The impact of CDS Auctions on recovery rates.
Samir Bel Mouhand 10020993
University of Amsterdam
Faculty of Economics and Business Master Thesis
July 2014
Abstract
This paper examines whether the auction mechanism introduced in 2005 to settle CDS contracts impacted recovery rates of defaulted bonds. Using a sample containing US firms filing for Chapter 11 in the period 2003-2014, this paper provides evidence that CDS auctions resulted in defaulted bonds underpricing. This underpricing is proportional to the Net Open Interest (NOI) over the net notional amount of protection bought. Whenever the NOI is to sell (buy), securities are underpriced (overpriced). These results are robust to using different market recovery rates for bonds not auctioned. This paper also shows that institutional trading through the market on the auction day itself tends to be higher in a eight day window surrounding the auction. In addition, the results show that prepackaged Chapter filings yield higher recovery rates.
Keywords: Bankruptcy, CDS, Default, Auction, Creditfixings, Bonds, Prepackaged, Net
Open Interest, Protection buyers.
Data availability: The dataset is described in section IV.2 of this paper.
MSc Business Economics, Finance Thesis supervisor: Dr. Rafael Matta
2
Acknowledgements:
I would like to thank my family and friends for supporting me throughout my studies. Many thanks to the staff of the Finance Group, which is dedicated to excellent research and teaching, of which I had the privileged to learn from. And lastly, a special thanks to my supervisor Dr. Rafael Matta, an excellent researcher with whom I had the pleasure to work with.
3 Table of Contents
I. Introduction 4
II. The Auction Mechanism 6
II.1 First Stage 6
II.2 Second Stage 7
II.3 Auction Protocol 10
III. Literature review 13
III.1 CDS Auctions 14
III.2 Recovery Rates 16
IV. Research Method 18
IV.1 Hypotheses 18
IV.2 Dataset 20
IV.2.A Data 20
IV.2.B Composition of the sample 22
IV.3 Methodology 23
V. Results 27
V.Results Regression one 27
V.1.A Baseline regression (1) results 27
V.1.B Reduced form regressions 31
V.1.C Auction induced trading 33
V.2. Robustness 36
V.2.A Robustness check 1 and 2 36
V.2.B Robustness check 3 (liquidity) 43
VI. Conclusion and discussion 43
Bibliography 46
Appendix 1: Definitions of the variables. 48
I. Introduction
This paper examines the impact of the auction mechanism introduced in 2005 to settle credit default swaps (CDS) contracts. A CDS contract is used to insure against the default of a specific firm ( reference entity). A protection buyer can insure the face value of a bond or (loan) by buying a CDS (LCDS), likewise a protection seller can sell a CDS for a periodic fee, against the obligation that it needs to make the buyer whole in case of a trigger event1. Specifically, in case of a default the protection buyer is entitled to the face value in exchange for the defaulted bond. This required that CDS were settled physically. However this settlement mechanism became impractical, since in some cases the notional amount of CDS contracts outstanding exceeded the notional amount of deliverable bonds outstanding due to speculators2. This means that in case of a credit event, protection buyers not holding the bond would rush to the market to buy the bond in order to get the face value and thus yielding inflated recovery rates. This reduces the net payment received by the protection buyer. This was the main motive3 for the International Swaps and Derivatives Association ( ISDA) in cooperation with Creditex and Markit to devise an auction mechanism that allowed for cash settlement4.The auction devised consists of two stages. The first stage allows for a replication of physical settlement, meaning dealers submit requests for physical settlement on behave of clients and themselves. This is than aggregated into the Net Open Interest (NOI). Dealers are also required to submit bid and offer quotes for predetermined notional amounts5 with usually the spread between the prices being equal to 2% of the face value. This information is then used to calculate the Initial Market Midpoint (IMM). Which acts as cap (floor) whenever the
NOI is to sell (buy).The first stage thus produces two key numbers, the NOI and IMM, which
is made available to the public and market participants simultaneously through creditfixings.com. In the second stage one sided limit bids/offers are made, whenever the
NOI is to sell(buy) dealers sent limit bids (offers). These limit orders are than combined with
the orders made in the first stage and used to clear the NOI. The bid or offer that clears the
NOI is set equal to final auction price. The final auction price is then used to settle all CDS
contracts. In case a protection buyer opted for a physical settlement in the first stage, the gross payoff is equal to the face value. While protection buyers that opted for cash settlement
1
The most common trigger events according to the ISDA master agreement include the following ones: Bankruptcy, Failure to Pay, Restructuring, Repudiation, Moratorium, Obligation Acceleration and Obligation Default.
2 Holding a bond is not a requirement for entering in a CDS contract.
3 Another motive is the recovery basis risk of parties holding bonds and CDS would be eliminated, because the auction would allow for positions to be closed out simultaneously.
4
In case of cash settlement the protection seller pays the buyer (par-recovery value bond). And the protection buyer does not need to deliver the bond.
5
get the face value minus the final auction price. Early empirical research on the auction mechanism suggested that the auction fairly prices defaulted bonds. More recent theoretical research, suggest that the auction is biased in pricing and allows for underpricing whenever the NOI is to sell. Specifically, Chernov et al. (2013) extend the model of Wilson (1979) and Back and Zender (1993) and show that CDS auctions are biased, large investors can influence the NOI and final price. The auction final price can be influenced because of the one-sided design of the second stage. Du and Zhu (2013), devised a model that shows that allocative efficiency is low under the current auction mechanism and propose a double auction design in which limit orders in the second stage can be submitted in both directions. Empirically, previous research has focused on pre and post auction recovery rates of bonds auctioned, and not on the impact of CDS auctions market wide. One might argue that the market overprices instead of auctions underpricing, since post auction market recovery rates tend to be lower than pre auction market recovery rates of auctioned bonds.
That is why this paper uses a more general empirical framework to see whether the CDS auctions have impacted recovery rates. Henceforth the main research question is: Do CDS auctions impact recovery rates? This paper contributes to a small body of research on CDS auctions. Previous empirical studies on CDS auctions include Chernov et al. (2013), Gupta and Sundaram (2012), Helwege et al.(2009), and Coudert and Gex (2010). Theoretical treatments of CDS auctions include Chernov et al. (2013) and Du and Zhu (2013). This paper is the first that tries to disentangle the effect of CDS auctions on recovery rates market wide. By using a sample of US firms filing for bankruptcy in the period 2003-2014, of which a subsample had their bonds auctioned. The results show that auctions effected recovery rates, auctioned bonds are underpriced whenever the NOI is sufficiently high enough (to sell) relative to Net Notional Amount of CDS Outstanding(NND), the opposite holds if the NOI is negative (to buy). This underpricing is reflected in the market recovery rates observed post auction but seems to be mitigated through post auction induced trading. This results are robust to using different proxies for market recovery rates on non-auctioned bonds. Also when excluding low liquid bonds the results indicate roughly the same. In addition results show, that auctions induce increased institutional trading and that prepackaged Chapter filings yield higher recovery rates.
The remainder of this paper is organized in the following way. Section II gives a full description of the CDS auction. In section III, previous research on CDS auctions and recovery rates is discussed. Subsequently in section IV the research method is outlined. It
6
contains the main hypothesis, the data collection procedure, characteristics of the sample and the methodology. In section V, the results are presented for all regressions performed including robustness checks. Finally section VI provides a summary of the main results and offers explanations.
II. The Auction Mechanism
This section discusses the auction mechanism in detail6. The first part concerns the first stage of the auction. The second part describes the second stage of the auction and gives an illustration of the mechanism. While the last part describes some aspects and rules of the auction protocol, in addition to some changes in the protocol that took place in 2009. The discussions below are based on material made available by the ISDA, Creditex and Markit.
II.1 First stage
Before the first stage starts, usually the day before the auction takes place, market participants that are either a net protection buyer or seller, can submit a physical settlement request to buy or sell a defaulted bond. In order for this to be accepted at least two conditions have to be met. First the relevant bond should be listed on the list of deliverable obligations published by the ISDA and determined by the Determinations Committee (DC). And second the relevant party submitting the request should have at least an equal opposite market position in terms of CDS notional exposure. So for example a net protection buyer who has bought net protection on 100 notional value (face value), can sell bonds up to 100 of notional value (face value) through a physical settlement request. Likewise a net protection seller that sold net protection on 100 notional value, can buy at most for 100 notional value of bonds through a physical settlement request. This means that market participants with truly offsetting positions (zero net protection) cannot do a physical settlement request. There is no mechanism in place that makes sure this condition is met, it is rather left over to the best knowledge of the relevant party. Each request filled with a dealer is a buy or sell for a chosen notional amount, dealers themselves also reserve the right to do a physical settlement on behave of themselves.
After the first round started dealer submit the sum of their physical settlement request ( sells minus buys) and in addition dealers make a two-way market by submitting bids and
6
The first single-name auction that followed the current was conducted on November 28, 2006 for the Dura credit event, before this, the auction format consisted of one stage. The one-stage format is not discussed since only 7 single- name auctions were conducted under the old format.
7
offers with a predetermined bid-ask spread on a given quotation size7. Both stated in the reference-entity specific protocol published before the auction takes place. The first stage usually lasts 15 minutes and it allows dealers to send their aggregated physical settlement amount coupled with their bids and offers for the given quotation size. After the end of the first stage, the administrators Creditex and Markit calculate the NOI and IMM. The NOI is calculated by aggregating al physical settlement requests by summing net sells and net buys and taking the difference. If the NOI is positive it means the Net Open Interest is to sell, which is usually the case. If however the NOI is negative, the Net Open Interest is to buy. The IMM is then calculated and derived from the bids and offers submitted by the dealers for a given quotation size. First the crossing/touching bids and offers are discarded. The ‘best half’ are than taken of bids and offers submitted and used to calculated the average, the best halves are considered to be the highest bids and lowest offers.
After the IMM is calculated the administrators assess whether any adjustment amounts have to be paid by participating dealers. A dealer pays an adjustment amount if it made a bid or offer that crossed and if either the bid made is higher than the IMM and the
NOI is to sell, or when the NOI is to buy and the offer made is lower than the IMM. In the
first round the IMM is already utilized as a cap or floor on the price. And any deviation is considered as an off market bid or offer. That is why dealers are than required to pay an adjustment amount to the ISDA, which is calculated as the product between the given quotation size and the difference between bid or offer submitted and the corresponding
IMM. Following all calculations, Creditex one of the administrators publishes the NOI, IMM
and the bids and offers made by the dealers on their website creditfixings.com8.
II.2 Second stage
After the publication of the first stage results, dealers and market participants usually wait two hours before the second stage starts9. If the NOI is zero, the auction final price is set equal to the IMM. Otherwise dealers are required to submit limit orders, if the NOI is to sell, dealers submit limit bids, if the NOI is to buy they submit limit offers. The quotation sizes accompanying those orders are not predetermined in the second stage, but rather set by dealers themselves. Dealers are allowed to submit limit orders on behave of clients and
7
The bid-ask spread is usually 2% of the face value. Quotation sizes range from $2 to $10 million, $2 million is the most common amount.
8
This information is made available to participating dealers and the general public in the same way, this is confirmed by the Legal department of the ISDA. This limits potential arbitrage payoffs through intraday trading.
9
The most common schedule is the following one; Round 1 09:45 till 10:00 AM , publication 10:30 AM and the second round starts at 12:45PM ending at 13:00PM. The time between rounds allows auction participants to engage in arbitrage through the secondary market, since the price of the bond is capped by the IMM.
8
themselves, no restrictions are made regarding the market positon in CDS. This means that market participants with zero market position can buy or sell the bond at the auction final price. Dealers have in total fifteen minutes to submit their orders, after which the second stage ends and administrators calculate the auction final price. In case of a NOI to sell the open interest is matched with market bids of the first and second stage, until it is filled up completely. This process starts with matching with the highest bid order and continuing with the remaining highest bids until the open interest is filled completely or the bids are exhausted. The last match yields the auction final price and clears the open interest. However in case of a NOI to sell this auction final price is capped at the IMM plus half of the pre specified spread used in stage one. This means that if the auction final price is higher than this cap, it is set equal to the cap. In case of a NOI to buy, the price has a floor equal to the IMM minus half of the spread. In this case a violation of the floor yields the floor as the auction final price.
Although never occurred, the ISDA set rules regarding a NOI not completely filled, this happens in case the NOI is to sell, whenever the bids made are exhausted and the NOI has not been cleared, the auction final price is than set at zero and bids will be filled on a pro-rata basis. In case the NOI is to sell and not cleared, the auction final price is set equal to the face value of the bond (100). After the auction final price is determined given the restrictions outlined, all information is published on creditfixings.com, this happens usually one hour after the second stage ends10.
As an example, the auction regarding Bowater conducted on 12 May 2009 is used to illustrate the auction mechanism. Table 1A lists the initial market bids and offers submitted by participating dealers in the first round, the quotation size in this auction was two million notional value. The NOI is to sell and amounts to $117.583 Million11. First the bids are ranked in descending order and the offers in ascending order this is done in Table 1B.
10
The second round usually starts at 12:45 PM NY time and ends at 13:00 PM, the auction final price and whether bids or offers of specific dealer were completely or partially filled is published at 14:00 PM.
11
The sum of buy physical request was 1 million while the sum of sell physical request was 118.583m in this auction. Yielding a NOI of 117.583m.
9
Table 1A: Initial market submissions, NOI is 117.583 and quotation size 2m.
Table 1B: List the bids (offers) descending (ascending) order.
Dealer Bid Offer
Bank of America 12.5 14.5 Barclays 12.5 14.5 BNP Paribas 13 15 Citibank 15.5 17.5 Credit Suisse 12.5 14.5 Deutsche Bank AG 14.5 16.5 Goldman Sachs 13 15
JP Morgan Chase Bank 13 15
Morgan Stanley 12.5 14.5
The Royal Bank of Scotland 14 16
UBS 13 15
Dealer Bid Offer
Crossed (bid>offer) 15.5 14.5 Touched (bid=offer) 14.5 14.5 Not tradable 14 14.5 Not tradable 13 14.5 Not tradable 13 15 Not tradable 13 15 Not tradable 13 15 Not tradable 12.5 15 Not tradable 12.5 16 Not tradable 12.5 16.5 Not tradable 12.5 17.5
10
Table 1B shows that two trades are going to be discarded, and since the NOI is to sell, the bid side is going to be penalized. First the bid that crossed is made by Citibank who needs to pay an adjustment amount of (15.5-IMM) times 2 million. For the touched bid made by Deutsche Bank AG, the penalty is (14.5-IMM) times 2 million. The IMM is computed as an average of the best half of the non-tradable pairs, in this example the average of the sum of two times 14.5, 14 and three times 13 and two times 15, resulting in an IMM of 14. So that the adjustment amounts that need to be paid by Citibank and Deutsche Bank AG are $30,000 and $10,000, respectively.
The IMM act as an cap for final auction price, since the spread is 2% of the face value, the cap is set at 15. In the second stage dealers submit their limit bid orders for a quotation size of their choosing, these bids are combined with bids made in the first stage and ranked descending order. The bids made in the first stage by Citibank and Deutsche Bank AG are set equal to the IMM and included in the list. Table 2 shows the list in which all bids ( stage one & two) are aggregated and ranked in descending order.
11
Table 2: List of Bids, * indicate bids of first stage and ^ indicate orders partially filled.
Dealer Bid Size
JP Morgan Chase Bank 15^ 35
Bank of America 15^ 3
JP Morgan Chase Bank 15^ 12
JP Morgan Chase Bank 15^ 10
Morgan Stanley 15^ 117.583
JP Morgan Chase Bank 15^ 20
Morgan Stanley 15^ 117
UBS 15^ 25
JP Morgan Chase Bank 15^ 117.583
JP Morgan Chase Bank 15^ 5
Morgan Stanley 14.5 25 Goldman Sachs 14.5 15 Citibank 14.5 6 Goldman Sachs 14.375 10 Deutsche Bank AG 14.25 3.135 Deutsche Bank AG 14.25 62 Goldman Sachs 14.25 117.583 Deutsche Bank AG* 14 2
The Royal Bank of Scotland* 14 2
Citibank* 14 2
JP Morgan Chase Bank 14 10
Barclays 14 10 Barclays 13.75 25 Barclays 13.5 25 Goldman Sachs* 13 2 UBS* 13 2 BNP Paribas* 13 2 Deutsche Bank AG 13 47
JP Morgan Chase Bank* 13 2
Bank of America* 12.5 2
Credit Suisse* 12.5 2
Morgan Stanley* 12.5 2
Barclays* 12.5 2
Barclays 12 10
JP Morgan Chase Bank 12 25
Citibank 12 10 Barclays 10.125 25 Credit Suisse 10 5 Barclays 9.125 25 Citibank 9 20 Credit Suisse 8 5
The Royal Bank of Scotland 7 10
Credit Suisse 6 5
12
The list shows that the highest bid is 15 and that the sum of the trade sizes exceeds the NOI of 117.583. Given that the cap is set at IMM plus half the spread (15), the bid is accepted and the NOI is completely filled at the price of 15. The auction final price is than set at 15. Since multiple dealers submitted a bid of 15, the bonds are allocated on a pro-rata basis. The final auction price and submission of order are than published on creditfixings.com.
13
II.3 Auction Protocol
The auction protocol was mainly implemented because of concerns regarding the size of outstanding CDS notional amounts relative to the notional amounts of bonds outstanding. In case of a trigger event this would cause inflated recovery rates, since CDS investors would rush to the market to buy the bond so they could settle the CDS contract. The protocol was implemented in 2005 and was on an optional basis, in case of a credit event, administrators would try to locate CDS investors and give them the option to adhere to the protocol. Adhering to the protocol would amend previously signed CDS contracts, allowing for the auction to determine the recovery rate on which the contracts are settled. Although participation rates were as high as 90%, covering almost all institutional CDS investors. There was a need for hardwiring the auction into each CDS contract. Tracking down al CDS investors and making the auction optional, became too inefficient for the administrators. That is why the “Big Bang” protocol was implemented in April 2009. This protocol supplemented the initial protocol and allowed for hardwiring the auction mechanism in previously signed and future CDS contracts. With the aim towards more standardization into the CDS market. It also introduced the Determinations Committee, a committee consisting of ten voting dealers, two non-voting dealers, five buy side voting member firms ( hedge fund or pension funds) and one non-voting buy-side member firm in addition to a non-voting secretary on behave of the ISDA. Each year the DC members are elected and changed, DC members vote on issues like; whether a credit event has occurred, whether an auction will be held and which bonds constitute deliverable12. On each issue the majority of the votes should be a at least 80% of the total votes for an issue to be passed. Otherwise the issue is passed to the External Review Committee (ERC). The ERC consists of three independent members and reserves the right to vote differently than the majority of the DC.
12
The DC seems to vote always unanimously on whether an auction should be held. Certainly whenever a bankruptcy triggers the contract and a sufficient amount of CDS contracts are outstanding the DC will vote for an auction to be held, in this sense the vote is more a formality.
14
III. Literature review
The following section describes the literature regarding CDS auctions, it also presents the main papers on recovery rates determinants and Chapter filings.
III.1 CDS Auctions
The first who were to study CDS auctions are Helewege et al.(2009). Specifically, they examine a number of auctions conducted until May 2009. They compare market recovery rates one day before the auction with the auction final prices and conclude that they tend to be close to each other13. They also examine the participation rate and NOI in addition to the
IMM. No statistical analysis is done, and of the 43 auctions studied only four actual CDS
auctions are based on the two-stage format, which started in November 2006 with the auction of Dura14. It gives more an overview and simplistic analysis of the auction, rather than an in depth investigation.
Coudert and Gex (2010) studied the auction mechanism in more detail. By analyzing 26 auctions, they show that on average pre and post auction prices are higher than the auction price. However, no in depth empirical results are derived, nor any explanations provided for the apparent mispricing. They examine individual auctions in more detail ( Lehman Brothers, Fannie Mae and Freddie Mac), but their overall focus is on the functioning of the CDS market in an economic downturn.
Gupta and Sundaram (2012) are the first that undertook a detailed empirical investigation. Using a sample 22 auctions for which they have data, they observe that pre and post auction price are higher on average than the auction final price. They use the return between the post and pre auction price as a dependent variable and show by controlling for pre-auction volume traded and auctioned generated public information, that the auction final price normalized by the pre-auction price is a significant predictor of the return. This is used as an indication that the auction is informative. They provide two explanations for the bias in auction prices, one is the Winner’s Curse and two is the strategic behavior employed by market participants through exercising market power. Gupta and Sundaram (2012) test both explanations. First they proxy the Winner’s Curse intensity by using the variance of the first stage price submissions made by dealers and test whether an increase in this proxy causes
13
This is only done for a small subset of their auction sample, since they use only available bond data from Trace.
15
bidders to make lower (higher) bids (offers) in the second stage. This is measured by the degree of bid-shading, a proxy introduced by Nyborg et al.(2002). Results show that the more dispersed the initial market submissions are the more conservative dealers react in the second stage. Other results weakly confirm the Wilson (1979) and Bank and Zender (1993) hypothesis, which suggest that strategic behavior by bidders can result in underpricing in divisible-good auctions. Gupta and Sundaram (2012), also find post auction volatility to increase, which is puzzling if the CDS auctions are informative it should decrease uncertainty.
Chernov et al.(2013) derive a theoretical model of the auction, which is mainly an extension of Wilson(1979), in which there is no asymmetric information, and where the post auction value of the bond is considered to be common knowledge, all participants have the same valuation of the bond. Chernov et al.(2013), show in this model that strategic behavior can result in overpricing and underpricing equilibria. However stating that underpricing is more common after analyzing a small set of auction data, and that the degree of underpricing is positively related to the NOI. Chernov et al.(2013) propose to introduce a pro-rata allocation rule and a conditional price cap to mitigate the mispricing. Another theoretical model devised by Du and Zhu (2013) shows that equilibria are possible in which recovery rates are biased. They argue that the current auction design prevents full participation of investors, which leads to biased recovery rates and inefficient allocations. There model differs from Chernov et al.(2013), since it assumes that traders have different values for owning the bonds and they take into account allocative efficiency. However Du and Zhu (2013), predict the exact opposite of what the data pattern seems to suggest. Specifically, they predict that the NOI is negatively related to the underpricing. The suggest using a double auction design the mitigate biases and allocative efficiency.
16
III.2 Recovery Rates
The earliest research on recovery rates has focused on credit risk models, which make assumptions on recovery value of bonds in case of default or distress. Black and Scholes (1973) and Merton(1974) proposed structural models, where the default probability of firms is impacted by the drivers of its asset value. The probability of default is mainly driven by the asset volatility. Whenever the value of assets is lower than the liabilities, a default occurs. The recovery rate is then the remaining market value of assets, an endogenous variable that is inversely related to the probability of default. This relation is also argued by Frye (2000) and Gordy (2003). Other authors extend the model of Merton (1974)15. By assuming that default can occur at any time between the issue and maturity date of debt, default is set off whenever asset value falls below a certain level. Moreover, bankruptcy cost arise exogenously in these models.
While structural models condition the default of a firm on its characteristics. Reduced-form models of credit risk make use of explicit assumptions16.Specifically, these models make assumptions on the dynamics of the default probability and recovery rate. The recovery rate is usually assumed to be exogenously and independent of the default probability. However both structural and reduced-form models failed explain yield spreads fully17. This is why research has been directed towards explaining the stochastic nature of recovery rates, so to provide evidence from past defaults. This would allow for better modelling of default risk.
In light of this Altman and Kishore (1996) showed by using a sample of defaulted bonds, that the original rating of a bond has no effect on its recovery rate, while seniority does. Hanson and Schuermann (2004) show similar results. Both papers provide evidence that recovery rates are lower during recessions and tend to be different across industries. Altman et al.(2005) also finds that average annual recovery rates and default rates are negatively correlated. Other macroeconomic variables such as GDP and GDP growth rate are not that predictive as suggested by early theoretical papers. Acharya et al.(2007) also show recovery rates to be differing across industries. Shleifer and Vishny (1992) argue that one of the reasons why recovery rates might differ across industries, is because, when a whole industry is in distress, it becomes harder for defaulted firms to sell assets to competitors.
15
Acharya et al.(2006), Goldstein et al.(2001) , Collin-Dufresne et al.(2001) and Mella-Barral and Perraudin (1997), are some examples.
16
Madan and Unal (1998), Lando (1998), Duffle and Singleton (1997) are some examples. 17
17
They also document the negative relationship between the default rate and the recovery rate. Bris et al.(2006) and Davydenko and Franks (2008) show that when there are differences between the rights creditors have and reorganization procedures applied, recovery rates tend to reflect these differences at the time of resolution18.
Jankowitsch et al.(2014) investigate recovery determinants empirically. The main motive is that previous research has focused on rough proxies for market recovery rates. For example Acharya et al.(2007) use the prices of the bonds at the time of emergence from default, which could influence their results. Since emerging from a default can take up a couple of years after the initial filing. Similar shortcomings are also recognized in other studies by Jankowitsch et al.(2014). They also analyse the impact of liquidity and use more bond- and firm specific variables. Providing a detailed analysis of determinants of recovery rates. More importantly, they find that the relevant trading period is 30 days after default19.
18
This is done by analysing ultimate recovery rates. The true recovered value of the defaulted bonds.
19
This also happens to be the settlement period for CDS contracts. Moreover, credit rating agencies like Moodies and Standards & Poor’s also use the 30 days period to calculate an implied recovery rate.
18
IV. Research Method
This section elaborates on the hypotheses, data collection procedure and the methodology used in this paper. The first part concerns the hypotheses tested and research questions to be answered. The data collection procedure is discussed in the second part of this section in addition to presenting the main descriptive statistics. The last part presents the methodology.
IV.1 Hypotheses
The small yet growing literature on CDS auctions seems to suggest that the auction format produces biased recovery rates on which CDS contracts are settled20. Although some empirical research is done on this, using mainly samples containing only auctioned bonds21. In depth empirical research while controlling for auctioned securities remains non-existent. Although the price patterns on days surrounding the auction day clearly suggest that the auction produces biased prices when using market recovery rates from auctioned securities only. This paper may still allow for better understanding of this bias, by analyzing recovery rates of a sample of firms filing for bankruptcy in the period 2003-2014. This sample contains a mix of firms with auctioned and non- auctioned bonds. And controlling for recovery determinants allows for better disentanglement of the auction effect. Moreover, recent literature for instance Jankowitsch et al.(2014). Provide a relevant trading period for defaulted bonds, which in turn allows for a better proxy for the market recovery rate on a given bond. The main research question in this paper, is whether CDS auctions impact recovery rates. Normally, it should not since the auction is devised to ensure fair and efficient pricing. However theoretical predictions made by Chernov et al. (2013), suggest that CDS auctions can result in biased recovery rates. Specifically, Chernov et al. (2013) suggest that either overpricing or underpricing equilibria are possible. But they found that the NOI is positively related to the level of underpricing, when using a small sample of auction data. This paper tests this relation by testing the following hypothesis.
H1A: The recovery rate of auctioned bonds is lower relative to non-auctioned bonds in case the NOI is positive and sufficiently high.
20
See Chernov et al.(2013) 21
19
Given the analyzed price patterns by previous studies, it is expected that NOI is negatively related to the recovery rate22. So specifically the next hypothesis is tested.
H1B: The NOI is negatively related to the recovery rate of a auctioned bond.
If the previous mentioned hypotheses are not rejected and thus the auction mechanism impacts recovery rates. The relationship found should not hold if the pre and post market recovery rate is substituted as the recovery rate of the auctioned bond. Yielding the following hypotheses.
H2A: The market recovery rate of auctioned bonds does not differ from non-auctioned bonds given that the NOI is positive and sufficiently high.
H2B: The NOI is not related to the market recovery rate of an auctioned bond.
These are the main hypotheses tested in this paper. They need to be all not rejected in order to conclude that the auction impacted recovery rates of defaulted securities. Specifically, hypotheses 1A and 1B relate underpricing or overpricing to the NOI, and predict it to be negatively related to the recovery rate. While hypotheses 2A and 2B, make sure that it is related to the auction mechanism, because the NOI cannot be related to pre or post market implied recovery rates of auctioned bonds. Coudert and Gex (2010), discuss the possibility of arbitrage trading by market participants briefly. Given the optional nature of the auction regarding physical and cash settlement. It allows for net protection sellers (likewise for net protection buyers) to potentially optimize between the two options, and whenever they expect the bond to be underpriced to opt for a physical settlement, so as to buy the bond at the auction final price. This is also true for second stage bidders ( in the second stage no restrictions are made regarding CDS market position), since they can get defaulted bonds, as long as they make bids that are close to the cap. As a result, it is expected that trading of defaulted bonds increase post-auction. Thus testing the following hypothesis23.
H3: Post auction trading increases relative to pre-auction trading .
22
See Gupta and Sundaram (2012), who show that for four auctions where the NOI was negative, the pre and post auction market recovery rate was higher.
23
Chernov et al. (2013), argues market participants can influence the price through the NOI and the orders they make, this would yield increased trading somewhat proportionial to the NOI.
20
Since hypotheses H1 till H2 imply that the mispricing is proportional to the NOI, which suggests the following; whenever the absolute value of the NOI greater is than zero, auction participants are able to exploit resulting mispricing through the secondary market. This yields another testable prediction, namely that trading right after the auction increases sharply, which is proportional to the absolute value of the NOI. This hypothesis is not explicitly stated but tested and part of hypothesis H3
IV.2 Dataset
This section describes the data collection procedure and how the recovery rates are determined, while also providing the main descriptive statistics and discussing them.
IV.2.A Data
This paper uses several data sources combined to analyze recovery rates. First, defaulted firms are identified through the Lopocky Database24. This database contains information on US firms filing for Chapter 7 and 11 over an extensive period. It gives a snapshot of the balance sheet of the firm in question at the moment of filing, in addition the database provides some filing characteristics. The most frequently occurring form of bankruptcy, is Chapter 11. This is not that surprising, according to Gilson et al. (1990), most bankruptcy procedures are ruled under Chapter 11. Bris et al. (2006) also shows that larger firms tend to file more often for Chapter 11 while smaller firms file more often for Chapter 7. The Lopocky
database only includes firms that have assets worth of more than 100 million measured in
1980 dollars and filled an annual report with Securities and Exchange Commission (SEC) not less than three years prior to the filing. This might explain the excess representation of Chapter 11 filings in the database. However, small firms are less likely to issue bonds, and no recovery rates can then be extracted for those firms in any case. By using the Lopocky
Database as a reference for firms going bankrupted, transaction data is collected for any
bonds issued by each specific firm. This data is collected through the Trade Reporting and
Compliance Engine (TRACE) database. The month after the filing date is used as relevant
trading period for the recoverable value, this is in line with Jankowitsch et al.(2014). Another reason is that the degree of information generated, might be equal across defaults for a specific length of a period ( in this case 30 days), and since the auction takes place after at most 30 days the credit event takes place, the 30 days is treated as a relevant benchmark. The
24
21
data collected consists of transaction size expressed in notional value, and bond characteristics. All bond prices that are extracted from TRACE are clean prices, since defaulted bonds are traded flat, so without exchange of accrued interest. Jankowitsch et
al.(2014) argues that bonds with complex structures have different prices during a default
scenario, which might bias the result. These are mainly bonds with embedded derivatives features and should be excluded from the analysis. However, Jankowitsch et al.(2014) also investigate other default types extending their sample and making it more probable to be confronted with this problem. This sample only includes simple straight bonds, simple puttable or callable25. Furthermore some restrictions are made for transaction data to be included. (1) A bond needs to have traded at least three trading days covering the 30 day period after the filing, otherwise it is marked as illiquid and excluded when performing a robustness check concerning liquidity 26.(2) Only trades were the trade size exceeds 100000 are considered, since these reflect institutional trades and are informative. Firms that have multiple bonds that do not differ in seniority, which tend to have prices that are close to each other, and are trading on the same day, are pooled together by using the trade size (notional value traded) to calculate the weighted average price. For all remaining bonds the price over the trading day is calculated as a weighted average of the trade prices using the notional value as a weight. The TRACE database reports trade sizes over 1 million notional value as 1MM+, to overcome this problem it is set equal to 1 million27. Finally this data is merged with the bankruptcy data from Lopocky. The merged dataset is supplemented with Compustat accounting data, since the Lopocky Database does not contain all accounts. The most recent fiscal year with respect to the filing date is used to acquire the data, with a half year gap between the fiscal date and the filling date set as a upper boundary.
For bonds that were auctioned, a slightly different approach is used. First the list of deliverable obligations is used to identify the specific bonds being auctioned28. Secondly,
TRACE transaction data is collected surrounding the auction date. Specifically, the
transaction data starting one day before the auction till the end of the auction week is collected. Since different auctions are held for different seniority levels for the same firm. The prices of specific deliverable bonds for the same auction should be close to each other, however sometimes the deviation is large, threatening to bias the result. Gupta and Sundaram
25
According to Jankowitsch et al.(2014), simple call and put options do not affect the analysis, since call options are deeply out-of-the money in default, and put options offer no advantage, since default events trigger (cross-) acceleration clauses.
26
Jankowitsch et al.(2014), uses five trades as a threshold for 30 days, in this paper trading days are used which is a stronger condition. 27
This is an often used approach see Jankowitsch et al.(2014). 28
22
(2012) made a similar observation and excluded those transactions, the same procedure is applied in this paper. Furthermore the same requirement is set regarding the notional value traded, only trade sizes above 100000, are considered as informative trades. The exclusion effect for auctioned bond has less impact, since institutional investors heavily trade these bonds especially surrounding the auction. In addition we sum notional values traded over different trading days in the respective time window, and doing so only for institutional trades( trades above 100000 notional). The market recovery rate on a given day is then given by using the notional value as a weight and calculating the weighted average price over a given trading day. These weighted average prices represent the market recovery rates for bundles of deliverable obligations, which follows previous research and is in line with the uniform pricing mechanism of the auction. Since the auction also prices a bundle of bonds at the same time and price for a specific reference entity.
In addition to this, some auctioned generated information is collected. The NOI and the auction final price is acquired through creditfixings.com. The Depository Trust & Clearing Corporation (DTCC) publishes the Net Notional CDS Outstanding (NND) for specific reference entities29. This database is used to gather the NND of firms that had their bonds auctioned. In addition the Mergent Fixed Income Securities Database is used to collect the Notional Amount of Bonds Outstanding (NAB). This auction data is than merged with the
Lopocky data. This process yields 166 specific recovery rates for several points in the time
window, for the complete sample. With a coverage of 144 reference entities that experienced a credit event through a bankruptcy filing30. The time period concerns 2003 till 2014, including some of the most recent auctions. This time period is chosen because getting reliable bond transaction data becomes harder for earlier years31.
IV.2.B Composition of the sample
Table 3 presents the descriptive statistics, in Appendix 1 an overview of all variables and computations can be found. The table does not contain descriptive statistics for the dependent variables nor the ones of some other variables used, these will be referred to throughout this paper.
The table shows that Leverage and Log Assets are matched with 166 different
29
The NND of the week prior to the auction of each reference-entity is collected. 30
The sample contain only chapter 11 filings, as expected, Jankowitsch et al.(2014) found that their sample contained 1% chapter 7 filings and excluded those. Like already argued these filings are rare.
31
23
recovery rates, for the additional accounting data acquired through Compustat this is 147 recovery rates, this has to do with the restriction regarding the gap of 6 month described under the data section. It also shows the NOI scaled by NND for the bonds that were auctioned. The average is 0.1753, meaning one average 17.53% of the amount of protection bought is settled through a physical settlement request. The maximum is 88.36% of net protection bought, requested to sell through a physical settlement. The sample also contains recovery rates on loans, these are acquired through creditfixings.com and constitute a small part of the sample (16 out of 166 and 147). Loans are auctioned to settle loan CDS (LCDS) contracts, but are not of primary interest in this paper. Furthermore, the sample contains 34 firms of which the bonds were auctioned, 10 of which filed a prepackaged Chapter 11. Overall the sample exists of 114 non auctioned market implied recovery rates.
Table 3. Descriptive statistics.
This table shows descriptive statistics for the sample containing US firms filing for bankruptcy 2003-2014. For definitions of all variables refer to the appendix. The table provides an overview of the number of observations, mean, standard deviation, median, minimum and maximum for each variable. In addition division between auctioned and not auctioned is provided.
(1) (2) (3) (4) (5) (6)
N Mean Standard Deviation Median Minimum Maximum
Leverage 166 1.1539 0.6467 0.9829 0.5575 5.6791 Log Assets 166 7.8327 1.5904 7.7386 5.6384 13.5229 Equity 147 0.0798 0.1503 0.0259 0 0.9877 Intangibility 147 0.1634 0.2123 0.0587 0 0.86397 Receivables 147 0.1398 0.1526 0.1037 0 0.8570 Profitability 147 0.1383 0.2550 0.0474 -0.0766 1.5820 LTDIssuance 147 0.6315 0.4181 0.8915 0 1 OI 36 0.1753 0.2712 0.1652 -0.3521 0.8836 Sample characteristics
Auction Bonds Auction Loans Not Auction Bonds Not Auction Loans Sum
N 36 16 114 0 166
No Firms 34 16 94 0 144
24
IV.3 Methodology
This part describes the identification strategies used to test hypotheses 1 till 3. First the definition of the market recovery rate following a default is formalized. For non-auctioned bonds, the following market recovery rate is proposed ∑ where
∑
, this is the weighted average of the trade prices over a given trading day, given the restrictions outlined under the data section. is based on the last trading day ( ) following the 30 days after the Chapter filing. In addition is also estimated and is based on one trading before the last trading day within the 30 day window ( ), finally ( ) is estimated. Bonds with no data to estimate ( ) are marked as illiquid and excluded in a robustness check on liquidity. For non-auctioned bonds, is equal to the auction final price. While is the market recovery rate one day before the auction. is also estimated and equals the market recovery rate one day after the auction. Finally is estimated, which equals the average market recovery rate over the remaining days of the week in which the auction is held. All market recovery rates are calculated through the following formula ∑ where ∑
and obey the restriction that the accompanying trade size exceeds the threshold.
For market recovery rates of auctioned bonds the trades of bonds identified as deliverable by the Determinations Committee are pooled together to calculate a market recovery rate. For non- auctioned bond this is only done for bonds, when they are issued by the same reference entity, share similar seniority level, and have roughly similar price levels32. The aim of this paper is to answer the question whether CDS auctions impacted recovery rates, up to now price patterns of deliverable bonds surrounding the auction day seem to imply it does. However previous research has neglected to provide a true disentanglement of the auction implied effect. That is why CDS auctions implied recovery rates in this paper are analyzed in a broader setting. This is done by controlling for the auction effect using the NOI scaled by NND. Chernov et al. (2013), already argued that this variable is positively related to the underpricing. In addition the treatment effect of the auction is isolated by including non-auctioned bonds of after and before the auction protocol (AP). The effect is than isolated by estimating the following two regression ( (1) and (2).
32
25 (1) (2)
Where Rt,i,j denotes the recovery rate on a specific bond i issued by firm j, t denotes which
recovery rate is used. For non- auctioned bonds this can only range from R0 till R-3. Whereas
for auctioned bonds this can be R0 ( i.e. the auction final price) or the market recovery rates R-1, R1 or Raw. The t index allows for different tests and comparisons across regressions. The
firm level vector of controls is represented by Zj and varies across firms but might be the
same for some recovery rates since there are less firms than the number of recovery rates. This is because recovery rates are split up according seniority and type security (loan or bond). For instance different auctions are being held on the same day for different defaulted securities of the same firm. This can be the case if a firm has issued bonds of different seniority levels or issued loans (LCDS settlement) in addition to bonds33. The vector Zj also
contains the dummy prepackagedj which is a filing characteristic, but differs across defaulted
firms. The dummy variable seniori is only one if it concerns a bond that is more senior than
another bond in the dataset that shares the same issuer. This is because seniority only impact recovery rates if they differ across bonds issued by the same issuer. The dummy variable
Bondi indicates whether it concerns a bond of which the recovery rate is explained. The
dummy variable APi indicates whether the recovery rate is implied after the auction protocol
was implemented. Surely, for all recovery rates of auctioned bonds this dummy is equal to one. The dummy variable Auctioni indicates whether the security was auctioned. Finally the
main difference between equation (1) and (2), is how the NOI scaled by NND is modeled34. In equation (1) this variable is modeled as a continuous variable and referred to as OIi , while
in equation (2) it is modelled as a binary variable equaling one whenever the ratio between NOI and NND is positive. If NOI is zero this interaction drops and the auction effect is equal to β2 with respect to bonds defaulting after the auction protocol. This is also true for a
negative NOI, that is why a continuous modelling might be more appropriate, since it is
33
Dura is an example for which two auctions were held, one to settle subordinated bonds and the other to settle senior bonds. 34
26
expected that any absolute value of NOI greater than zero, will result in mispricing. However both variants are estimated in a reduced- form fashion for convenience reasons ( easier to interpret). In equation (1) the difference between recovery rates of auctioned bonds with OIi
being zero and auctioned bonds with absolute value OIi greater than zero is given by β1, the
difference with non- auctioned bonds after the auction protocol is given by the sum of β1 and β2 . Finally the difference with non-auctioned bonds before the auction protocol is given by
the sum of β1, β2 and β3. For equation (2) it is similar except that β1 is the difference with
auctioned bonds when OIi is greater than zero. In both equations auctioned loans are captured
by β0, since only auction data is used for loans, the recovery rate for each loan in each
regression is R0 (i.e. the auction final price). Equation (1) and (2) also allow testing for other
differences in a straightforward way. For instance the difference between the recovery rate of a non-auction bond after and before the auction protocol is equal to β3.
This paper also investigates whether post auction trading increase, more specifically if
OIi predicts a degree of mispricing whether post auction trading increases proportional to OIi..
This is done by using a small window around the auction day. Data is collected on institutional trading in auctioned bonds, the time window is somewhat similar to the one of market recovery rates for auctioned bonds discussed above. Specifically the time window consists of, one day before the auction day, right after the auction closes but still on the auction day (this is not done for the market recovery rate), one day after the auction day and the average of the remaining days of the week. For these points in the time window, notional value traded is aggregated35. Other important variables are the Notional Amount of Bonds Outstanding (NAB) and the number of deliverable obligations. These variables differ across auctions. The NAB is used to scale the notional amount of traded, and the number of deliverable obligations is used to control for, in addition NND is also controlled for. This yields the following equation (3) to be estimated:
(3) | |
In this equation (3), Auctiondayi is a dummy that indicates whether NotionalSCi , is being
traded on the auction day right after the auction. This variable is interacted with the absolute value of NOI over NND (|OIi|). Likewise Dayafteri indicates whether it concerns notional
35
27
value traded one day after the auction. And AWi indicates whether it is an average notional
value traded over the remaining days following the auction day ( dayafter is excluded for calculating this average). In this regression β0 captures pre auction trading levels, the sum of β3 and β4, captures effect of auction day trading with |OIi| greater than zero. The coefficients, β5 and β6 capture the effect of the day after and the remaining days, respectively.
V. Results
This section presents the results of regressions (1) till (3). First the main baseline regression (1) is performed which included all variables and yields 147 observations. Subsequently, reduced forms of both equation (1) and (2) are estimated and compared. This section also includes the results for the estimation of equation (3). At the end this section shows results of different robustness checks performed, two that use different market implied recovery rates for non- auctioned bonds and one that uses only liquid non-auctioned bonds and excludes illiquid ones.
V.1 Results Regression one
This section contains the main results derived using regression (1), it than shows results of reduced forms of both regression (1) and (2). Finally the results of regression (3) are presented and discussed.
V.1.A Baseline regression (1) results
In Table 4 the baseline regression is estimated, for non-auctioned bonds the recovery rate is set equal to R0 ( i.e. the last trading day in the 30 day period after the filing), this is the case in
all four regressions. The difference across columns is the recovery rate used for auctioned bonds, in column (1) it is the market recovery day one day before the auction, namely R-1,
column (2) uses the auction final price as a recovery rate for the auctioned bonds (R0). While
column (3) uses the market recovery rate one day after the auction for auctioned bonds (R1).
In column (4) the average market recovery rate of the remaining days of the auction week is used (Raw). Appendix two contains a correlation table for the independent variables and
descriptive statistics for the dependent variables used in the coming regressions. There seems to be no multicollinearity problems, with correlation being at most 0.5 across all variables.
28
Table 5, contains descriptive statistics for the dependent variables used in Table 4, it shows that the average recovery rate is lowest for column (2) the regression were R0 is used for the
auctioned bond ( i.e. the auction final price). This is not so surprising and in line with previous studies. Table 4 estimates regression (1) as outlined in the methodology part and thus models the NOI over NND as a continuous variable. The number of observation equals 147, since not for all firms Compustat data can be acquired to supplement the Lopocky data, see Table 3 for details. The results in column (2), indicate that OIi, is negatively related to the
recovery rate and significant at 5%. The coefficient of the interaction term (b) is negative and significantly different from zero. This coefficient measures the difference between the
auctioned bond with zero NOI and auctioned bond with a non-zero NOI. The coefficient is equal to -23.4399, this means that if OIi, is equal to one, which can only be true if all net
protection buyers make use of the physical settlement option within the auction. The underpricing is the most severe and equals 23.44% (on 100 face value) compared to a bond that is auctioned with a zero NOI. Compared to non- auctioned bonds after the protocol implementation the underpricing is less severe, since the difference is equal to the sum of (a) and (b). The F statistic of this sum is provided on the bottom row. For column two, it
indicates that the sum is significantly different from zero, indicating that relatively to non- auctioned bonds after the protocol implementation, auction bonds with a sufficiently high enough OIi, are underpriced. The coefficient (a) is insignificantly different from zero
indicating that there is no significant difference between non auctioned bonds post AP and auctioned bonds with zero OIi,, which can only be true if NOI is zero. In this case the auction
final price is equal to the IMM and the market clears directly after the first stage. Moreover, the results indicate that there is no significant difference between non auctioned bonds before and after the auction protocol implementation. This means that an auctioned bond with OIi,
sufficiently high enough, is underpriced with respect to all different bonds. Given the small size of the sample, a lot of the controls are insignificant, however leverage is negatively related to the recovery rate and significant at 1%. Senior bonds and prepackaged filings produce higher recovery rates both significant at 1%. And bonds produce lower recovery rates compared to loans (measured by the coefficient of bond) which is also significant at 1%36.
36
29
Table 4. Baseline regression equation (1)
This table shows the results of regression equation (1) as formulated in section IV.3. Where the recovery rate for non-auctioned bonds is taken as the last trading day in one month after the filing and used for all four regressions (R0). The auction recovery rate is R0, R-1 is the market recovery rate one day before the auction and only applicable for bonds that were auctioned. Raw is the average recovery rate per day over the remaining week after the auction. All the control variables are defined as in Appendix 1. The sample contains US firms filing for bankruptcy in the period 2003-2014.The T-statistics are presented between parentheses. The standard errors are clustered on the firm level.
(1) (2) (3) (4)
Recovery rate auctioned
bonds: R-1 R0 R1 Raw Leverage -7.2129*** -6.7419*** -7.1157*** -7.2267*** (-3.29) (-3.04) (-3.22) (3.28) Log Assets -0.9213 -0.7469 -0.8648 -0.9236 (-0.47) (-0.38) (-0.44) (-0.47) Equity 14.8219 14.8385 14.8991 14.8793 (0.69) (0.70) (0.70) (0.70) Intangibility -2.7042 -3.3234 -3.3545 -3.3474 (-0.27) (-0.32) (-0.32) (-0.32) Receivables -6.6322 -7.5189 -7.2632 -7.3628 (-0.43) (-0.49) (-0.47) (-0.48) Profitability 13.5613 12.8649 12.6349 13.2889 (1.38) (1.33) (1.31) (1.37) LTDIssuance -8.6664 -8.1043 -8.2135 -8.3177 (-1.51) (-1.41) (-1.43) (-1.45) Senior 31.9261*** 32.06*** 32.1845*** 32.1352*** (3.95) (4.01) (4.03) (4.02) Prepackaged 16.8467*** 17.2529*** 17.2585*** 17.2410*** (3.70) (3.76) (3.76) (3.76) Bond -30.1201*** -29.5548*** -29.9659*** -30.1201*** (-3.24) (-3.20) (-3.24) (-3.25) Bond*AP -7.5293 -7.5779 -7.5885 -7.5758*** (-1.26) (-1.27) (-1.27) (-1.27) (a) Bond*AP*Auction 6.1381 5.030 4.8396 5.0741 (0.95) (0.77) (0.75) (0.79) (b) Bond*AP*Auction*OI -18.51486* -23.4399** -18.5339* -18.0252* (-1.88) (-2.38) (-1.83) (-1.79) Constant 75.1214*** 72.5737*** 74.3186*** 75.0245*** (3.35) (3.25) (3.32) (3.35) Observations 147 147 147 147 R-squared 0.3710 0.3797 0.3763 0.3764
P-value of F test for (a+b)=0 0.1233 0.0191 0.0963 0.1102
F-statistic (2.40) (5.63)** 2.85* (2.59)
*
30
Table 4 shows that the regression performed in column (2) slightly outperforms the others in terms of explanatory power. Also, the apparent difference between auctioned bonds with a zero and nonzero NOI noticeable in column (2) dies of in column (1), suggesting that the market recovery rate implied by the market one day before the auction takes place, is the fair value. The magnitude of coefficient (b) also decreases by roughly 5% of the face value. Hence the F-statistic testing the sum of (a) and (b) sharply decreases, yielding no significance. This dying off effect is also apparent for auctioned bond market recovery rates post auction. This is shown in column (3) and (4) the underpricing seems to persist a bit one day after the auction, however dying of completely towards the end of the week. An explanation for this could be provided by testing regression equation (3), which tests whether trading increases on the auction day and whether this is proportional to the absolute value of
OIi,.
This would indicate that right after the auction the trading is dominated by auction participants, who tend to trade at prices close the auction final price. This is because, they have a couple of days to settle the physical settlement request they engaged in, which allows for auction induced trades. Net protection sellers buys bonds at the auction final price, while a net protection buyer sells bonds at the auction final price, but that does not necessarily mean that the net protection buyer holds the bonds. Since a settlement period of three days exist following the auction, participants can speculate on the price.
Moreover, auction participants can request to buy the bond in the second stage by submitting limit bids through participating dealers ( participating dealers also reserve the right to trade on their own account), this at least allows for some arbitrage opportunities , since auction participants know that limit order close to the IMM get filled. By observing the market prices or locating counterparties through the secondary market auction participants can lock in a profit. However, the prices produced by the auction are also observable by non- auction participants. So investors that have a higher valuation of the bond, might want to profit from the underpricing by offering less. This might explain the delayed correcting of recovery rates as observable in column (3), while towards the end of the week only investors with higher valuation trade in the specific bonds. Graph 1, plots the mean recovery rate for a sample of auctions conducted. Day 0 gives the auction final price, day -1 gives the implied mean market recovery rate one day before the auction, similarly for day 1, however day 2 represents the mean of the average recovery rates over the remaining days of the week. This graph confirms the trend in Table 4.
31
Graph 1: Mean recovery rate of auctioned bonds.
V.1.B Reduced form regressions.
The next table shows the results of the reduced form regressions, the reduced form regression equation (2) is also estimated. This is to see how it captures the variation, and whether it passes the different tests. Overall the table shows that the average explanatory power is roughly 30%. So excluding insignificant variables from the main regression does not take away that much explanatory power. For each different recovery rate of the auctioned bonds, two reduced form regressions are estimated, one using a continuous modelling of OIi, while
in the other one, OIi is modeled as a binary variable.
Column (3) and (4) report the results where the recovery rates for auctioned bonds is equal to the auction final price, the coefficient on (b), for (3) is greater in magnitude than in column (4), however this variable is modeled as a continuous variable. The sum of a and b seem to be significantly different from zero for column (3) at the 5 percent level, and for column (4) at the 10% level, the continuous variant seems the capture the variation slightly better.
Notice that the magnitude of (a) changes across the two regressions quite a lot. This because Bond*AP*Auction In column (4) incorporates bonds that were auctioned and had a negative NOI, which have higher auctioned produced recovery rates than the associated market recovery rate pre and post. Overall the results seems to confirm the findings of Table 4 in general, with the underpricing fading away nearing the end of the week. However these results do suggest a bigger effect of NOI, which could be due to the reduced form. Like Table 4, Leverage, Senior, Bond and Prepackaged are all significant at the 1% level across all columns 24 25 26 27 28 Me a n R e co ve ry -1 0 1 2 Day
