The economic magnitude of this relation is sizable, a one-standard deviation increase in the volatility of funding flows raises the propensity in CDS selling by about 40%. Moreover, higher CDS trading liquidity predicts more CDS usage, and significantly more so for CDS selling than for buying. a one-standard deviation increase in the proxy for CDS liquidity doubles the likelihood of CDS selling at the fund-issuer-period level, but only increase that of CDS buying by 10%. On the other hand, higher bond trading liquidity predicts more CDS buying but 5 See, for example, “Fund managers: Assets or liabilities,” The Economist, August 2, 2014. 6 not selling. The asymmetry reflects the distinct motives for CDS selling vs. buying. While CDS selling substitutes bond buying (hence CDS, rather than bond, liquidity is important), CDS buying complements bond buying due to the predominant hedging motive (hence both CDS and bond liquidity explains the long positions). Empirical research on institutional investors’ CDS investments is relatively scant (see a recent survey by Augustin, Subrahmanyam, Tang, and Wang, 2014). The only other paper in finance that analyzes the use of credit default swaps is Adam and Guettler (2012). While the data and sample have some overlap, the two papers ask quite different research questions. Adam and Guettler (2012) focus on how mutual funds use CDS to increase fund risk in order to game the convex incentive implied by the tournament, and their analysis is mostly based on CDS holdings aggregated at the fund level. Our study instead explores how mutual funds choose to buy and sell CDS contracts on different reference entities, how their investment strategy fits the theoretical predictions regarding both liquidity management and risk taking, and how the funds’ individual risk taking interacts with system-wide risk. I. Institutional Background and Sample Overview A. Institutional background In our study, a “mutual fund” is an investment company registered under the 1940 Investment Company Act and could be either an open-end or a closed-end fund. CDS are now commonly held by fixed-income mutual funds despite the fact that derivatives traditionally did not make up a significant portion of fund holdings (Koski and Pontiff, 1999). A mutual fund may buy CDS, in which case it seeks protection by paying a yearly premium until a pre-defined credit event occurs or until the contract expires; or it may sell CDS, in which case it receives the premium but assumes the loss in case of insolvency. When a mutual fund sells CDS, it receives credit exposure to the reference entity without holding the underlying bonds—that is, it creates a synthetic bond that enhances the yields on the bond that the CDS 7 protects. When a mutual fund buys CDS, it may do so to offset some of the credit risk it takes in its long bond positions, or to bet on its pessimistic view about the financial health of the reference entity. There is no legal restraint specifically targeting CDS holdings by mutual funds. Under the Investment Company Act of 1940, the general restrictions that could potentially apply on CDS positions come from four sources. First, CDS positions usually count toward the limit on total illiquid investments made by a fund (no more than 15% of all investments). Second, the embedded leverage in CDS contracts subject them to the aggregate limit on a fund’s actual and implied leverage (up to 300% of asset coverage). Third, the diversification requirement prohibits concentrated single counterparty exposure (below 5% of total assets). And finally, the full commitment requirement states that the notional amount of total derivatives may not exceed 100% of the total value of the fund. Given that the market value of CDS contracts are at or close to zero at initiation and are a small percentage of the notional amount in all but the most extreme cases, these restrictions are quite modest and were never binding in our sample period (and are very unlikely to be binding in general). B. Data collection Figure 1 shows the structure of the data collected from the applicable security filings. Mutual funds are first organized in “families” containing a group of funds that are sponsored by the same investment management firm, such as PIMCO. Mutual funds are required to file Form N-Q semiannually with the SEC to disclose their complete portfolio holdings. In addition, the mandatory filings of annual and semi-annual reports, Form N-CSR, to shareholders also contain the funds’ securities holdings. The two types of filings span all four quarters in a year. Both forms are filed at one level below, or at the “series trust” or “shared trust” level, such as “PIMCO Funds.” A series trust is a legal entity consisting of a complex of independently managed funds that have the same sponsor, that share distribution and branding efforts, and that often have unitary (or overlapped) boards. An N-Q or N-CSR filing contains detailed portfolio information (usually recorded at the quarter end) for each fund which represents a 8 distinct portfolio. CDS positions are disclosed in these original forms but are not included in most processed commercial databases such the Thomson Reuters Ownership database, and are thus available only via manual collection. Our sample construction starts with a search of all N-Q and N-CSR filings on the SEC EDGAR servers for portfolio with period-end dates in 2007 and 2009. The three-year period was chosen to capture the effect of the financial crisis. For each filing, we identify CDS positions using the following searching keywords: “Credit Default”, “Default Swap”, “CDS”, “Default Contract” and “Default Protection” following Adam and Guettler (2012). Given the purpose of our research, it is reasonable to exclude “accidental” CDS users. Hence we apply a filter that requires a fund (portfolio) to have at least 200 CDS positions or have a total notional amount of $400 million in order to be included in our sample. Such a filtered search results in 66,495 CDS holding positions on single name entities in 284 funds in 60 trust series (filers) affiliated with 33 fund families. From the portfolio disclosure we are able to record, for each CDS position, the reference entity, the counterparty, the notional amount, and whether the position was a buy or a sell. We also retrieve fund-level information, such as the total net assets (TNA) from the same source. Using the CUSIP as well as the names of the funds and their affiliated families, we are able to obtain (or construct) more fund-level variables such as the open-/closed-end status, returns, and fund flows from CRSP and Morningstar. [Insert Figure 1 here.] While our key data source is similar to that used in Guettler and Adam (2012), the samples in the two studies are constructed in quite different ways. Instead of aggregating CDS holdings for each fund in each period as in Guettler and Adam (2012), we focus on individual CDS positions and collect more detailed information about the individual holdings (notional amount, buy/sell, etc.) as well as the reference entities (size of the firm and its CDS spreads, etc.) Moreover, our sample contains all mutual funds that are regularly hold CDS position rather than just focus on the top fixed income funds. The 9 wider spectrum of funds allows us to explore different incentives and behavior among large, medium, and small mutual funds in the CDS market. C. Sample overview Table 1 presents an overview of our sample. Panel A shows the quarterly time-series patterns of CDS holdings at the fund family, series trust, and fund level. During the three year period, the notional amount of mutual fund single-entity CDS holdings increased from $13.1 billion at the beginning of 2007 to a peak of $29.2 billion in the second quarter of 2008 before descending smoothly to $16.0 billion by the end of 2009. About 58% of the positions involve a sale of CDS, indicating that, on the whole, mutual funds use CDS to seek additional risk taking. This pattern confirms a similar finding in Adam and Guettler (2012). [Insert Table 1 here.] The aggregate statistics, however, mask the huge cross-sectional differences. The distribution of CDS positions are highly skewed among funds in multiple dimensions. For ease of discussion, we divide all 33 mutual fund families in our sample into three tiers sorted by total CDS notional amount —leader (the PIMCO fund complex), large (the next nine mutual funds families after PIMCO; henceforth, the “Next 9”), and small fund families (the remaining 23 fund families in our sample; henceforth, the “Rest 23”). Panel B shows the breakdown by mutual fund families. PIMCO is the unambiguous leader among fixed-income investment managers. Its funds account for 62% of the total notional amount of CDS contracts held by all mutual funds. The “Next 9” group account for 24% of the total CDS holdings. The remaining 14% goes to the “Rest 23” families. Concentration of CDS usage among the top/large players aside, the sell-buy imbalance also exhibits an opposite pattern between the large and small players: while 63.8% of PIMCO’s CDS positions are on the short side (i.e., they sell protection), the same figures for the “Next 9” and the “Rest 23” are 57.7% and 35.2%, respectively. Though it has been known that mutual funds as a group were net 10 sellers of CDS during the sample period, it is worth pointing out that the net selling of single-entity CDS contracts by mutual funds is driven by the largest players and is actually not the typical behavior among individual funds. In fact, the “median” fund (among the “Rest 23”) bought more credit protection than it sold. Panel C of Table 1 demonstrates concentrated CDS activities along a different dimension: the underlying reference entities. Out of the 443 reference entities, the top 50 (100) account for 64% (78%) of the total notional amount. Moreover, large financial institutions constitute a disproportionately large share among the top reference entities: There are 27 large financial institutions among the top 50 reference entities, and among these 27, eight were designated as “Global Systemically Important Financial Institutions” (G-SIBs) by the FSB in 2009 and 2010. If fact, six of the eight G-SIBs headquartered in the U.S. are among the top 50 reference entities.6 Thus, mutual funds, especially the largest of them, could potentially incur large losses if those financial institutions fail. To the extent that mutual funds are among the primary forms of savings and investments for households, the pattern suggests that the mutual fund industry tends to make the “systemically important” financial institutions even more so. Panel D of Table 1 classifies mutual funds’ investment in CDS contracts into different categories based on the underlying purpose of the CDS contract. There is not a norm in the literature in inferring the purpose of CDS trading based on periodic holdings data. We classify the disclosures such that an “offsetting” buy is a CDS long position that can be matched to a sell position on the same reference entity by the same fund in the same quarter (on the same N-Q filing). A pair of offsetting positions is usually used to bet on the slopes of the term structure. A “hedging” buy represents CDS long positions where the same fund has a long position in the same underlying bonds during the same period. We only match positions rather than the exact amount in classifying offsetting and hedging purchases. Finally, a “speculative” buy represents the remaining long positions on which the funds would profit from the 6 The eight U.S. headquartered G-SIBs are: Bank of America Corp, Citigroup Inc, Goldman Sachs Group Inc, JP Morgan Chase & Co, Morgan Stanley, and Wells Fargo & Co. 11 financial failures of the reference entities. Panel D indicates that while large players (PIMCO in particular) sell more CDS than they buy; they are less speculative within their long CDS positions. For example, while 17% of PIMCO’s buy positions are classified as speculative, the same proportion for the “Rest” group is 56%. All panels in Table 1 combined indicate that the aggregate statistics of mutual funds’ holdings of CDS contracts generally do not reflect the behavior of typical funds because of the different strategies utilized by the large (especially funds from the leading fund family, PIMCO) and small funds. Overall, PIMCO funds use CDS contracts to seek levered credit exposure in large companies, and in disproportionately large and systematically important financial situations. Funds from small mutual fund families, other the other hand, are net buyers of CDS protection and therefore reduce their credit exposure using the derivatives. The behavior of the middle group is somewhere in between the two extremes, but has more similarity to the strategy used by PIMCO. Figure 2 displays the conditional empirical distribution of selling and buying intensity at the fund level, defined as the notional amount of net selling aggregated over all single reference entity contracts, scaled by the funds’ total net assets. When we equal-weight all funds, the distribution appears to be quite symmetric, with both the average and median close to zero at –0.03 percent and –0.02 percent, respectively. To reconcile our summary statistics with those in Adam and Guettler (2012), we compute the average total CDS notional amount and net selling intensity among the top 100 fixed income funds to be 3.84 percent and –0.84 percent, both of which are smaller than the equivalent numbers in Adam and Guettler (2011) (6.16 percent and –1.67 percent). The difference is mainly due to the inclusion of CDS positions on index products and sovereign debts in their sample. [Insert Figure 2 here.] The summary statistics of the main variables used in the analyses, both at the fund-quarter level, and fund-reference entity-quarter level, are reported in Table 2. [Insert Table 2 here.] 12 II. Mutual Fund CDS Holdings: Liquidity Management This section analyzes the first motivation for mutual funds to take CDS positions: liquidity management and risk raking. The CDS market allows mutual funds from large and median sized fund families to gain additional exposure to corporate credit risk. The same exposure could be accomplished by simply investing in the underlying bonds. Oehmke and Zawadowski (2014b) propose that funds have an important liquidity incentive to choose CDS over the bonds issued by the reference entities in order to obtain roughly equivalent credit exposure. In particular, if bonds are relatively illiquid relative to CDS contracts, then investors with more need for liquidity-driven trades should have a preference for short positions in CDS over long positions in the underlying bonds given the prevailing bond-CDS basis in the market. In all analyses, the subscripts i, j, t serve as indices for fund, CDS reference entity, and time period, respectively. A. Fund Level Analysis First, we assess the relation between CDS selling intensity and two proxies for mutual funds’ liquidity needs at the fund level. Results are reported in Table 3. In Panel A, the dependent variable in the analyses is a dummy variable for a fund to hold any CDS position during a period. The relevant sample is thus all fund-quarter observations where CDS usage is a possibility. To construct such a sampe of “potential” users of CDS, we resort to the Lipper fund style categories and include all funds from 37 out of the 182 categories in which at least one fund was a CDS user in our sample. Such a procedure results in about 20,000 fund-quarter observations, out of which 994 fund-quarter pairs hold CDS sell positions, 935 hold CDS buy positions, and 673 hold both. [Insert Table 3 here.] The first proxy for funds’ liquidity needs, is Flow volatility, defined as the standard deviation of estimated monthly fund flows (which is the return adjusted change in fund asset value, as is commonly 13 used in the literature) during the 24-month window ending in the same month as the portfolio period end date corresponding to the filing. The flow measure reflects the basic open-ended nature of mutual funds. Most open-end mutual funds offer daily liquidity to investors who can buy new shares or redeem shares from the fund at the funds’ NAV until just before the market closes. Providing investors with such a service imposes on the funds’ liquidity management, requiring these funds to keep adequate cash reserves and invest some of the fund assets in a set of securities that boast adequate liquidity to trade in and out upon short notice. Edelen (1999) shows that providing this liquidity service is quite costly for even mutual funds that primarily invest in the U.S. public equity market, and Chen, Goldstein, and Jiang (2010) showed how the complementarities among investors due to the open-end structure can impose challenges on the funds’ liquidity management. The challenge of accommodating fund flows increases with their unpredictability, which the flow volatility measure captures. Given the general lack of liquidity among corporate bonds (Edwards, Harris, and Piwowar, 2007; Bao, Pan, and Wang, 2011) and the liquidity advantage of the CDS market (Oehmke and Zawadowski, 2013a), the CDS market should be a more desirable venue for credit risk exposure for mutual funds with higher flow-driven, or external, liquidity needs. The second proxy for funds’ liquidity needs is Portfolio turnover, the annualized fund portfolio turnover rate, calculated as the lesser of the total amount of new securities purchased or the amount of securities sold, divided by the total net asset value (NAV) of the fund. The variable was reported in the CRSP Mutual Fund database. Turnover could be forced by fund flows, or by internal motives due to discretionary trading. The portfolio turnover rate is commonly considered a proxy for the active management of mutual funds (in either stock selection or market timing), and more active portfolio management gives rise higher portfolio-driven, or internal, liquidity needs. Finally, Flow volatility and Portfolio turnover are modestly correlated (with a correlation coefficient of 9.6%), suggesting that they capture quite distinctive aspect of funds’ liquidity needs. 14 The first two columns in Panel A of Table 3 adopts the standard logistic regression. Apart from the key variables proxying for funds’ liquidity needs, we include common control variables such as fund size (logarithm of total net assets), fund age (logarithm of years since inception), and funds’ performance rank (from 0, or worst, to 100, or best) within their respective Lipper fund style categories during the previous year. The regressions further include dummy variables for the four periods, as well as for the 37 Lipper fund style categories. The logit coefficients are “log ratio of odds ratios” (henceforth, simply “log odds ratios” as commonly used). In our context, the exponentiated coefficients indicate the multiple of the ratio Prob(CDS Usage)/[Prob(No CDS Usage)] relative to the base level due to a one-unit change in the regressors. Consistent with theoretical predictions, coefficients on both Flow volatility and Portfolio turnover turn out to be significantly (at the 1% level) positive. The economic magnitude is sizable, too. A onestandard deviation increase in Flow volatility (3.8 percentage points) is associated with an odds ratio for any CDS selling of 1.60. Relative to the unconditional probability of CDS selling (4.6%), this implies an incremental probability of 2.5 percentage points.7 Due to the small unconditional probability of CDS usage among all mutual funds, the odds ratios are roughly the same as the multiples of probability. Similarly, a one-standard deviation increase in Portfolio turnover is associated with roughly a 2.3 percentage point increase in the probability of CDS selling. The first two columns of Panel A show very similar coefficients for CDS selling and buying, which could be due to the large overlap of the two outcome variables. That is, the majority (54%) of fund-quarter observations where it holds CDS positions the fund also engages in both buying and selling. We use a multinomial logit model as an attempt to differential the determinants for CDS selling from those for buying. The last three columns in Panel A report results from one multinomial logit regression 7 The detailed procedure of calculation, using column (1) of Table 2 Panel A (CDS selling) as an example, is as follows: The base line odds ratio is Prob(CDS Selling)/[Prob(No CDS Selling)]= 994/(21840-994) = 0.048. A one standard deviation increase in Flow volatility increases the odds ratio to 0.076 (=0.048* exp(12.39*0.038)), which implies that Prob(CDS Selling)=0.071 (=0.076/(1+0.076), or an incremental probability of 0.025 (=0.071-0.046). The same calculation applies to other discussions of odds ratios. 15 where the baseline outcome is no-buy-and-no-sell. The coefficients in the “buy-no-sell” column (column (4)) are the log odds ratio for a fund to hold some buy position but no sell position during a period, relative to the baseline outcome, for a one unit change in a regressor. The coefficients for “both-buy-andsell” and “sell-no-buy” follow analogously. Overall, the coefficients suggest that the relation between CDS usage and fund liquidity needs is more driven by selling rather than buying. For example, the key coefficient of Flow volatility, is not statistically different between the “both-buy-and-sell” and “sell-no-buy” outcomes, but is significantly different (at the 10% level) between the “both-buy-and-sell”/ “sell-no-buy” and “buy-no-sell” outcomes. Not surprisingly, larger and older funds are more likely to engage in CDS investments. Prior performance does not predict CDS usage.8 Panel B of Table 3 proceeds with analyzing the determinants for the intensity of, rather than propensity to, CDS usage by mutual funds. Here the dependent variable is gross selling (or buying) intensity, defined as the total notional amount of CDS selling (or buying), scaled by the fund TNA during the period. The estimation method is the tobit regression to accommodate the large number of fundperiod observations with zero CDS holdings. Results again support the liquidity management hypotheses as both proxies for fund liquidity needs are significantly (at the 1% level) positive. A one-standard deviation increase in Flow volatility is associated with a 1.61 (1.12) percentage points increase in CDS selling (buying) intensity, both sizable relative to the unconditional average intensity of 0.13 percentage points for selling and 0.14 for buying. To summarize, results in Table 3 are highly consistent with predictions from Oehmke and Zawadowski (2014a) that CDS serves as an effective tool for institutional investors to management liquidity needs because the CDS market tends to be more liquid than the underlying reference bonds. B. Reference Entity Level Analysis 8 This no-result does not contract Adam and Guettler (2012) because Adam and Guettler (2012) analyze the motives for mutual funds to resort to CDS investment based on interim relative performance within a year. 16 To sharpen the test, we relate CDS usage directly to security-level liquidity needs by analyzing the determinants of CDS usage by funds at the issuer (reference entity) level, incorporating the effects of bond as well as CDS contracts characteristics, mostly importantly, their respective liquidity. Results are reported in Table 4. As in Table 3, Panel A performs logit regressions to analyze the propensity to holding CDS positions (buy or sell), and Panel B adopts tobit regressions to further assess the intensity of CDS holdings. Moreover, the regressions in Table 4 also incorporate dummy variables for the four periods, as well as for the 37 Lipper fund style categories. [Insert Table 4 here.] Panel A of Table 4 affirm the relation between CDS holdings and fund liquidity needs at the issuer level. More important from this analysis is the relation between CDS holdings and the liquidity measure for CDS securities and that for the underlying bonds. For the former, we follow the literature to settle on #CDS Contracts, defined as the number of quoted CDS contracts by the issuer as covered by Markit during the period. There are usually multiple CDS contracts traded on the reference entity, varying in both term structure (from six months to ten years) and contractual terms related to the definition of trigger events and deliverable obligations.9 #CDS Contracts proxies for the liquidity of an issuer’s CDS securities by capturing the density of contingency coverage for the credit risk of a reference entity.10 For bond liquidity, we follow the literature to adopt the Bond turnover variable, defined as the ratio of monthly dollar trading volume of bonds of the issuers over the issuance amount, using data from TRACE. When there are multiple bond issues for a given issuer-period, we pick the one with the largest issuance dollar amount. Somewhat surprisingly, the correlation between #CDS contract and Bond turnover is close to zero (−0.006), suggesting that the CDS and the underlying bond market could have 9 The common categories include “full restructuring” (FR), “modified restructuring” (MR), “modified-modified restructuring” (MM), and “no restructuring” (NR). The five-year MR contracts are usually the most liquid. 10 Another commonly used CDS liquidity measure is the number of dealers providing quotes on a reference entity, as covered by Markit. In our sample we find that #CDS Contracts entails more within sample variation and hence sharper results than # Dealers. 17 their relative advantages in trading liquidity. Finally, we add as control variable the asset size (in logarithm) of the issuer. As expected, #CDS contracts significantly (at the 1% level) predicts higher propensity of CDS usage by mutual funds. A one-standard deviation change in #CDS contracts (1.25 contracts) is associated with an odds ratio of 2.40 (1.07) for CDS selling (buying). Given the tiny unconditional probability of CDS selling/buying at a fund-issuer-period level (about 0.07% for selling and 0.13% for buying), the odds ratios imply that a one-standard deviation increase in the CDS liquidity proxy more than double the likelihood of a CDS selling while increases the likelihood of CDS buying only by 7%. Though both effects are statistically significant, one cannot fail to notice that CDS liquidity is far more effective predicting CDS selling than buying. The difference between the coefficients on #CDS contracts is significant (at the 1% level) between CDS buying and selling. Such evidence offers direct support to Oehmke and Zawadowski’s (2014a) predication that mutual funds may sell the liquid CDS contracts in lieu of bond long positions to achieve the same credit risk exposure. The same theoretical prediction for CDS buying is much weaker, and so is the empirical relation. Coefficients on Bond turnover are overall insignificant, expect in the specification explaining “Both buy and sell.” There, a one-standard deviation increase in Bond turnover (0.071) is associated with an odds ratio of 1.11 of (roughly an 11% increase in) the likelihood for a fund to be in both buying and selling of CDS contracts. It thus indicates that bond liquidity also make the derivative contract more desirable in general, but is not specifically related to the buying or selling of CDS. Results from tobit regressions for CDS selling/buying intensity reported in Panel B affirm the same economic relations suggested by those in Panel A. A one-standard deviation increase in #CDS contract is associated with a 3.51 (0.23) percentage points increase in gross selling (buying) intensity, both of which are statistically significant (at the 1% level) and economically substantial given the nearzero unconditional average CDS holdings at the fund-issuer-period level. The asymmetry between the effects of CDS liquidity on CDS selling and buying is also salient, and the difference between the two 18 coefficients is also statistically significant at the 1% level. The proxy for bond liquidity, on the other hand, does not significantly predict CDS selling or buying intensity. III. Mutual Fund CDS Holdings: Lead-follow and risk taking Given PIMCO’s absolute leader status among fixed income mutual funds and the relatively new CDS market, we hypothesize a lead-follow pattern between PIMCO and other funds in both CDS investments and in the directional betting. That is, a non-PIMCO fund is more likely to initiate a buying or selling CDS on a new reference entity if PIMCO disclosed a large position in the same reference entity and in the same direction during the previous period. We analyze the possibilities of both following PIMCO’s selling and following PIMCO’s buying. A. Following PIMCO’s CDS Selling We conduct both unconditional and conditional analyses to explore a lead-follow pattern in CDS selling. First, we ask whether a non-PIMCO mutual fund is more likely to initiate a new net selling position on a reference entity if PIMCO had disclosed a large net selling position in the same reference entity in the previous quarter. More specifically, the sample for the unconditional analysis is the “universe” of all reference entities, which consists of all reference entities that ever appear in our sample, excluding the net selling positions that a fund already held in the previous period. We run a probit regression at the fund-reference entity-period level where the dependent variable, New Sellingi,j,t is a dummy variable equal to one if fund i discloses CDS net selling in the reference entity j in period t and the fund did not disclose any net selling position j in period t-1. If a reference entity j is not among the disclosed net selling positions of fund i in period t, then New Sellingi,j,t is coded zero. The key independent variable, PIMCO Leadj,t-1, is a dummy variable equal to one if both of the following two conditions are met: (1) The reference entity is among the top 50 net selling positions during period t-1 by PIMCO funds, and (2) PIMCO’s selling position in the entity is larger than that by any other mutual fund 19 families in notional dollar amount in period t-1. Control variables include the following: The assets of the reference entity (in log); the five-year CDS spread on the reference entity using data from Markit (a leading data provider for the CDS market); and the total net assets of the fund (in log) as disclosed in the N-Q filing. All control variables are recorded at the end of the previous quarter. The results are presented in Table 5, where Panels A and B analyze the “Next 9” and “Rest 23” mutual fund families separately. [Insert Table 5 here.] Results in both panels of Table 5 show that the independent variable of key interest, PIMCO Leadj,t-1 is significant at less than the 5% level in all but one specification for the funds from the “Next 9” fund families, and is uniformly highly significant in all specifications for funds from the “Rest 23” fund families. When PIMCO displayed a large selling position in a previous period on a particular reference entity, the probability that a “Next 9” fund will initiate a selling position in the same entity, other things equal, increases by 0.8, 1.2, and 0.7 percentage points during 2007 (the latter three quarters), 2008, and 2009, respectively, if the fund did not already hold a position in the entity. Such incremental probabilities are sizable, given the unconditional probabilities for the funds to initiate a net selling position in a new reference entity in the given years (95, 82, and 16 basis points respectively). The corresponding incremental probabilities are lower for the “Rest 23” funds, at 0.5 – 1.0 percentage points. However, these numbers are even more significant relative to the much lower unconditional probability (about 1128 basis points) for this group of funds to initiate selling in a new reference entity. The evidence so far suggests that non-PIMCO mutual funds often follow the direction that their industry leader took in that they are more likely to take credit exposure on a new reference entity after observing a major bet made by PIMCO in a previous disclosure period. This unconditional relation is further affirmed by a conditional one: conditional on a new position being initiated by a fund in a period, it is far more likely to be a selling rather than buying position if PIMCO had a large selling position in the previous period. Such an analysis is more applicable if a portfolio manager conducts research into a 20 particular reference entity for possible CDS transactions, and takes into account PIMCO’s position in deciding her own trading directions. Table 6 reports the results from this analysis. [Insert Table 6 here.] The framework for the analysis reported in Table 6 is very similar to that in Table 5 except that the sample of fund-reference entity-periods is restricted to reference entities the fund hold in period t. The dependent variable is an indicator variable set to 1 if the position reflects a net selling of CDS, and is coded zero if the position is a net-buy or involves an equal notional amount in buying and selling. Though the majority of the new positions by non-PIMCO funds during the five semi-years starting in the second half of 2007 are buying (60 percent), the probability of the position being selling increases by 20 to 42 percentage points among the “Next 9” funds and 16 to 57 percentage points among the “Rest 23” if PIMCO had a large net selling position in the reference entity during the previous period. The coefficients are significant at 1% level across all specifications, supporting the hypothesis that PIMCO’s lead has a strong impact on other funds’ directional betting in CDS selling. The sign and magnitude of the coefficients on the control variables are of interest on their own. First, mutual funds are far more likely to initiate new CDS short positions on large entities on during the sample period. The top reference entities in our sample include Ford Motor, General Motors, Procter & Gamble, General Electric Capital Corp, American International Group, Morgan Stanley, Lehman Brothers, Goldman Sachs, and Citigroup Inc. That is, the largest players in the mutual fund sector were making concentrated bets on the good financial health of business behemoths in other sectors. A perception (which was ex pose justified) that certain large firms are “too big to fail” or “too systemic to fail” might have influenced these players. The second notable relation is that mutual funds initiated and sold new CDS contracts on reference entities that already appeared to be risky, as measured by the spreads on the most liquid five-year contracts with modified restructuring terms as reported in Markit. The conditional relation is much stronger than the unconditional one. 21 Interestingly, the tendency for mutual funds to initiate new bets on relatively risky firms (as captured by the lagged five-year CDS spreads) peaked in 2007 before the financial crisis unfolded. The “Next 9” funds continued to favor selling new reference entities with relatively high CDS spreads in 2008, but the preference was weaker than in 2007 and more or less disappeared by 2009. The “Rest 23” funds, on the other hand, started with a much lower sensitivity to risk in their new directional betting in 2007 and 2008, and ended in about the same level as the “Next 9” by 2009. B. Following PIMCO’s CDS Buying We conduct analogous analyses on the lead-follow pattern in CDS buying. We follow our earlier classification scheme (see Table 1 Panel D) to analyze speculative buys and hedging buys separately. The small “offsetting” category (4.0 percent of the sample total notional amount) is ignored because such positions do not represent the type of risk taking we are analyzing. We start with speculative buys where the long positions in CDS are not offset by any long positions in the bonds of the same issuer. Following the same set up, the dependent variable now becomes New Speculative Buyingi,j,t ,a dummy variable equal to one if fund i discloses CDS buying in the reference entity j in period t and the fund did not disclose any buying position j in period t-1. The key independent variable of interest, PIMCO Leadj,t-1, is revised to be a dummy variable equal to one if both of the following two conditions are met: (1) The reference entity is among the top 50 speculative buying positions during period t-1 by PIMCO funds, and (2) PIMCO’s speculative buying position in the entity is larger than that by any other mutual fund family in notional dollar amount in period t-1. The same set of control variables is included as in Table 5. Results are reported in Table 7. [Insert Table 7 here.] It turns out that a large speculative buy by PIMCO in a previous period increases the probability that other funds initiate new speculative buy positions in the same reference entities by 0.4 – 0.9 percentage points for the “Next 9” funds and 0.3 – 0.5 percentage points for the “Rest 23” funds. The 22 coefficients on PIMCO Leadj,t-1 are significant at the 1% level across all specifications. The economic magnitude is also sizable given that the unconditional probability that the other funds will initiate a new speculative CDS buy position in a given period ranges from 0.2-0.3 percent. Both the sell and buy sides convey the same herding pattern for risk taking: First, prior to the crisis, funds were more likely to exhibit herding behavior in which they took the same risk rather than to exhibit herding in which they took the same protection. Additionally, large mutual funds were more aggressive in following PIMCO’s lead than small ones, presumably because the former were more closely benchmarked, explicitly or implicitly, to the industry leader. Second, the financial crisis (and the looming losses) reduced mutual funds’ tendency to follow their leader’s risk behavior, but increased their tendency to take similar precautionary positions. Such pro-cyclical herding in risk taking constitutes an amplifying force in the systemic risk of all financial institutions. IV. Conclusion Exploring a comprehensive dataset of mutual funds’ quarterly holdings of credit default swap (CDS) contracts in 2007-2009, we test the motivations for institutional investors to hold CDS positions as modeled by Oehmke and Zawadowski (2013) and document evidence supporting predictions of Rajan (2006) that asset managers facing relative performance evaluations are incentivized to take “hidden tail risk” by selling CDS contracts. Additionally, we find that asset managers are prone to herding behavior which concentrates risk in the financial system. Further, the CDS market increases the correlation of financial health between the largest mutual funds and the largest and systemically important firms (especially financial institutions). 23 References: Adam, Tim and Andre Guettler, 2012. The use of Credit Default Swaps in Fund Tournaments, working paper, Humboldt University and EBS Business School. Atkeson, Andrew G., Andrea L. Eisfeldt, and Pierre-Olivier Weill, 2014. Entry and exit in OTC derivatives markets, working paper, University of California Los Angeles. Augustin, Patrick, Marti Subrahmanyam, and Sarah Qian Wang, 2014. Credit default swap (CDS): A survey. Foundations and Trends in Finance, forthcoming. Bao, Jack, Jun Pan, and Jiang Wang, 2011. The illiquidity of corporate bonds, Journal of Finance 65, 911-946. Chen, Qi, Itay Goldstein, and Wei Jiang, 2010. Payoff complementarities and financial fragility: Evidence from mutual fund outflows, Journal of Financial Economics 97, 239-262. Chevalier, Judith A., and Glenn Ellison, 1997. Risk taking by mutual funds as a response to incentives, Journal of Political Economy 105, 1167-1200. Edelen, Roger, 1999. Investor flows and the assessed performance of open-end fund managers, Journal of Financial Economics 53, 439-466. Edwards, Amy K., Lawrence E. Harris, and Michael S. Piwowar, 2007. Corporate bond market transaction costs and transparency, Journal of Finance 62, 1421-1451. Feroli, Michael, Anil K Kashyap, Kermit Schoenholtz, and Hyun Song Shin, 2014. Market tantrums and monetary policy, working paper, University of Chicago, New York University, and Princeton University. Jarrow, Robert A., 2011. The economics of credit default swaps, Annual Review of Financial Economics 3, 235-257. Koski, Jennifer Lynch and Jeffrey Pontiff, 1999. How are derivatives used? Evidence from the mutual fund industry, Journal of Finance 54, 791-816. Oehmke, Martin and Adam Zawadowski, 2014a. The anatomy of the CDS market, working paper, Columbia University and Boston University. Oehmke, Martin and Adam Zawadowski, 2014b. Synthetic or real? The equilibrium effects of credit default swaps on bond markets, working paper, Columbia University and Boston University. Rajan, Raghuram G., 2006. Has finance made the world riskier? European Financial Management 12, 499-533. Stultz, René M., 2010. Credit Default Swaps and the credit crisis. Journal of Economic Perspectives 24, 74-92. 24 Figure 1: Structure of Mutual Funds’ CDS Holdings Data This chart illustrates the structure of mutual funds’ CDS holdings data disclosed in N-Q, N-CSR and N-CSRS forms. A form is filed at the CIK (Series Trust) level where multiple filers could be affiliated with the same mutual fund family (or complex). The disclosure reveals holdings at the fund (portfolio) level where each series trust often encompasses several funds with similar or related investment strategies. Fund Family PIMCO PIMCO Funds CIK (Series trust) Fund (Portfolio) CDS holding Total Return Fund ... Low Duration Fund ... 25 PIMCO Variable Insurance Trust ... All Asset Portfolio ... ... ... Display: Details of CDS holdings (Filer: PIMCO Funds, CIK = 810893) Counterparty Bear Stearns Cos I Merrill Lynch & Co Bk of America Corp Deutsche Bk AG JPMorgan Chase Bk Citigroup Inc Deutsche Bk AG Reference Entity General Motors Corp. 7.125% due 07/15/2013 General Motors Corp. 7.125% due 07/15/2013 General Motors Corp. 7.125% due 07/15/2013 General Motors Corp. 7.125% due 07/15/2013 General Motors Corp. 7.125% due 07/15/2013 General Motors Corp. 7.125% due 07/15/2013 General Motors Corp. 7.125% due 07/15/2013 Buy/Sell Notional Amount ($1,000) Mark-to-market ($1,000) Period Ending Fund name Sell 11,700 -249 6/30/2008 TOTAL RETURN FUND Sell 25,000 -249 6/30/2008 TOTAL RETURN FUND Sell 30,300 -9,841 6/30/2008 TOTAL RETURN FUND Sell 6,400 -2,079 6/30/2008 TOTAL RETURN FUND Sell 7,000 -2,265 6/30/2008 TOTAL RETURN FUND Sell 6,400 -2,063 6/30/2008 TOTAL RETURN FUND Sell 14,600 -4,706 6/30/2008 TOTAL RETURN FUND 26 Figure 2. Distribution of Mutual Funds’ CDS Net Selling The three charts plot the distributions of CDS net selling, gross selling, and gross buying intensity at the fund (portfolio) level. For each fund in each period, we aggregate the notional amount of each CDS contract (buy or sell) for each reference entity. There are a total of 284 funds represented in the charts Panel A plots the distribution of net selling (where a buy is considered a negative sell), scaled by the funds’ total net assets (TNA), and expressed in percentage points: 𝑁𝑒𝑡 𝑠𝑒𝑙𝑙𝑖𝑛𝑔 = 𝑡𝑜𝑡𝑎𝑙 𝑠𝑒𝑙𝑙𝑖𝑛𝑔 𝑛𝑜𝑡𝑖𝑜𝑛𝑎𝑙 𝑎𝑚𝑜𝑢𝑛𝑡 − 𝑡𝑜𝑡𝑎𝑙 𝑏𝑢𝑦𝑖𝑛𝑔 𝑛𝑜𝑡𝑖𝑜𝑛𝑎𝑙 𝑎𝑚𝑜𝑢𝑛𝑡 × 100 𝑇𝑁𝐴 Similarly, Panels B and C plot gross selling and gross buying, respectively, in percentage points of the funds’ TNA. Panel A: Net selling 20.0% 18.0% 16.0% 12.0% 10.0% 8.0% 6.0% 4.0% 2.0% Sell Intensity 27 11, 12 > 12 9, 10 7, 8 5, 6 3, 4 1, 2 -1, 0 -3, -2 -5, -4 -7, -6 -9, -8 -11, -10 0.0% < -12 Frequency 14.0% Panel B: Gross selling 35.0% 30.0% Frequency 25.0% 20.0% 15.0% 10.0% 5.0% 8, 9 9, 10 10, 11 11, 12 > 12 8, 9 9, 10 10, 11 11, 12 > 12 7, 8 6, 7 5, 6 4, 5 3, 4 2, 3 1, 2 0, 1 Buy only 0.0% Gross Sell Intensity Panel C: Gross buying 35.0% 30.0% 20.0% 15.0% 10.0% 5.0% 7, 8 6, 7 5, 6 4, 5 3, 4 2, 3 1, 2 0, 1 0.0% Sell only Frequency 25.0% Gross Buy Intensity 28 Table 1. Table 1 provides an overview of CDS positions disclosed by mutual funds at the quarterly frequency for 2007-2009. Panel A shows CDS trading amounts in 2007, 2008 and 2009. #Positions is the number of CDS holdings on record. Amount is the notional amount of CDS contracts, and Sell Amount is the notional amount of contracts for which the mutual funds took the short side. #Families is number of mutual fund families. #Trusts is the number of series trusts (filers with unique CIK). #Funds is the number of mutual funds (portfolios). #Entities is the number of different CDS reference entity names. Each year is divided into the first and second half year. Panel B reports CDS notional amounts from 3 groups of mutual fund families: PIMCO (the top fund family by total CDS notional amount), the next 9, and the remaining fund families by CDS holdings. Panel C reports the CDS positions of the three groups of reference entities, sorted by their total CDS notional amount in the sample. Panel D further classifies CDS buying into three categories: An “offsetting” buy is a CDS long position that can be matched to a sell position on the same reference entity by the same fund during the same period. A “hedging” buy represents CDS long positions where the same fund has holdings in the same underlying bonds during the same period. We only match positions rather than the exact amount in classifying offsetting and hedging buys. Finally, a “speculative” buy represents the remainder of the long positions. Panel A. Mutual fund CDS holding by period 3098 4468 5886 6509 7281 9329 7667 6305 4922 3732 3868 3430 Amount ($1,000s) 13,154,975 15,898,276 20,034,953 22,064,073 24,317,488 29,161,554 25,989,193 22,565,593 21,158,307 15,392,066 15,694,105 15,976,014 Sell Amount ($1,000s) 8,535,859 9,315,313 13,654,085 14,306,067 15,024,279 17,443,943 15,986,209 11,880,053 9,560,698 8,080,759 8,446,495 9,145,908 66495 241,406,597 141,379,668 Period #Positions 2007Q1 2007Q2 2007Q3 2007Q4 2008Q1 2008Q2 2008Q3 2008Q4 2009Q1 2009Q2 2009Q3 2009Q4 2007-2009 29 #Families #Trusts #Funds #Entities 28 30 32 33 33 33 33 33 31 32 29 31 51 56 57 58 58 58 55 58 54 48 49 49 152 203 208 214 219 228 210 197 189 170 173 168 290 335 345 350 377 396 384 354 327 291 272 257 33 60 284 445 Panel B. Mutual Fund CDS holding by mutual fund family #Positions Amount ($1,000s) % of total Sell Amount ($1,000s) Buy Amount ($1,000s) All 66495 241,406,597 100% 141,379,668 100,026,929 PIMCO Next 9 Rest 26943 23312 16240 149,074,496 57,445,008 34,887,093 62% 24% 14% 95,106,951 33,149,319 13,123,398 53,967,545 24,295,689 21,763,695 Panel C. Mutual Fund CDS holding by CDS reference entities #Positions Amount ($1,000s) % of total Sell Amount ($1,000s) Buy amount ($1,000s) Avg spread in 2007 Avg spread in 2008 Avg spread in 2009 All 66495 241,406,597 100% 141,379,668 100,026,929 0.0259 0.0847 0.0849 Top 50 Next 50 Rest 345 32326 10906 23263 154,945,903 32,928,095 53,532,599 64% 14% 22% 114,762,631 7,885,987 18,731,050 40,183,271 25,042,108 34,801,549 0.0294 0.0144 0.0268 0.1072 0.0339 0.0767 0.1199 0.0205 0.0522 Panel D. Mutual fund CDS holdings by classified purposes Sell Amount ($1,000s) Buy: Offsetting ($1,000s) Buy: Hedging ($1,000s) Buy: Speculation ($1,000s) PIMCO 95,106,951 3,524,628 36,914,185 13,528,732 Next 9 33,149,319 1,509,001 7,497,575 15,289,112 Rest 13,123,398 319,601 9,862,411 11,581,684 Total 141,379,668 5,353,230 54,274,171 40,399,528 30 Table 2. Summary Statistics of Main Variables The sample in this Table includes all funds that are actual or potential CDS investors. For each mutual fund category by the Lipper class, if there is one fund holds any CDS position during any quarter, all fund-quarter observations from the Lipper class are included in our sample. This table reports the number of observations, mean, standard deviation, and values at the 25th, 50th (median), and 75th percentiles, of the main variables both at the fund-period and fund-issuer-period level. Flow volatility is the monthly standard deviation of estimated fund flows (measured as the change in return-adjusted fund asset value, as commonly used in the literature) during the 24-month window ending in same month as the portfolio period end date corresponding to the N-Q, N-CSR and N-CSRS filing. Portfolio turnover is a fund’s annualized portfolio turnover rate during the last year. Fund size is the logarithm of the total net assets of the fund that holds the position. Fund age is the logarithm of number of years since the fund first offered. Performance rank is the past 12-month performance ranking for mutual fund, within its Lipper classification. Any CDS sell (buy) is a dummy variable if a fund holds any CDS selling (buying) position in a given reference entity during a period. Gross sell (buy) intensity is the ratio of total notional amount of CDS selling (buying) by a fund in a given reference entity during a period, scaled by a fund’s total net assets during the same period. Variable name # Obs Mean Std Dev 25th Median 75th 19,454 19,454 19,454 19,454 19,454 19,454 19,454 19,454 19,454 19,454 19,454 0.052 1.09 1438.9 5.563 14.518 2.416 0.491 0.051 0.049 0.144 0.159 0.037 1.572 5990.3 1.838 12.212 0.722 0.278 0.221 0.215 1.098 1.149 0.025 0.31 77.9 4.355 7.145 1.966 0.254 0 0 0 0 0.043 0.62 251.4 5.527 12.089 2.492 0.488 0 0 0 0 0.068 1.21 885.6 6.786 17.022 2.835 0.727 0 0 0 0 Fund-period level Flow volatility Portfolio turnover Fund size ($ million) Log Fund size Fund age (years) Log Fund age Performance rank Any CDS sell Any CDS buy Gross sell intensity Gross buy intensity 31 Variable name # Obs Mean Std Dev 25th 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 4,147,385 0.05 1.075 1464.1 5.571 14.624 2.422 0.49 0.058 10.178 9.87 0.001 0.001 0.003 0.004 0.035 1.446 6060.5 1.83 12.426 0.72 0.278 0.071 1.296 1.472 0.027 0.037 0.166 0.166 0.023 0.32 78.3 4.361 7.252 1.981 0.253 0.013 10 8.858 0 0 0 0 Median 75th Fund-issuer-period level Flow volatility Portfolio turnover Fund size ($ million) Log Fund size Fund age (years) Log Fund age Performance rank Bond turnover # CDS contracts Assets reference entity Any CDS sell Any CDS buy Gross sell intensity Gross buy intensity 32 0.04 0.62 248.2 5.514 12.071 2.491 0.487 0.036 11 9.666 0 0 0 0 0.064 1.21 889.5 6.791 17.011 2.834 0.725 0.072 11 10.509 0 0 0 0 Table 3. CDS Selling and Funding Liquidity: Fund Level Analysis This table explores the fund-level relation between CDS usage and mutual funds’ funding liquidity needs and other fund-level characteristics using logit regressions. The sample includes all mutual funds (whether they hold CDS positions or not) that could potentially invest in CDS. Operationally we include all fund-quarter observations from 37 out of the Lipper fund style categories in which at least one fund holds any CDS position in any quarter. Panel A relates the propensity to hold CDS buy and/or sell positions to fund characteristics. The first two columns adopt logit regressions where the dependent variable is a dummy variable equal to one if a fund holds any CDS sell (column (1)) or CDS buy (column (2)) position during a period. The last three columns report results from a multinomial logit regression which estimates the risk ratio of each of the three parallel outcomes at the fund-period level: Buy no sell, both buy and sell, and sell no buy, relative to the base outcome of no buy or sell. In all columns, the reported coefficients represent the logarithm of the odds ratio of a particular outcome (relative to the base outcome) for a one-unit change in the corresponding covariate. Panel B relates the intensity of CDS usage to the same set of independent variable using tobit regressions. The dependent variables are Gross sell intensity and Gross buy intensity, defined at the total notional amount of all CDS sell (or buy) positions during a fund-period, scaled by a fund’s total net assets. The two key independent variables are proxies for mutual fund liquidity needs. The first is Flow volatility, the monthly standard deviation of estimated fund flows (measured as the change in return-adjusted fund asset value, as commonly used in the literature) during the 24-month window ending in same month as the portfolio period end date corresponding to the N-Q, N-CSR and N-CSRS filing. The second is Portfolio turnover, a fund’s annualized portfolio turnover rate during the last year. Control variables include Fund size, the logarithm of the total net assets of the fund that holds the position; Fund age, the logarithm of number of years since the fund first offered; Performance rank, the past 12-month performance ranking for mutual fund, within its Lipper classification. All regressions include dummy variables for time periods and Lipper fund categories for which the coefficients are suppressed. Standard errors adjust for heteroskedasticity and clustering at the fund level. *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level respectively. 33 Panel A. Propensity to CDS usage by mutual funds Dependent variable: Flow volatility Portfolio Turnover Log Fund size Log Fund age Performance rank # observations % (Dep var =1) Pseudo R squared Logit Any sell Any buy (1) (2) 12.51*** (2.442) 0.294*** (0.0552) 0.555*** (0.0595) 0.192 (0.151) -0.206 (0.232) 8.919*** (2.494) 0.279*** (0.0538) 0.430*** (0.0553) 0.103 (0.155) -0.403* (0.214) 19,137 5.22% 0.2390 19,454 4.87% 0.1968 Logit, vs. no buy no sell Buy no sell Both buy and sell Sell no buy (4) (5) (6) 3.055 (3.757) 0.123*** (0.0456) 0.261*** (0.0628) 0.0492 (0.200) 0.106 (0.295) 16,000 1.64% 0.1133 12.44*** (2.786) 0.316*** (0.0615) 0.535*** (0.0665) 0.151 (0.173) -0.522** (0.259) 13.57*** (3.984) 0.168*** (0.0611) 0.623*** (0.0875) 0.340 (0.239) 0.554 (0.394) 18,565 3.68% 0.2418 15,291 2.06% 0.2380 Panel B: Intensity of CDS usage by mutual funds Dependent variable: Gross sell intensity (1) Gross buy intensity (2) 38.65*** (9.285) 0.913*** (0.171) 1.606*** (0.219) 0.418 (0.435) -1.121 (0.701) 26.83*** (8.831) 1.093*** (0.207) 1.369*** (0.188) 0.0386 (0.517) -1.471** (0.724) 19,454 0.1427 19,454 0.1241 Flow volatility Portfolio Turnover Log Fund size Log Fund age Performance rank # observations Pseudo R squared 34 Table 4. CDS Selling and Funding Liquidity: Reference Entity Level Analysis This table explores the reference entity-level relation between CDS usage and both fund-level and reference entity level characteristics using logit regressions. The sample includes all potential CDS positions by all mutual funds which could potentially invest in CDS positions. The “potential” CDS positions include all reference entities that appear in our sample. The set of mutual funds are the same as in Table 2. Panel A relates the propensity to hold CDS buy and/or sell positions to fund and reference entity characteristics using logit regressions. Panel B relates the intensity of CDS usage to the same set of independent variable using tobit regressions. In both panels the specification are the same as in the corresponding panels of Table 2, except that all variables are recorded at the fund-reference entity-period level. The fund-level variables are defined the same way as in Table 2. The key reference entity-level independent variables are the proxies for bond and CDS liquidity. The first is Bond turnover defined as the ratio of monthly dollar trading volume of bonds of the issuers over the issuance amount. When there are multiple bond issues for a given issuer-period, we pick the one with the largest issuance dollar amount. The second is #CDS contracts defined as the number of quoted CDS by the issuer as covered by Markit during the period. The additional reference-entity level control variables is Assets Reference Entity, the logarithm of the reference entity’s assets. All regressions include dummy variables for time periods and Lipper fund categories for which the coefficients are suppressed. Standard errors adjust for heteroskedasticity and clustering at the fund level. *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level respectively. 35 Panel A: Propensity to CDS usage by mutual funds Dependent variable: Flow volatility Portfolio turnover Log Fund size Log Fund age Performance rank Bond turnover # CDS contracts Assets reference entity # observations % (Dep var =1) Pseudo R squared Any sell Any buy Any sell Any buy (1) (2) (3) (4) Buy no sell (4) 14.00*** 7.520** 7.645** (3.389) (3.091) (3.156) 0.233*** 0.328*** 0.329*** (0.0559) (0.0530) (0.0541) 0.653*** 0.751*** 0.750*** (0.0868) (0.142) (0.145) -0.00736 -0.169 -0.166 (0.181) (0.165) (0.168) 0.220 -0.218 -0.219 (0.332) (0.419) (0.428) 0.221 0.263 0.220 0.268 0.0174 (0.221) (0.252) (0.224) (0.256) (0.275) 0.724*** 0.0770*** 0.732*** 0.0788*** 0.0631*** (0.0556) (0.0178) (0.0555) (0.0182) (0.0163) 0.589*** 0.137*** 0.596*** 0.141*** 0.0943*** (0.0417) (0.0266) (0.0431) (0.0275) (0.0318) 4,094,042 4,117,080 4,094,042 4,117,080 0.075% 0.1455 0.140% 0.1085 0.075% 0.2285 36 0.140% 0.2455 Both buy and sell (5) Sell no buy (6) 6.852 (4.429) 0.328*** (0.0657) 0.857*** (0.186) -0.224 (0.236) -0.0572 (0.579) 2.005*** (0.510) 0.568*** (0.140) 0.706*** (0.0881) 15.01*** (3.455) 0.222*** (0.0594) 0.644*** (0.0867) 0.0173 (0.189) 0.295 (0.346) -0.132 (0.231) 0.770*** (0.0644) 0.581*** (0.0422) 4,073,070 2,756,756 4,047,326 0.131% 0.2425 0.015% 0.2596 0.065% 0.2176 Panel B: Intensity of CDS usage by mutual funds Dependent variable: Gross sell intensity Gross buy intensity Gross sell intensity Gross buy intensity (1) (2) (3) (4) Flow volatility 1.457 (0.962) 2.895*** (0.342) 2.596*** (0.281) 0.861 (0.729) 0.229*** (0.0668) 0.434*** (0.102) 57.98*** (16.00) 1.016*** (0.250) 2.593*** (0.340) 0.0797 (0.749) -0.00815 (1.369) 1.656* (0.957) 2.899*** (0.348) 2.603*** (0.284) 4,147,385 0.1143 4,147,385 0.0831 4,147,385 0.1714 Portfolio turnover Log Fund size Log Fund age Performance rank Bond turnover # CDS contracts Assets reference entity # observations Pseudo R squared 37 22.24** (9.404) 1.066*** (0.202) 1.988*** (0.257) -0.601 (0.476) -1.334 (1.330) 1.028 (0.732) 0.238*** (0.0735) 0.448*** (0.0996) 4,147,385 0.1735 Table 5. Lead-Follow in CDS New Net Selling: Unconditional Analysis This table reports results of logit regressions for the probability of non-PIMCO mutual funds’ net selling of CDS on new reference entities. The unit of observation is at the fund-reference entity-period (quarter) level, and the universe of reference entities in any period t consists of all potential entities (including any reference entity that ever appears in our sample) for which a fund did not already have a net short position on in period t-1. The dependent variable is a dummy variable equal to one if a fund takes a new net selling CDS position in a reference entity (the new position is defined as a CDS net selling position by a mutual fund in period t on a reference entity for which the mutual fund did not have a net selling position in period t-1). The key independent variable, PIMCO Lead is a dummy variable equal to one if the following two conditions are met: (1) The entity is among the top 50 net selling positions during period t1 by PIMCO funds, and (2) PIMCO’s position in the entity is larger than that by any other mutual fund family in notional dollar amount. Control variables include the following: The assets of the reference entity (in log) in 2006; the five-year CDS spread on the reference entity in period t-1 using data from Markit; and the total net assets of the fund (in log) as disclosed in the N-Q, N-CSR and N-CSRS filing. All regressions include quarterly dummies. Panels A and B analyze the “Next 9” and “Rest 23” mutual fund families, respectively. Standard errors clustered at the fund level are reported in parentheses, and *, ** , and *** indicate statistical significance at the 10%, 5%, and 1% level respectively. Panel A: The next 9 mutual fund families 2007 (1) PIMCO Lead (t-1) 2009 (3) (4) (5) (6) 1.512*** (0.0760) 104,587 0.934*** (0.0951) 0.272*** (0.0221) 6.415*** (0.333) -0.0886*** (0.0281) 87,999 2.137*** (0.187) 77,546 0.619*** (0.0996) 0.352*** (0.0261) 38.14*** (1.471) -0.00151 (0.0321) 65,150 79,172 1.698*** (0.216) 0.445*** (0.0523) 0.628* (0.339) 0.425*** (0.0830) 64,451 0.950% 0.953% 0.823% 0.819% 0.148% 0.163% 0.0428 0.1228 0.0502 0.0864 0.0636 0.1403 1.405*** (0.0840) Assets Reference Entity (log) Spread-5year (t-1) Fund TNA (log) # observations % (Dep var =1) Pseudo R squared 2008 (2) 38 Panel B: The remaining 23 mutual fund families 2007 (1) (2) PIMCO Lead (t-1) 2009 (4) (5) (6) 2.003*** (0.0990) 165,411 1.109*** (0.123) 0.433*** (0.0293) 2.979*** (0.605) 0.133*** (0.0390) 139,143 2.852*** (0.188) 117,795 1.356*** (0.136) 0.410*** (0.0350) 24.67*** (2.468) 0.168*** (0.0478) 98,854 114,831 2.273*** (0.217) 0.441*** (0.0520) 0.459 (0.305) 0.338*** (0.0688) 93,472 0.284% 0.281% 0.258% 0.268% 0.106% 0.119% 0.0510 0.1063 0.0628 0.1031 0.1366 0.2017 1.932*** (0.112) Assets Reference Entity (log) Spread-5year (t-1) Fund TNA (log) # observations % (Dep var =1) Pseudo R squared 2008 (3) 39 Table 6. Lead-Follow in CDS on New Net Selling: Conditional Analysis This table reports the results of logit regressions for the probability of non-PIMCO mutual funds’ net selling of CDS on new reference entities. The unit of observation is at the fund-reference entity-period (quarter) level. The universe of reference entities consists of all reference entities held by a fund in any period t for which the fund did not already have a net short position in period t-1. The dependent variable is a dummy variable equal to one if a fund takes a new net selling CDS position in a reference entity (the new position is defined as a CDS net selling position by a mutual fund in period t on a reference entity for which the mutual fund did not have a net selling position in period t-1). The key independent variable, PIMCO Lead is a dummy variable equal to one if the following two conditions are met: (1) the entity is among the top 50 net selling positions during period t-1 by PIMCO funds, and (2) PIMCO’s position in the entity is larger than that by any other mutual fund family in notional dollar amount. Control variables include the following: The assets of the reference entity (in log) in 2006; the five-year CDS spread on the reference entity in period t-1 using data from Markit; and the total net assets of the fund (in log) as disclosed in the N-Q, N-CSR and N-CSRS filing. Panels A and B analyze the “Next 9” and “Rest 23” mutual fund families respectively. Standard errors clustered at the fund level are reported in parentheses, and *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level respectively. Panel A: The next 9 mutual fund families 2007 (1) PIMCO Lead (t-1) 1.653*** (0.141) Assets Reference Entity (log) Spread-5year (t-1) Fund TNA (log) # observations % (Dep var =1) Pseudo R squared 2,216 2008 2009 (2) (3) (4) (5) (6) 0.855*** (0.199) 0.388*** (0.0453) 90.68*** (5.085) -0.135** (0.0581) 1,932 1.556*** (0.105) 0.703*** (0.150) 0.228*** (0.0322) 18.44*** (1.183) -0.353*** (0.0421) 3,075 3.432*** (0.288) 2.249*** (0.354) 0.618*** (0.0858) 3.761*** (1.168) 0.0245 (0.119) 981 3,521 1,112 33.3% 32.1% 24.5% 23.4% 10.5% 10.7% 0.1015 0.3614 0.0689 0.2004 0.2339 0.3417 40 Panel B: The remaining 23 mutual fund families 2007 (1) (2) PIMCO Lead (t-1) 2009 (4) (5) (6) 2.190*** (0.123) 3,557 1.071*** (0.181) 0.297*** (0.0407) 14.21*** (1.713) 0.134*** (0.0506) 3,229 3.510*** (0.222) 2,123 1.103*** (0.219) 0.147*** (0.0507) 39.77*** (4.709) 0.289*** (0.0648) 1,927 2,207 3.340*** (0.329) 0.0508 (0.0774) 1.605** (0.658) 0.235** (0.0934) 1,980 15.7% 14.4% 12.0% 11.6% 5.5% 5.6% 0.1012 0.1789 0.1357 0.1811 0.3115 0.3439 1.964*** (0.141) Assets Reference Entity (log) Spread-5year (t-1) Fund TNA (log) # observations % (Dep var =1) Pseudo R squared 2008 (3) 41 Table 7. Lead-Follow in CDS Speculative Buying on New Reference Entities: Unconditional Analysis This table reports the results from logit regressions for the probability of non-PIMCO mutual funds’ speculative buys of CDS on new reference entities. The unit of observation is at the fund-reference entityperiod (quarter) level, and the universe of reference entities in any period t consists of all potential entities (including any reference entity that ever appears in our sample) for which a fund did not already have a speculative buy position in period t-1. The dependent variable is a dummy variable equal to one if a fund takes a new speculative buying CDS position in a reference entity (a new position is defined as a CDS speculative buying position by a mutual fund in period t on a reference entity for which the mutual fund did not have a speculative buy position in period t-1). The key independent variable, PIMCO Lead is a dummy variable equal to one if the following two conditions are met: (1) The entity is among the top 50 speculative buying positions during period t-1 by PIMCO funds, and (2) PIMCO’s position in the entity is larger than that by any other mutual fund family in notional dollar amount. Control variables include the following: The assets of the reference entity (in log) in 2006; the five-year CDS spread on the reference entity in period t-1 using data from Markit; and the total net assets of the fund (in log) as disclosed in the N-Q, N-CSR and N-CSRS filing. Panels A and B analyze the “Next 9” and “Rest 23” mutual fund families respectively. Standard errors clustered at the fund level are reported in parentheses, and *, **, and *** indicate statistical significance at the 10%, 5%, and 1% level respectively. Panel A: The next 9 mutual fund families 2007 (1) PIMCO Lead (t-1) 0.629*** (0.143) Assets Reference Entity (log) Spread-5year (t-1) Fund TNA (log) # observations % (Dep var =1) Pseudo R squared 85,480 0.649% 0.0050 2008 2009 (2) (3) (4) (5) (6) 0.529*** (0.150) -0.195*** (0.0349) -40.16*** (5.462) 0.146*** (0.0369) 71,708 0.708% 0.0253 0.500*** (0.156) 0.558*** (0.162) 0.0771** (0.0326) -3.180** (1.449) 0.118*** (0.0397) 99,877 0.370% 0.0229 0.708** (0.283) 0.728** (0.289) 0.0779 (0.0630) -0.964 (1.139) 0.272*** (0.0829) 70,995 0.139% 0.0255 42 118,865 0.364% 0.0171 87,275 0.131% 0.0118 Panel B: The remaining 23 mutual fund families 2007 (1) (2) PIMCO Lead (t-1) 1.183*** (0.116) Assets Reference Entity (log) Spread-5year (t-1) Fund TNA (log) # observations % (Dep var =1) Pseudo R squared 117,330 0.480% 0.0158 2008 1.193*** (0.121) 0.0480* (0.0290) -30.85*** (4.998) -0.0477 (0.0336) 98,359 0.530% 0.0282 43 2009 (3) (4) (5) (6) 0.473*** (0.141) 0.434*** (0.142) -0.0402 (0.0311) -11.56*** (2.237) -0.172*** (0.0318) 145,288 0.339% 0.0410 1.001*** (0.167) 0.853*** (0.178) -0.0300 (0.0438) -3.206*** (1.142) -0.0474 (0.0481) 101,437 0.238% 0.0423 172,937 0.312% 0.0304 124,746 0.212% 0.0373