The counterparty risk of ETPS : is counterparty risk priced in the market?

(1)

Master Thesis

2017

THE COUNTERPARTY RISK OF ETPS – IS COUNTERPARTY

RISK PRICED IN THE MARKET?

JOEP BRAAS - 10899685

(2)

1

ABSTRACT

Counterparty risk on Exchange Traded Products (ETPs) has been the subject of multiple papers. For example the composition of the assets backing an ETP, have been researched with a view to collateral shortfall. Also potential harmful implications of the financial innovation within the ETP asset class has been theoretically put apart. However little research have been done so far on the market impact in ETPs during times of increased financial distress. This paper seeks to empirically proof that counterparty risk on ETPs exists by assessing the relationship of credit worthiness of the issuer’s parent company. The research is done by regressing the CDS of the ETP’s ultimate parent institution on the bid/ask spread of the ETPs. In addition, differences in effects between ETP types are assessed. This paper uses data from 483 Exchange Traded Funds (ETFs) and Exchange Traded Notes (ETNs) in the form of panel data. A Logit model is used to estimate the probability on a “high bid/ask spread” on the condition of a “high CDS level” by taking percentiles of both its distributions. The model accounts for commonly known factors that have shown to be of significance for the bid/ask spread like volume, market- and FX-volatility. The CDS is of significance for both ETNs and ETFs. The effect is strengthened if the ETP is an ETN. Also the effect is strengthened for ETFs that perform Securities Lending or use a Synthetic replication scheme compared to ETFs that do not apply this. This result shows that counterparty risk on ETPs is recognised by market participants, and can be used for further research on the underlying causes of counterparty risk.

(3)

2

1. INTRODUCTION

Exchange Traded Products (ETPs) are financial instruments traded on a stock exchange whereby typically the aim is to provide the same return as a specified benchmark or asset. Liquidity is provided by authorised participants and market makers. The product has a passive investment strategy and is a cost-effective way of gaining exposure to a benchmark. (ETF Securities (2017), ESMA (2015))

As we have seen in earlier financial crises, innovation in structured products has shown to be a potential kick-starter of global financial turmoil. A lack of transparency on the underlying assets backing structured products combined with the complexity of certain structures, make counterparty risk assessment more difficult (CGFS, 2008). These assessments have been made for example after the sub-prime crisis of 2008, where mortgage backed securities created excessive degrees of leverage, intransparency, complexity and interconnectivity. The exchange-traded product (ETP) and its features meet these characteristics, especially when a synthetic replication scheme is applied or securities lending is performed. But ETPs can also contribute to the propagation of local anomalies, such as the Flash Crash on May 6, 2010. After the flash crash, Ben-David et al. (2011) presented evidence that the ETP served as a conduit for shock transmission between the e-mini S&P500 futures and the equity markets, thereby exposing the level of underlying liquidity in times of stress. Concerns were raised by, amongst others, the Financial Stability Board (2011), International Monetary Fund (2011), Ramaswamy (2011), Bank of England (2010) and FSB (2011) and risks related to ETPs have been subject to various research. This thesis will focus on one of the conclusions by Ramaswamy (2011): the counterparty risk related to an ETP can trigger a run on the ETP in periods of heightened counterparty risk.

The ETP has become one of the primary products of not only institutional investors such as pension funds but also central banks and less sophisticated retail investors. The ETP used to be a plain vanilla, flexible, transparent product, traded similar like stocks on-exchange. However this turned into a more complex financial product by the inclusion of derivatives and different fund structures. Investors seeking for cheap ways to gain exposures to illiquid and exotic markets, required more innovative solutions from the ETP issuers. In addition, there are incentives for the sell-side

(5)

4 institutions for financial innovation in the ETP space due to regulatory requirements such as liquidity coverage ratios. Synthetic replication schemes, using total return swaps, allow ETP issuers to exactly mimic the reference index for a fixed price. In addition, the swap counterparty is able to perform balance sheet management by allocating certain assets as collateral towards the issuer. Amongst others, the innovation led to higher competition, lower fund management fees and access for the investor to a wider range of financial indices.

The benefits of the synthetic replication and other complex financial structures come at a cost. According to Ramaswamy (2011) it increases the risk to instability in the financial market. Large size liquidation of synthetic ETPs in periods of higher counterparty risk not only puts pressure on the liquidity of the ETP and its underlying assets. Also it could force swap counterparties to sell the collateral, often illiquid assets, which could hinder correct functioning of the market. For example due to the liquidation of the ETP, the issuer is required to reduce the total notional on which the total return swap is agreed. This implies that the swap counterparty will repay the issuer in cash and the collateral basket will be returned from the issuer to the swap counterparty (i.e. the swap counterparty will be hit on his balance sheet and in terms of liquidity).

Synthetic replication schemes protect the investor from the tracking error, the difference between the return of the ETP and the return of the reference index, but in return the synthetic replication scheme exposes the investors to counterparty risk towards the swap providers. It provides an extra risk dimension investors should take into consideration. (Bank of England (2010), FSB (2011), Ramaswamy, S (2011)).

The conclusion from previous research on the counterparty on ETPs varies. Ramaswamy (2011) did show some examples where the collateral composition of the ETP (usually over-collateralized) has very little overlap with the composition of the index however Hurling et al. (2014) did show that the allegations made by international agencies about high counterparty risk on ETPs can be unjustified as the collateral matches the characteristics of the reference index (in their sample with db-X trackers where the characteristics are equity or fixed income). Later on, in a renewed research, Hurling et al. (2016) show that ETP flows respond significantly to changes in counterparty risk, where the default probability is measured by the CDS. Theoretically there seems to be more than sufficient ground to state that there is

(6)

5 significant counterparty risk in holding an ETP, especially one that uses a synthetic replication scheme and/or is uncollateralized. Elia (2012), Naumenko and Chystiakova (2014) found that synthetic ETPs have a higher tracking error when compared with the funds that follow a physical replication method and suggest that this could be explained by a higher degree of counterparty risk as the synthetic characteristic should provide the ETP a smaller tracking error.

Empirically the impact of counterparty risk could be difficult to investigate since a systemic crash, fortunately, doesn’t occur often. Nonetheless, the area that could show signs that this risk exists and investors are requesting a risk premium for it, is if it is priced in the market. Therefore we will examine whether there is an impact on the market prices based on changes in the credit default spread of the ultimate parent financial institution. A similar approach of Hurling et al. (2016) will be used, accept that it will be applied on market prices instead of fund flows. Given the disagreement within the field and the possible complexity of the actual workings per ETP structure, determining differences in market impact within these high level variables will be valuable for further research on this topic.

This paper first sets out the theoretical framework in the chapter Literature review. Subsequently the section Research question and methodology discusses the choice of sample and the hypothesis as well. The Data section provides data descriptive and a description of the dataset being used. The Result section contains the results from the Panel and Logit model. The Robustness section describes various checks that have been performed to test the robustness of the estimated model. The section For further research provides an overview of areas for improvement from this research and ideas for more in depth research. The paper is wrapped up in the section Conclusions.

(7)

6

2. LITERATURE REVIEW

The theoretical framework of this thesis is built on a working paper by Ramaswamy (2011), the Financial Stability Board (2011) and the Bank of England (2010) which highlights the counterparty risk of ETP’s and the resulting risk to the financial system as a whole. These papers has been picked up by several regulating entities, international agencies and the press. However, as indicated by Hurling et al. (2016) who researched the relationship between the CDS and the fund’s assets under management, very few empirical research have been done so far on the impact in market prices. I have identified multiple variables that determine the structure (or lack) of collateralisation of the ETP: ETP type, replication type and structure and securities lending. The level of collateralisation of the ETP and the legal structure around it has shown to be an important factor in the aforementioned papers.

ETP types:

There are three types of ETPs: 1) Exchange Traded Fund (ETF), 2) Exchange Traded Commodity (ETC) and 3) Exchange Traded Notes (ETN). The most important characteristics with regard to counterparty risk are displayed in table 1. Theoretically ETFs are least risky being an open-end investment fund with segregated assets in a SPV (special purpose vehicle) entity. In addition, UCITS regulation sets more strict rules in terms of transparency, level of collateralisation and diversification in invested funds. ETCs (theoretically riskier) and ETNs (theoretically most risky) are, contrary to ETFs, debt instruments. The ETC is collateralised with segregated assets in a SPV and the ETN is an unsecured debt instrument issued either directly by the financial institution (market practice in the US) or in a SPV (market practice in Europe and APAC). (Hill et al. (2015), ETF (2017) and ETF Securities (2017)).

Type Type Entity Collateralized Seg. assets UCITS

ETF Investment vehicle SPV Y Y Y

ETC Debt security SPV Y Y N

ETN Debt security Bank or SPV N N N

(8)

7 Replication type:

On a high level, there are two methods the ETPs will seek to replicate its benchmark: (1) the physical ETP, where the issuer uses the cash of the investor to buy the constituents of the benchmark and therefore mimicking the index, and (2) the synthetic ETP, where the ETP issuer uses derivatives, such as a total return swap, to replicate the return of the benchmark. Within the synthetic ETPs, there are two ways the issuer can enter into a collateralized total return swap: the funded swap and the unfunded swap which each contain different characteristics (see graph 1).

Graph 1: Funded and unfunded structure

The unfunded swap is the most commonly used structure. The swap provider (commonly the parent institution) commits to deliver the return of the index and sells a basket of securities to the ETP issuer. Once the marked to market of the collateral basket gets below a certain threshold compared to the marked to market of the index, the swap provider will deliver additional collateral. In the funded swap model the ETP issuer transfers cash to the swap provider who in return enters into an agreement to pay the return of an index plus the principal amount at a future date. To offset the credit risk for the ETP issuer, the swap provider places collateral in a segregated account at a 3rd party (custodian). The main difference between the unfunded and the funded swap is that for the unfunded swap the collateral lays directly with the ETP issuer. In case of an ETN, the swap (if any) is uncollateralized. In

(9)

8 case of a physical replication scheme, the same assets as the ETP deems to track lay directly with the issuer (if no securities lending is performed). This shows that ETP’s using a synthetic replication scheme theoretically could run more counterparty risk. (Hill et al. (2015), Blocher and Whaley (2015), Ramaswamy, S. (2011)).

Hurling et al. (2016) did show that the fund outflows are significantly higher for synthetic replication schemes than for physical replication schemes if the issuer has an increased default probability. Also Elia (2012), Naumenko and Chystiakova (2014) suggest that the higher degree of tracking error in ETPs using synthetic replication scheme compared to physical is caused by the higher degree of counterparty risk. Securities lending:

Securities lending is a transaction used by institutions to make better use of their assets. Typically it is a bilateral transaction where one party transfers securities, such as equities or bonds, to the other party. It is common that in return the lender receives collateral (cash or non-cash) from the borrower. ETP issuers can use this technique to generate additional income. (Hill et al., 2015)

Risks can arise from securities lending due to increased interconnectedness and opacity, according to Dive et al. (2011). Re-hypothecation (lending out of assets that are lend or received as collateral) of the collateral received by lenders will make the situation even more complex. In the case of Lehman’s default in September 2008, it led to a complex bankruptcy case in the U.K. as the collateral received on the lent out assets where re-hypothecated. Consequently the assets remained frozen until the structure was disentangled, according to Singh and Aitken (2010).

Ramaswamy (2011) focussed on the synthetic replication scheme and Hurling et al. (2016) didn’t research the impact on ETPs performing securities lending as well. Theoretically securities lending could cause the same increased exposures as a swap would cause and therefore it is included in this research. The issuer encloses whether the ETP’s assets are used for securities lending, this will be used as a variable in the regression.

Creditworthiness:

As indicated in the Introduction section, there are incentives for the issuer (both in terms of revenues as in balance sheet management) to allocate as much swap business as allowed to its parent company or affiliated bank (Ramaswamy, 2011). In

(10)

9 this research the ultimate parent company is used to derive the credit worthiness data form.

RiskMetrics (2000) describes counterparty risk for an ETP investor as an equivalent to the risk that the ETP is not or not fully collateralized at the point of default of the issuer and/or swap provider. It is the current replacement cost marked to market, plus an add-on to reflect potential future exposure. The FED and the UK FSA apply different factors for the add-on for different asset types such as equity, fixed income and commodity but also geographical exposure. Alternatively, the historical or implied volatility can be used to take a measure of potential future exposure.

Hull (2012) provides the following on default probabilities: 𝜆̅ = 𝑠

1 − 𝑅

Where 𝜆̅ is the average default (hazard) rate per year, 𝑠 is the default spread and 𝑅 the recovery rate. We can conclude from this that two factors are of importance for the ETP holder: 1) the default spread and 2) the recovery rate. Hurling et al. (2016) uses a different equation but with similar factors. They take the probability of a default on the condition of a collateral shortfall (i.e. recovery rate unequal to 1):

𝑆 = 𝔼(∆ | ∆ > 0, 𝐷 = 1)

Where 𝑆 is the required risk premium for the ETP holder, ∆ is the collateral shortfall and 𝐷 is the default of the issuer of counterparty of the issuer.

For 𝑠 or the default spread, we will take the credit default spread (CDS). It is market practice to use the change in 5-years CDS as a measure for the probability of default, which is also used by Hurling et al. (2016). Here the assumption is made that, separate from the collateralization, the claim of the issuer on the swap provider in case of a default, all share the same credit seniority and the CDS is representative of this seniority. Also we will assume that the 5-years CDS is a good representative of the default spread (regardless of the tenor). The 5-years CDS is the most commonly tenor and therefor most likely to give a reliable signal on the credit worthiness. (Deutsche Bank, 2012)

Collateral shortfall is an equivalent of the recovery rate from standard credit risk literature. Due to simplification we can treat them as equal is this research which will

(11)

10 be explained later. The collateral shortfall or recovery rate works slightly differently for ETFs and ETNs.

1) For ETFs the collateral shortfall could arise when the characteristics of the collateral do not match the characteristics of the benchmark and a different return than expected (the benchmark index) is achieved. In other words, the marked-to-market value of the collateral basket is less than the net asset value of the fund. Hurling et al. (2014) show that the collateral has the same (high level) characteristics as the ETP, however Ramaswamy (2011) showed us (in more detail) otherwise. It is a UCITS requirement for the issuer to perform daily collateral management which implies that a mismatch could arise at least over a short timeframe.

2) The ETN is an unsecured debt instrument and thus will relate on the general recovery rate of the issuer for unsecured debt holders.

Recovery rates can be estimated using the averages per rating or by solving for the implied recovery rates from bond prices and default spreads (Hull 2012). Collateral shortfall can be modelled as well. For example by modelling the expected returns for the collateral baskets versus the expected return of the ETF. These methods are omitted in this research however as it can be approached by a simplified approach first. If the CDS is proved to be significant, it implies that the recovery rate is unequal to 1. With a recovery rate of 1, the probability of default would not have an impact. Subsequent models on estimating the recovery rate or collateral shortfall will be feed for further research.

Volatility:

Previous empirical studies (Chiang (1998), Elton, Gruber, Comer, and Li (2002), Frino and Gallagher (2002)) identify the main factors that give rise to price discrepancies, including transaction costs, index-composition changes, corporate activity, fund cash flows, the reinvestment of dividends and index replication strategies. The significance of these variables could be explained by the technical difficulties (such as tax and dividends) when measuring the tracking error and therefore are less interesting for this study. Instead, we use the volatility of the market index and the volatility of the FX rate (Shin and Soydemir (2010)) as variables in order to account for the effect of an increase in overall volatility that could give rise for an higher degree of price discrepancy.

(12)

11 Market impact:

In order for an ETP to track the benchmark index, the price is being maintained by authorised participants (APs). An AP can be seen as a market maker who is constantly quoting a buy and a sell price. Contrary to market making in stocks or options, an AP is able to create or redeem the ETP. This implies that the AP is able to exchange a basket of stocks or cash for the ETP (or visa versa). A typical transaction is that the market maker buys (or sells) the ETP on-exchange and hedges it with the index futures up until the position is unwind through creations or redemptions at the issuer. This implies that, if the return of the ETP will move in line with the index, the market maker is hedged. However, if the market maker would hold the ETP and the ETP issuer or its swap counterparts would default, the holder of the ETP is entitled to the collateral posted at the issuer, with possibly a different return from the index. Crisis experience has also shown that the collateral assets pledged by a failed swap counterparty could be frozen by a bankruptcy administrator even when they are held in segregated accounts (Fender et al., 2008). This implies that there is a risk the ETP holder will receive a different return on his ETP than on his index futures which can result is a loss. In addition, the market maker would have a significant part of his assets stuck in a frozen account and therefore would suffer from opportunity loss during times money should be made. In case of uncollateralized ETPs, the investor is depending on the bankruptcy settlement.

Theoretically, the aforementioned are incentives for the market maker to keep their long positions in ETP’s with stressed issuers as small as possible which can be achieved by adjusting their bid ask spread skewed to sell. Other market participants would also be more likely to sell their long positions and thus providing additional cause for a wider spread. Nollen et al. (2002) show that a significant factor of the bid/ask spread is the inventory-holding premium. This is a premium that represents compensation for price risk borne by the market maker while the security is held in inventory. This implies that it can be expected that market makers request an higher risk premium (i.e. wider bid/ask) in case of an increase in counterparty risk.

(13)

12

3. RESEARCH QUESTION AND METHODOLOGY

This chapter focusses on the research question and how the choice of sample has been determined. Also the methodology is explained, a panel data model and a logit model. Subsequently the hypothesis for each model is provided.

3.2 Research question

This research will focus on the following research question:

What is the relationship between counterparty risk on ETPs and the credit worthiness of the issuer or parent company?

In addition, it is tested whether the following factors have an effect on the relationship: ETP type, replication type and securities lending.

3.2 Choice of sample

We are investigating the impact of credit worthiness on ETPs and possible differences between different characteristics, therefore we have tried to use a wide as possible scope to arrive at sufficient ETPs per category. Bloomberg is used to determine the initial scope of the research, containing a list of all 8,522 listed ETPs being traded word-wide in April, 2017.

The use of weekly data allows for the detection of market impact during a relatively short period but also enables us to have a wide timeframe. As most ETPs are liquid products, taking monthly data increases the chance of the impact remaining unnoticed and thus not significant. Daily data is not used as it would make the dataset too large for this research (or less ETPs would have had to be used). Implications for not using daily (or even intra-day) data are given in recommendations for further research.

In order to have a wide variety of issuers included during a timeframe with sufficient variability in default risk, checks have been performed on the number of issuers with

(14)

13 default risk data in the years from 2008 to 2014. A good distribution has been found by taking the following selection: 483 ETPs over period January’13 – May’17 (N=483, T=231). More information on the data characteristics can be found in the data section.

Insufficient ETPs have been found in the dataset for the category ETC (compared to the number of selected ETNs and ETFs). The ETC class is smaller compared to the others and contained a high amount of missing data fields. Also many data fields were missing for other ETPs. Reasons for this are the date of listing, listing of the parent institution and market segmentation in terms of trading.

The highest market impact is expected to be in the ETN class, versus the lowest impact is expected to be in the ETF class, therefore the absence of ETC is not obstructive for this research. Three different datasets are constructed in order to analyse differences between the characteristics: 1) the ETF dataset, 2) the ETN dataset and 3) the combined or all dataset.

3.3 Choice of methodology

A similar model to the research from Hurling et al. (2016), who showed a significant impact of CDS on the fund outflow, will be adopted. Hurling applies a Probit estimation to determine significant differences in probabilities. Instead of using the random effects model estimating a Probit equation as Hurling did, we have used the Maximum Likelihood Effects to estimate a Logit model as prescribed by Brooks (2008). A Logit model should give very similar results to the Probit model as the densities are very similar. However, Brooks (2012) suggests to use a Logit model if the split between dependant variables is very unbalanced. As we are testing exceptional events, the data will be unbalanced and thus the Logit model is preferred.

The Logit model is a limited dependent variable model. A continuous dependant variable is transformed to a dichotomous variable to calculate the probability of an event occurring. By doing so, we try to avoid being unable to fit a possibly linear equation to a non-linear relation which could happen using OLS. Also it overcomes problems that linear probability models could have, giving a negative probability or a probability higher than 1 as outcome. Nollen et al. (2002) modelled the risk-factor

(15)

14 in the bid/ask spread as an option with stochastic time to expiration, thus non-linear, which goes beyond this research. Hypothetically a relation between default risk and the bid/ask spread only exists in a high-risk state and thus could be an additive to the complex factors of the bid/ask in the “normal” state. This research is interested in whether there is a significant increase in probability of an “high bid-ask”-state during a “high default”-state. Robustness checks show that it is likely that the process is not linear, therefore we have searched for an alternative model, the Logit model.

The Logit model estimating the dependent variable having a value of 1, is stated by the following equation.

𝑃(𝑌 = 1) = 𝐹(𝑧 ) = 1 1 + 𝑒 =

1

1 + 𝑒 ( , )

The function is the cumulative logistic distribution function. The goodness-of-fit can be derived from the McFadden’s R squared. The higher the R squared, the better the goodness-of-fit. The coefficients require a different interpretation from a linear regression model. A change in 𝑋 has a change on 𝑧 by the degree of coefficient 𝛽 (e.g. 𝑋 going from value 0 to 1, in- or decreases 𝑧 by 𝛽 ). Since the model follows a S-shaped distribution, a change in one of the variables cannot be linearly interpreted (e.g. 𝑋 going up by 2 does not imply that 𝑃(𝑌 = 1) increases by 2 ∗ 𝛽 % ). It is common that the mean or median values are inputted for the other variables to compare the z-values and corresponding probabilities by changing only one of the variables. This allows us to interpret the impact of a change in one single variable. Also a linear model is estimated to determine the degree of impact based on the levels or first differences. Considering the panel nature of the data set, two obvious statistical models are the fixed effect model and the random effects model. The Hausman-test can be performed to check which model is suitable. The linear model is estimated as per below:

(16)

15

3.4 Hypothesis

Two hypothesis are tested, one for the panel model and one for the logit model. Both models are applied on three datasets: ETF, ETN and a combined dataset.

Panel model:

The following hypothesis will be tested in threefold, one for ETF, one for ETN and one for the aggregated dataset. Also, it will be tested whether a difference in ETP type is significant.

𝐵𝑖𝑑/𝐴𝑠𝑘, = 𝛽 𝐶𝐷𝑆, + 𝛽 𝐶𝐷𝑆, ∗ 𝐸𝑇𝑁 + 𝛽 𝐸𝑇𝑁 + 𝛽 𝐶𝐷𝑆, ∗ 𝑆𝑤𝑎𝑝 + 𝛽 𝑆𝑤𝑎𝑝 + 𝛽 𝐶𝐷𝑆,

∗ 𝐹𝑢𝑛𝑑𝑒𝑑 + 𝛽 𝐹𝑢𝑛𝑑𝑒𝑑 + 𝛽 𝐶𝐷𝑆, ∗ 𝑆𝑒𝑐𝐿𝑒𝑛𝑑 + 𝛽 𝑆𝑒𝑐𝐿𝑒𝑛𝑑 + 𝛽 𝜎 , , + 𝛼

+ 𝜀,

Where 𝐵𝑖𝑑/𝐴𝑠𝑘, is the level of bid/ask spread, 𝐸𝑇𝑁 is 1 if the ETP is an ETN and zero

otherwise, 𝑆𝑤𝑎𝑝 is 1 if swaps are used (synthetic replication) or zero for physical replication, 𝐶𝐷𝑆, is the level of CDS, 𝐹𝑢𝑛𝑑𝑒𝑑 is funded (1) or unfunded (0), 𝑆𝑒𝑐𝐿𝑒𝑛𝑑 is

securities lending yes (1) or no (0), 𝜎 is the historical volatility of the ETP, the FX rate and the price of the volatility index (VIXX). For each dataset (all, ETFs and ETNs) only the relevant dummy variables are included to avoid the dummy variable trap. For example, for the all-dataset, only 𝐸𝑇𝑁 is included as an interactive term and as a dummy variable. For the ETF dataset, 𝑆𝑤𝑎𝑝 , 𝐹𝑢𝑛𝑑𝑒𝑑 and 𝑆𝑒𝑐𝐿𝑒𝑛𝑑 are included as interactive terms and dummy variables.

Logit model:

The Logit model has been estimated as per the following equation.

𝑃(𝐵𝑖𝑑𝐴𝑠𝑘 = 1) = 1

1 + 𝑒 ( ∗ , )

Where 𝐵𝑖𝑑𝐴𝑠𝑘 takes the value of 1 if the bid/ask spread is above a certain percentile from its distribution for that specific ETP, 𝐶𝐷𝑆 takes the value 1 if the CDS is above a certain percentile from its distribution for that specific ETP and 𝐸𝑇𝑁 takes the value of 1 if the product is an ETN and zero if it is an ETF. 𝐸𝑇𝑁 interacts with 𝐶𝐷𝑆 such that it can be tested if the impact of an increased default risk is higher for ETN products. The below hypothesis is tested for each of the variables. For example, the hypothesis tests whether there is a significant decreased or increased probability on an

(17)

16 observation that belongs to the higher distribution of the bid/ask spread if that observation does or does not contain an “high default risk”-state (i.e. a conditional probability). In other words, is there a change in probability of a high bid/ask spread if a condition is added to the probability.

𝐻 : 𝛽 = 0 𝑣𝑒𝑟𝑠𝑢𝑠 𝐻 : 𝛽 ≠ 0

The null-hypothesis would imply that the probability doesn’t change significantly when a condition is added. For example, if one out of ten observations is considered to be a high bid/ask spread, the probability of observing a high bid/ask doesn’t change on the condition of being an ETN, it remains to be 10% in this example. If the null-hypothesis is rejected, the probability of observing a high bid/ask changes significantly on the condition of being an ETN.

In summary, the theoretical framework prescribes that, given the collateralisation of the ETF, the CDS should not be of significance given the recovery rate of 1. In case of an ETN, the CDS is expected to significant given that an ETN is a unsecured debt instrument and thus has a recovery rate of less than 1. Thus the significance of any of the dummy variables in interaction terms would imply that the recovery rate for that type of ETP is less than 1 and there is counterparty risk for the ETP holder.

(18)

17

4. DATA

This chapter provides a brief summary on the dataset and its most important variables. Some practical issues were encountered when the dataset has been constructed (as highlighted in the choice of sample). The best possible dataset at disposal has been used. That the impact of a non-perfect sample distribution can be accepted is explained below.

ETPs:

A total number of 483 ETPs have been included in the sample (see table 5 in the appendix) which can be broken down into 37 ETNs and 447 ETFs. The data is derived from Bloomberg. The complete universe of 8,522 ETPs has been used as a starting point after which a selection was made based on checks of presence of all variables. From the selected ETPs there are 15 that are part of the top 20 ETPs based on assets under management (table 9 in the appendix) which indicates that proper liquid ETPs are part of the selection. The reason more ETFs have been selected is that the ETF asset class can be broken down into more sub-classes due to its different characteristics. The ETF class consists of 381 physical replicated (of which 277 apply securities lending versus 104 with no lending) and 65 swap based ETFs (9 unfunded versus 56 funded). 37 ETNs are included in the sample. Bloomberg did provide similar characteristics to the ETNs as for the ETF, this has been ignored due to the relatively small sample size of the ETN. In addition, considering that the ETN is a debt instrument, it can be reasoned that the replication type isn’t relevant as in case of a default, the creditor doesn’t have a first pledge on the “physical” assets held back by the issuer (if any).

The lower degree of included ETPs that are swap based and ETNs can be explained by the figures shown by Morningstar (2016), the physical replication structure overtook the swap based replication from 45% physical ETFs in 2011 to 80% in 2016. For example, ETF Strategy (2012) reports that Deutsche Bank and Credit Suisse converted respectively 18 and 7 swap based ETFs to physical replication structures in 2012 and 2013. We have performed checks on the db-X trackers that switched regime and they all changed their ticker code after the switch. This implies that a possible change in replication strategy does not impact the accuracy of the

(19)

18 dataset. An assessment of what the impact is of a change in replication strategy within an ETP, is interesting for further research.

The ETN growth stalled compared to the ETF growth (graph 4). In addition, the daily exchange traded volume of the ETF on average exceeds the volume of the ETN (EUR 44mm versus EUR 6.4mm). We consider the dataset relatively representative to the ETP population for the use of this research.

Issuers:

ETPs from 37 different parent institutions are part of the sample. The geographical location of the issuer’s ultimate parent is shown in table 6. The ultimate parent has been derived per ETP and is relevant as it has been assumed in the theoretical framework that most swap and securities lending business is allocated to the parent institution. The number of ETPs is concentrated in US listed institutions with 294 ETPs versus 134 European, 42 Asian and 13 other. This is at an acceptable representable level of the population which has USD 2.911B AUM in the US, 673B in Europe and 353B in Asia. (Blackrock, 2016)

CDS:

The cumulated CDS level has been plotted in graph 2 where we see that there is variance in CDS level over time (maximum of 700,000 points to minimum of 300,000 points). This shows it is likely that the selected issuers and timeframe contain sufficient variance to conduct the research.

At first review of the graph, there might be a significant correlation between the overall CDS level and the VIXX which indicates that multicollinearity issues could arise. We will elaborate on multicollinearity issues in the Robustness chapter.

(20)

19 Graph 2: Plot of the cumulated CDS level of the sample and the VIXX

0.00 5.00 10.00 15.00 20.00 25.00 0 100,000 200,000 300,000 400,000 500,000 600,000 700,000 800,000 01 /0 1/ 20 13 01 /0 4/ 20 13 01 /0 7/ 20 13 01 /1 0/ 20 13 01 /0 1/ 20 14 01 /0 4/ 20 14 01 /0 7/ 20 14 01 /1 0/ 20 14 01 /0 1/ 20 15 01 /0 4/ 20 15 01 /0 7/ 20 15 01 /1 0/ 20 15 01 /0 1/ 20 16 01 /0 4/ 20 16 01 /0 7/ 20 16 01 /1 0/ 20 16 01 /0 1/ 20 17 01 /0 4/ 20 17

Cumulated CDS levels

CDS VIX

(21)

20 Bid/ask spread:

The dataset has been analysed using the maximum bid/ask spread (expressed as % of the NAV) per ETP. Approximately 50 observations higher than 200% have been corrected to previous day’s value (in order to keep the panel data in balance). A bid/ask spread of 200% would imply that the bid is at zero and/or the ask is more than twice as large as the mid-price which is unrealistic.

A distribution of the dataset is displayed in graph 3. Remaining possible outliers are deemed to be sufficiently small. There are 619 observations with a bid/ask exceeding the 1%. The impact of possible outliers is smaller when we use the distribution of the bid/ask spread which is an additional plus of using the logit model.

Graph 3: Distribution of bid/ask spread (expressed as % of the NAV). 619 observations exceed the 1%. 0 2000 4000 6000 8000 10000 12000 0. 00 % 0. 03 % 0. 06 % 0. 09 % 0. 12 % 0. 15 % 0. 18 % 0. 21 % 0. 24 % 0. 27 % 0. 30 % 0. 33 % 0. 36 % 0. 39 % 0. 42 % 0. 45 % 0. 48 % 0. 51 % 0. 54 % 0. 57 % 0. 60 % 0. 63 % 0. 66 % 0. 69 % 0. 72 % 0. 75 % 0. 78 % 0. 81 % 0. 84 % 0. 87 % 0. 90 % 0. 93 % 0. 96 % 0. 99 %

Distribution of bid/ask spread

(22)

21

5. RESULTS

This chapter provides an overview of the regression and corresponding conclusions that can be made based on the results. First the results from the panel model are discussed after which the Logit model is discussed.

5.2 Panel model

The panel model has been estimated by taking the first differences of the variables. When applying the Augmented Dickey Fuller test on the CDS levels, it is shown that the variable has a unit root. Therefore the estimated model on levels can lead to a spurious regression and will not be used to derive further conclusions from. First differences are created to make all the variables stationary (Augmented Dickey Fuller null hypothesis is rejected for all first-differences variables). Natural logarithms are taken since all levels are greater than zero. (Brooks 2012)

The Hausman test has been performed on the model to test whether the random effects or the fixed effects model is appropriate. When estimating the models using the first differences of the variables, the Hausman test provides a p-value of 0.9818 which indicates that the Random Effects model is appropriate, the null hypothesis is not rejected.

During the research it became apparent that the process for the bid/ask is non-linear, hence the need for an alternative model which was found in the logit model. This conclusion is mostly drawn on the finding of an increase in significance when squared levels are added to the model (table 7 in the appendix). A more detailed analyses on how this conclusion was drawn can be found in the robustness section. Conclusions that can be drawn from the Panel model are listed below. The regression is displayed in table 8.

The variables used in the model are the logs from: CDS is the 5-years term expressed in percentage points, the Net Asset Value (NAV) and average daily volume (Volume), the VIXX closing prices are taken as an indication of overall market

(23)

22 volatility, the FX volatility has been calculated based on the daily closing prices and the ETP (simple) volatility has been calculated based on the weekly closing prices. Significance is found for lnCDS, lnETP-volatility, lnVolume, (all 1% confidence level), lnFX-volatility (5%) and lnVIXX (10%). The signs of the coefficients are as expected. The CDS, VIXX, ETP- and FX-volatility are positively correlated with the bid/ask. This can be interpreted that market makers require a higher premium when there is more uncertainty in the market. Volume is negatively correlated with the bid/ask which is as expected, a smaller bid/ask can be expected when there is more liquidity. No significance is found for the ETN dummy variables in interaction terms with the CDS. This implies that no significant difference is found in the estimated linear model between the different ETP types.

A R-squared of 0.21% has been found which is very low. Although multiple variables (volatilities and volume) have been added to the model based on literature research trying to achieve an larger R squared, the gross of the variance is still left to be explained by other variables. Also no significant difference between ETNs and ETFs is found while we expect that there are differences based on the literature research. The panel model contains multiple concerns which led to the adoption of an alternative model, the Logit model. The panel model shows to contain a very low R squared value. In addition, the volatility variables and intercept became significant when their respective squared values were added to the model. These findings are discussed in more detail in the Robustness section.

(24)

23

5.3 Logit model

The following transformations have been made in order to apply the Logit model: the bid/ask spread (calculated as percentage from the NAV) takes the value of 1 if the bid/ask is larger than the 80th_{percentile of that ETP’s distribution and zero otherwise,}

the CDS takes the value of 1 if the 5-years CDS level is larger than the 80th_percentile

of that ETP’s distribution and zero otherwise. The variables ETN, Swap, Funded, Sec Lending are binary variables with the value of 1 if that characteristic is applicable for the ETP and zero otherwise.

ETN*CDS is an interaction term where we expect that the impact of the CDS is only applicable for ETNs (the control group), thus the independent CDS variable is not significant in the null-hypothesis. For the other variables: swap, funded and securities lending, it is tested whether these provide any significant change in the probabilities. All interactions terms are expected to have positive coefficients. ETN has been included as a single variable to check for correctness of the model and is expected not to be significant. The dummy variables have been included only as interactive terms in the ETF datasets. The regression result is provided in table 2.

The logit model estimates the probability of the dependent variable having a value of 1, i.e. an increased level of bid/ask spread (larger than the 80th_{percentile of the}

distribution) called “high bid/ask” hereafter. The 80th_{percentile has been assessed as}

a proper level. Hurling et al. (2016) used the 70th_{percentile which has been added as}

a robustness check in table 8. We have taken a slightly higher percentile in order to visually show more clearly the changes in probabilities. Similar results are found at different percentile levels. If no relation between the variables would exist, the probability of observing an “high bid/ask” on the condition of an “high CDS” is slightly smaller than 20% (due to the 80th_{percentile), 𝑃(𝐵𝑖𝑑𝐴𝑠𝑘 = 1|𝐶𝐷𝑆 = 1) ~ 20%.}

The probability on a high bid/ask would be 20% for all other possible combinations as well if no relationship would exist. Hence this probability equation can be supplemented with the other variables. The model has been estimated on three datasets to allow for dummy variable testing: All, ETF and ETN.

(25)

24

This table displays the regression results of the Logit model. Weekly data is used from the period Jan’13 – May’17. CDS is from the ETP’s Issuer’s Ultimate Parent Institution. CDS takes the value of 1 if the observation is larger than the 80th_{percentile of the ETP’s distribution}

and zero otherwise. ETN, SWAP and Sec Lending take the value of 1 if that characteristic is applicable for that observation and zero otherwise.

The dependent variable is the Bid/Ask which takes the value of 1 if the observation is larger than the 80th_{percentile of the ETP’s distribution and zero otherwise. The Model has}

been estimated using Maximum Likelihood Effects.

The Z-statistic is displayed in brackets. Significance levels are stated as: *** 1%, ** 5% and * 10%.

Dataset ALL ETF ETN 70th Perc.

Observations 111,573 103,026 8,547 111,573 Coefficient Intercept -1.5876 -1.5876 -1.6263 -1.1084 (-171.97) *** (-171.97) *** (-49.91) *** (-129.25) *** CDS 0.8466 0.4767 0.9650 0.7889 (47.57) *** (12.83) *** (15.87) *** (54.48) *** CDS*ETN 0.1184 0.1237 (1.87) * (2.36) ** ETN -0.0388 -0.0392 (-1.14) (-1.25) CDS*SWAP 0.3007 (5.86) *** CDS*FUNDED -0.0558 (-1.62) CDS*SEC LENDING 0.5549 (14.26) *** McFadden R sq. 0.0214 0.0229 0.0280 0.0240

Table 2: Logit model

For the All-dataset we find that the CDS is significant at a 1% confidence level and the ETP type (ETN variable) is a significant factor at a 10% confidence level. The result can be interpreted as per the following. For example, the probability of observing a high bid/ask on the condition of an ETN with high CDS, increases by 15.48% (in absolute terms). The probability decreases by 2.40% on the condition of an ETF and no high CDS is observed. This is calculated by inserting the sample mean values in the following formula to arrive at the base probability (unconditional probability) and changing the variable values based on the scenario being tested to arrive at the conditional probability. For example for an ETN with high CDS is calculated as per: 𝑃(𝐵𝑖𝑑𝐴𝑠𝑘 = 1| 𝐶𝐷𝑆 = 1, 𝐸𝑇𝑁 = 1) = ₍ _. _. _∗ _. _∗ _. _{∗ )}= 34.85%

(26)

25 We can reject the null hypothesis for ETP type when using a 5% confidence level as the alternative hypothesis has a beta higher than zero (one-sided test), i.e. there is a significant higher probability on a high bid/ask for ETNs versus ETFs on the condition of an high CDS state. The literature review showed us that the ETN is a unsecured debt product versus the ETF being a collateralized fund. This result supports that the ETN is deemed to be more risky during times of financial distress. Given the 10% significance level for a two-sided test, there are indications that this finding might not be very representative for the whole population, which would be feed for further research. For the ETF dataset, the model has been estimated using the CDS and interactive terms CDS*swap, CDS*funded and CDS*securities lending. Similar as for the ETN dataset, the probability increases significantly at a 1% confidence level that an high bid/ask is observed under the condition of an high CDS. Although this is expected based on the literature review (for example due to a collateral mismatch), it is not what is advertised by the industry and what regulators strive for. The significance of the CDS variable implies that it is also significant for the ETFs sample as a whole, i.e. also the ones that apply a physical replication scheme. Physical ETFs are deemed to be fully collateralized at all times by exactly the same assets as the benchmark it tracks. Theoretically there should not be an impact in the market. In case of a default, the ETF holder is entitled to the assets held in a segregated account. The significance of the CDS could indicate that the market maker is still running a counterparty risk on the issuer even though the fund is fully collateralized by the assets that it deems to track. A possible explanation is the technical workings of ETF. For example, an in-kind creation, the market maker delivers the basket to the issuer “free of payment” after which the issuer delivers the ETF to the market maker “free of payment” (ETF.com, 2017). During this settlement process, the market maker is exposed to counterparty risk. Another explanation is that the market maker would have its assets stuck in the fund in case of a default, leading to opportunity losses (Singh and Aitken, 2010).

Within the ETF dataset, the probability increases significantly at a 1% confidence level for the securities lending and swap interactive terms. There is a higher probability of observing a high bid/ask in a high CDS state when the ETF engages in securities lending or swaps than if it would not performing securities lending or swaps. This implies that performing swaps and/or securities lending provides additional risk for the market maker. A common characteristic of the swap and securities lending

(27)

26 transaction is that there is a transfer of assets between two parties which implies that other assets are in possession of the issuer than the original assets being used for tracking purposes. This thus provide rise for additional counterparty risk (Ramaswamy (2014), Hurling et al. (2016) amongst others). As discussed in the theoretical framework, the swap and securities lending transaction is likely to be performed with the parent institution. ETF regulation however, require that an ETF is collateralized. When the product would be fully collateralized, theoretically the CDS should not be of any impact as it would cause the recovery rate to be 1 (i.e. there is sufficient collateral to repay the ETF holder in case of a default of the swap or securities lending counterparty). The significance of securities lending and swap proofs that the recovery rate is less than 1. Of course the level of collateralization of the ETF depends on the market prices of its products in position. When the collateral basket moves out of sync with the NAV of the ETF, there is a possible shortfall for the ETF holder. Hurling et al (2014) found that the collateral basket matches the characteristics of the ETP, however our research tends to support the concerns from Ramaswamy (2011) and other international agencies that there is a possible mismatch in collateral. Possible explanations are: mismatch in expected value of the collateral and/or intransparent collateral holdings (Ramaswamy, 2011). Intransparent collateral holdings make it difficult for the market maker to unwind its position in case of a default as he is uncertain about his actual holdings. Other possible explanations are topics for further research.

The funded interactive term is not significant which indicates that market makers do not request a risk premium in case of a high CDS state for a funded swap structure compared to the unfunded structure. A possible explanation is that it isn’t much different in terms of counterparty risk if the assets lay directly with the issuer or with a 3rd_{party custodian (see funded/unfunded structure in theoretical framework). This is}

an interesting finding as the funded/unfunded structure is being touched upon in much detail in previous research.

As a check, the model has been estimated for the ETN dataset with CDS only. The CDS is found to be significant with a 1% confidence level. The probability of an high bid/ask increases on the condition of an high CDS. This result is as expected, the risk for the ETN holder increases with the level of CDS as the probability increases for the ETN holder that he will not be able to redeem his position or will be fully repaid by the issuer in case of a default. The ETN being an unsecured debt instrument, the price of

(28)

27 the note should be related to the credit worthiness of the issuer and/or its parent institution.

The model is also estimated using Volume, the VIXX, FX and NAV volatility which shows similar signs and significance levels as for the panel model. To enable intuitive interpretation of the results, these have been left out in the different scenarios as they are not relevant for this research. The absolute change in probability for a set of scenarios have been calculated in table 3. The largest increase in probability in case of a high CDS is on the condition of an swap based ETF with securities lending. The probability increases in this scenario from 19.35% to 43.51% (+24.16%). The lowest probability increase in case of a high CDS is for the physical ETF without securities lending. In this instance the probability on a high bid/ask increases only from 19.35% to 24.66% (+5.31%). This implies that a physical ETF without securities lending is least prone to counterparty risk. The probability on a high bid/ask if none of the scenarios apply (i.e. no high CDS, no ETN, swap or securities lending) decreases by 2.55% on average.

This table displays the absolute changes in probabilities estimated by the Logit model for a selected set of conditional probability scenarios. The unconditional probability on a high bid/ask spread is slightly smaller than 20%.

Dataset Scenario Change in probability

ETP: CDS: Y, ETN: Y 15.48%

CDS: Y, ETN: N 12.91%

CDS: N, ETN: N -2.40%

ETF: CDS: Y, SWAP: Y, SEC LEND: Y 24.16%

CDS: Y, SWAP: N, SEC LEND: Y 16.96%

CDS: Y, SWAP: Y, SEC LEND: N 11.31%

CDS: Y, SWAP: N, SEC LEND: N 5.31%

CDS: N, SWAP: N, SEC LEND: N -2.46%

ETN: CDS: Y 14.82%

CDS: N -2.79%

(29)

28

6. ROBUSTNESS CHECKS

Multiple checks are performed to test the model on robustness. First a preliminary research has been performed on a set of db-X trackers. This led to the addition of the variable Volume to improve the explanatory power of the model. Small additional checks, such as testing of significance of lagged values like CDS t-1 and VIXX t-1, adding and removing variables, etc. are not discussed in detail as no interesting outcome is found. The most interesting checks on robustness are displayed below: Functional form:

First the model has been estimated using the levels of the variables. The regression result is displayed in table 7. The level model has a relatively high R-squared value of 27.90%. No significance has been found in the variables of interest. As expected, the bid/ask is highly impacted by the level of NAV (at a 1% confidence level). It is obvious that an higher bid/ask spread is observed simultaneously with higher trading price to account for the cost of transacting per EUR (Bollen et al, 2004). The market makers requests higher risk premium as risk is worn of more capital (a % on every EUR invested).

A common methodology for testing the functional form of the model is by taking the squared value of all independent variables in levels (CDS, NAV, VIXX, volume, ETP- and FX-volatility). Significance of one of the squared values could indicate that the process is non-linear and the functional form should be revised. Even though the squared CDS level is not significant, FX volatility and the VIXX did become significant when the squared values are added. Also the intercept is now significant and the adjusted R squared increased from 27.90% to 28.85%. It is safe to say that, supporting the initial suspicions on non-linearity due to other papers on the bid/ask, there is a need of functional form redesign of the OLS when testing for significance of the CDS on the bid/ask. The Logit model was found to be a proper replacement for the linear model.

Multicollinearity:

Brooks (2012) mentions multiple signals that could indicate (near) multicollinearity. Our McFadden R squared value is relatively small (2% - 4% depending on the model and dataset, values between 20% - 40% are deemed to be an excellent fit) and

(30)

29 there is no significant impact on the R squared when we remove or add variables. This is an indication that there is no (near) multicollinearity.

The most significant correlation between variables (table 4) is observed between the CDS and the VIXX with a rho value of 0.2423. A certain degree of correlation was expected beforehand due to the relationship between average CDS levels, the VIXX (a measure of volatility of the market) and an higher degree of uncertainty in the market during times of stress. For example, a correlation could be expected between financial institutions during the Euro and subsequent crises which started in 2009.

We deem the degree of correlation to leave sufficient room for the “idiosyncratic” movement of the financial institution’s specific CDS and its impact on the bid/ask. Considering the different significance levels of the CDS and VIXX variables and a decreased value of adjusted R squared values when removing the CDS, we deem multicollinearity not an major issue although this will be touched upon in the chapter for further research.

The second highest correlation is between the VIXX and the NAV (value of -0.1281). This is as expected as it is common financial theory that stock prices are negatively correlated with the implied volatility. Similar to with the VIXX and CDS, the degree of correlation is deemed to be sufficiently low for this research.

CDS NAV ETP volatility Volume VIXX FX volatility

CDS 1.0000 - - - - - NAV -0.0757 1.0000 - - - - NAV volatility 0.0513 -0.0303 1.0000 - - - Volume 0.0248 -0.0601 0.0349 1.0000 - - VIXX 0.2423 -0.1281 0.0727 0.0779 1.0000 - FX volatility 0.0449 -0.0124 0.0189 0.0063 0.0487 1.0000

(31)

30 Alternative datasets:

As discussed in the results section, the model has been tested on a total of three datasets: All, ETN and ETF. All three datasets provide very similar results that are in line with expectations. That the model has been tested on different parts of the data with similar outcome, strengthens our feeling that the test results will be replicated when a completely different set of ETPs is used. Thus the dataset is a good representation of the population.

Size of the “on-exchange” market:

An important part of the liquidity in ETPs is in the Over-The-Counter market (OTC). The difficulty with the OTC market is that figures aren’t easily accessible. Vanguard (2015) states that the on-exchange traded volume for ETFs, in contrast to single stocks, does not provide the full picture of its liquidity due to its creation and redemption workings and OTC market. When we compare the fund outflow with the trade volume, we see that the fund outflow can exceed the daily traded volume which indicates that OTC trading took place. FT (2016) estimates the ratio of on-exchange vs OTC to be 30:70 in Europe and 70:30 in the U.S.. This is partially due to reg. NMS in the U.S. and market segmentation in Europe. What should be kept in mind is that conclusions from this research might be biased due to the absence of OTC data. Using on-exchange data however is the “best of the rest” solutions and the bid/ask is the actual price that could have been traded on.

(32)

31

7. FOR FURTHER RESEARCH

This chapter discusses different aspects of this research that are of interest for further research and/or should be further investigated before full conclusions can be made. Alternative dataset and model:

The estimated linear model did not show a good fit and the Logit model on the 80th

percentile dummy showed a better but still not optimal fit. In order to predict the impact of the credit worthiness on the market, a different (non-linear) model must be adopted. This model should also factor in any potential multicollinearity between the idiosyncratic risk of the issuer and the overall market. An idea at hindsight is to compare different existing models that explain the bid/ask and factor in credit worthiness of the issuer.

In order to research additional characteristics or factors, a different dataset must be used, also to allow for including ETCs. The significance of the CDS on ETNs was only proven at 5% confidence level (one-sided) which was relatively low compared to the other factors. Another research might proof differently although the theory supports our findings.

Potential significant factors:

During this research it became apparent that there are more potentially significant factors in the ETP. The significance of the CDS on ETFs using a physical replication scheme imply that there is a potential loss given default while theoretically the ETF is fully collateralized by the exact product it deems to track. Hill et al. (2015) provide a good and complete source of information. To highlight one of the potential characteristics: the transaction from the point of the market maker is different for in-kind creations/redemptions than for cash creations/redemptions. Also the default procedure could give rise to potential losses and the regulatory framework in which the issuer operates might provide insights in this matter. The countries in which most ETFs are issues are: United States, Germany, Ireland and Luxembourg. (Ramaswamy, 2012)

(33)

32 Time horizon:

It will be interesting to perform an event study on a selection of ETPs whose issuer’s credit worthiness stood under a high amount of stress and look into the intra-day data. An possible scenario is that institutional holders of ETPs (e.g. pension funds) are allowed to only hold ETPs from issuers with a certain degree of credit worthiness (e.g. CDS below 400bps or only >A-rating). A research can be performed on the best practices within the market on these limits and then select situations in which these limits are crossed. Subsequently intra-day data can be assessed as the impact could potentially only appear during a small timeframe (due to creation/redemption mechanics) after which the normal liquid markets return. In these instances, a more accurate economic meaning could be addressed to the cost of liquidation during times of stress.

Level of collateralization:

In this research the level and composition of the collateral of the ETP (if applicable) has been ignored. The collateral of the ETP might explain the significance of the CDS from our research. Hurling et al. (2016) found that the collateral composition is not optimal. Bartolini et al. (2011) found that the assets pledged in the securities lending market are mainly assets with low collateral value. Brandt and Kavajecz (2004) found that assets are pledged for which the demand is low in the secondary market. These papers make it apparent that the collateral composition is not optimal. Also it is plausible that issuers or regulators apply different standards in these aspects. The market impact can thus be better modelled if these factors are taken into account. Also it might show useful for regulators to know which methods are more or less susceptible for counterparty risk.

One of the implications of an ETP sell-off according to Ramaswamy (2012) is that it will cause a sell-off in the “illiquid” assets pledged initially by the swap counterpart at the issuer. Once the collateral of the ETP has been identified and a sell-off event in the ETP occurs, it can be investigated whether there is a volatility spill-over to the collateral assets.

(34)

33

8. CONCLUSION

This paper focusses on the relationship between counterparty risk on ETPs (Exchange Traded Products) and the credit worthiness of the issuer or parent company. The differences in impact between ETP characteristics is investigated as it would provide useful insights for further research. The key characteristics that are identified are: ETP type (ETF versus ETN; ETC is excluded), replication type (swap versus physical; funded versus unfunded) and securities lending. These factors have been subject of previous research and are under the attention of regulators. A set of weekly data has been used of 483 ETPs over the period Januari’13 – May’17.

First a linear model has been estimated using Random effects which has shown to be inaccurate. Non-linearity of the process has been suspected due to significance of the squared variables. An alternative model, Logit model, is selected as proper alternative. This was also applied by Hurling et al. (2016) who performed a similar research on fund outflows. The Logit model is applied such that it can be estimated what the probability is of observing a “high bid/ask spread”. The main explanatory variable of interest is the CDS of the issuer’s parent institution. The ETP characteristics and CDS are interaction terms. To enable interpretation of the results, the 80th

percentile has been determined as “high” for the ETP’s bid/ask distribution and for the issuer’s parent institution’s distribution of the CDS. Three datasets have been used to allow for dummy variable testing (ETF, ETN and a combined (all)). The model has been controlled for by including variables volume and market-, FX and ETP-volatility. The significance of the CDS and it’s interaction terms proof that the loss given default is unequal to zero. A high bid/ask implies for the ETP holder that there is more cost of liquidating the position. Potentially it could even imply that there is no market at all and the ETP holder remains exposed to not only the fund exposure but also to the counterparty risk.

Our findings support the literature review which shows that the ETN, being an unsecured debt product, versus the ETF, being a collateralized fund, is impacted to a larger degree by credit worthiness. The probability of observing a high bid/ask increases significantly when a high CDS is observed on the condition of an ETN product. The ETN can be seen as the control group as it was expected that the CDS would only be of significance in combination with the ETN, but it was not expected

(35)

34 that the CDS was significant as an independent term as this implies that the CDS is significant for ETFs as well, i.e. ETFs are prone to counterparty risk. The significance of the CDS is also shown in the ETF dataset and thus applies to ETFs that apply physical replication schemes. No concerns were found in the literature review for physical ETFs and therefore this result comes as a surprise. The significance of the CDS for physical ETFs could indicate that the market maker is still running a counterparty risk on the issuer even though the fund is fully collateralized by the assets that it deems to track. A possible explanation is the technical workings of ETF, settlement risk or lays within the default procedure.

Significance is found for Swap based ETFs and ETFs that perform Securities Lending. One common characteristic of the swap and securities lending transaction is that there is a transfer of assets between two parties which implies that other assets are in possession of the issuer than the original assets being used for tracking purposes. This potential mismatch thus provides additional counterparty risk which can be factored in in the high bid/ask. The theoretical framework did provide conjectures that the swap and securities lending provide additional counterparty risk but, to our knowledge, before this research the market impact has not been proven yet. Our research tends to support the concerns from Ramaswamy (2011) and other international agencies that there is a possible mismatch in collateral which can explain the significance of the CDS.

The funded interactive term is found not to be of significance which indicates that there is no difference in probability of observing a high bid/ask in case of a high CDS state for a funded swap structure compared to the unfunded structure. A possible explanation is that it isn’t much different in terms of counterparty risk if the assets lay directly with the issuer or with a 3rd_{party custodian. This is an interesting finding as the}

funded/unfunded structure is being touched upon in much detail in previous research.

Initially the most important limitation of this paper is the fit of the functional form. This has been worked around by the adoption of the Logit model. By using the Logit model however, the degree of change in the bid/ask itself cannot be estimated. It is only capable of estimating the probability of a “high bid/ask” occurring. This also implies that less economic value can be addressed to the term increased counterparty risk. In order for an ETP holder to accurately estimate the increased cost

The counterparty risk of ETPS : is counterparty risk priced in the market?

Master Thesis

2017