The visible hand: benchmarks, regulation, and liquidity

According to recent theoretical work, a more transparent and precise benchmark assessment should positively impact liquidity in the underlying market. We exploit a benchmark regime change in the $289 trillion interest rate swaps market to test this prediction. Utilizing proprietary electronic order book data, we find improved liquidity effects in the USD swaps market following the transition to the regulated ICE Swap Rate. Regulations that improve the methodology and oversight of benchmarks can, therefore, impact markets positively. Conservative estimates of direct savings in a single swap tenor on one trading platform are in the region of $4 million $7 million.


Introduction
Benchmarks are critical to the efficient functioning of markets.Many industries, but particularly the financial services industry, use benchmarks to settle contracts, monitor trade execution, and signal sentiment in the market.
Until recently, benchmarks were not subject to any regulatory supervision.However, following a number of scandals in the late 2000s and early 2010s that exposed several instances of misconduct, regulatory authorities around the world have spent the best part of a decade introducing rules to prevent the recurrence of such misconduct.In the United Kingdom, the Financial Conduct Authority (FCA) started regulating the London Inter-bank Offered Rate (LIBOR) in 2013, then arguably the most important benchmark in the world, and seven more benchmarks followed in 2015. 1 International bodies such as the International Organization of Securities Commissions (IOSCO) and the Financial Stability Board (FSB) produced a series of reports with recommendations on how to regulate benchmarks. 2 The European Union introduced extensive regulation of benchmarks in early 2018.The global regulatory efforts in the benchmark space have therefore been particularly intense.
In this paper, we present empirical evidence that supports the efforts made by regulatory authorities in recent years.Enhancements to transparency and precision with which benchmarks are measured can have economically significant market outcomes, as demonstrated by our empirical analysis of the regulatory intervention in one of the most important benchmarks in financial markets.
We build on Duffie et al. (2017), who show that the introduction of a benchmark improves the trade matching process in opaque over-the-counter (OTC) markets, enhancing social welfare as it improves the information available to traders and reduces their search costs leading to increased price transparency. 3For this reason, a benchmark that encourages dealers to compete aggressively for the best price prompts more efficient dealer-trader matching and increases the volume of beneficial transactions.Increased inter-dealer competition improves market liquidity and reduces transaction costs.Specifically, "the most efficient dealers can use a benchmark as a 'price transparency weapon' that drives inefficient competitors out of the market" (Duffie et al., 2017, p. 3).However, the model setup in Duffie et al. (2017) only considers a market that shifts from a state without a benchmark to one with a benchmark.Aquilina  and Pirrone (2020) develop a stylized model to examine the economic effects of an increase in the "quality" of an unregulated benchmark.This speaks directly to the empirical setup of this paper as improvements in benchmark quality are a likely result of the regulations mentioned above.Appropriate regulatory intervention reduces the noise introduced by imperfections in the assessment process through an increase in the precision of the benchmark fixing; for instance, by establishing stricter assessment instructions, controls, and oversight.Benchmarks send a more precise signal to the market, thus leading to better market outcomes.As a consequence, Aquilina and Pirrone (2020) expect that the quality of the underlying market improves as prices become less noisy and market participation increases.The result will be higher liquidity.
We test the theoretical prediction that a well-designed benchmark regime change (BRC) will have positive effects on the liquidity of the underlying market, using a natural experiment set out by the FCA in 2015.Specifically, we exploit the March 31, 2015 transition from the unregulated panel-based ISDAFIX benchmark to the regulated market-based ICE Swap Rate-a fundamental transformation of the benchmark, which is central to the $289 trillion swaps market and used, for example, in hedging interest rate risk.The BRC, induced by the FCA, introduced controls and regulatory oversight, as well as a new assessment methodology-the transparency gains of which should reproduce effects similar to the efficiency improvements in Duffie et al. (2017) and the reduction of noise in Aquilina  and Pirrone (2020).
We study proprietary order book data and show that market liquidity improves following the BRC, as measured by quoted spreads, depth, and execution costs.Spreads narrow significantly-14%.Despite the fact that quoted depth at the best bid and offer decreases, the overall depth in the order book across 10 levels increases, and the execution costs of standard market size orders decrease.For the purpose of estimating transaction costs in markets with regular quote updates but few trades like the inter-dealer swaps market, we develop a simple measure termed "fill spread".As an aggregate measure of the combined effects on spreads and depth, the proxied roundtrip costs of completing a buy transaction and a sell transaction decrease by roughly 11% following the BRC.Difference-indifferences test results show that the significant increase in liquidity is more pronounced for benchmark-grade swaps, for which a regulated benchmark rate is assessed daily, compared to non-benchmark-grade swaps following the transition to the new benchmark regime.The positive effects on the liquidity of benchmark-grade swaps are over and above other influences, such as increases in venue participation by so-called "streamers." 4 We therefore confirm a causal link between the improvement in on-platform execution costs and the regulatory intervention of the FCA.Well-designed policy interventions can indeed be beneficial for the functioning of financial markets (Stiglitz, 1993; Barth et al., 2013).Our results are robust to controlling for a multitude of confounding effects, such as volatility and macroeconomic events, alternative definitions of our dependent variables, the inclusion of a wide set of swap tenors and alternative multivariate specifications.Moreover, we endogenously test for structural breaks in the time series of the liquidity measures employed and identify significant breaks that align with the BRC.
Furthermore, we find indicative evidence suggesting that the BRC has a positive effect on the representativeness and accuracy of the benchmark rate, measured as the differential between the proxied on-platform execution price and the benchmark rate.At the end of the assessment, and at the time of publication, the benchmark rates under the new regime are between 22% and 68% closer to market prices.Finally, we test the accuracy of the benchmark price as a predictor of the price prevailing in the market after the benchmark assessment.We compare the goodness-of-fit measures before and after the BRC and find an improvement following the regulatory intervention.
Our paper contributes to the research stream on financial benchmarks and their interactions with the underlying markets.Previous studies focus on the trading patterns of financial products around the assessment periods of short-term loans, precious metals, oil, and foreign exchange benchmarks.Abrantes- Metz et al. (2012) study the market dynamics around the setting of the benchmark for short-term interest rates, and find patterns suggestive of anticompetitive behavior in the 1-month LIBOR rate.Monticini and Thornton  (2013) analyze the conjecture that some panel participants have understated their LIBOR submissions and present evidence that this behavior has likely led to a reduction in the reported rate.Meanwhile, Fouquau and Spieser (2015) apply a novel technique that allows them to detect possible cartels.Their findings are underscored by the regulators' fining of banks for their involvement in the 2012 LIBOR manipulation scandal.Recent examinations of commodities markets have also indicated patterns of exploitation of benchmark processes.Caminschi and Heaney (2014) deduce that information leaks from the physical London PM Gold price fixing into the gold derivatives market ahead of the official price publication.Frino et al. (2017) report similar evidence of a consistent price trend in the Brent futures in the direction of the benchmark outcome during the Platts Dated Brent assessment.Finally, Osler and Turnbull (2019)  and Evans (2018) focus on foreign exchange and the WM/Reuters London 4 p.m. foreign exchange (FX) fix.While the former models dealer behavior around benchmark price assessments and derives trading patterns that suggest collusion among participating dealers, the latter finds currency price movements that align with collusive activities.
The literature stream on benchmark manipulation and price patterns around the times of assessments has led to a set of theoretical papers focusing on the design and reform of financial benchmarks and the benchmarks' value for financial markets (Duffie and Stein,  2015; Perkins and Mortby, 2015; Duffie et al., 2017; Eisl et al., 2017; Coulter et al., 2018; Duffie and Dworczak, 2018). 5For instance, in addition to Duffie et al. (2017), who describe the importance of benchmarks for financial markets, Duffie and Stein (2015) argue that robust benchmarks should be based on concluded transactions and not market participants' subjective judgments.The reformed ICE Swap Rate that is the focus of our analysis takes a step in the right direction, being computed from tradable and transparent electronic quotes.The authors also acknowledge the vital role of regulators in supporting effective transitions to better benchmarks.Furthermore, Duffie and Dworczak (2018) study the computation of transaction-based reference rates and make suggestions on the optimal design.Coulter et al. (2018) and Eisl et al. (2017) investigate different assessment procedures and make specific recommendations for the reform of LIBOR.
With this study, we make three key contributions to the growing literature on benchmarks.Firstly, we provide first-hand evidence in support of the many regulatory interventions and the beneficial impact that they can have on both benchmarks and markets.Second, we test the predictions of recent theoretical advancements in the benchmark space within an empirical framework.In addition to demonstrating the effects of transparent and regulated benchmarks on market quality, the proprietary full order book dataset, covering roughly 38% of the electronic inter-dealer USD interest rate swaps (IRS) market, allows us to directly analyze and document the microstructure of the world's largest derivatives market for the first time in the academic literature.Thirdly, we add to the debate on the impact of regulatory interventions on the efficient functioning of financial markets.
The remainder of this paper is organized as follows: In Section 2, we further corroborate the rationale and study design, and in Section 3 describe the institutional background, introduce the data, and provide descriptive statistics on the electronic trading of swaps.In Section 4, we discuss the main results, while additional robustness tests are in Section 5. We conclude in Section 6.

The positive effects of a more transparent and precise benchmark
As described above, Duffie et al. (2017) predict that the introduction of a benchmark improves market efficiency and liquidity.The authors, however, categorize an opaque OTC market with no central limit order book and thus no pre-trade price transparency.Our setting is slightly different.Several regulated trading venues offer transparent order book facilities and did so even before the BRC.Nevertheless, fragmentation in terms of multiple venues, trading mechanisms (electronic versus voice), as well as between dealers and their clients, means that the market is still somewhat opaque.Hence, it is reasonable to assume that only sophisticated market participants with simultaneous access to the trading venues have a comprehensive view of the market.As a result, the market is still dominated by large dealer banks.At the same time, many buy-side clients that trade through dealers do not have the required level of sophistication to navigate the fragmentation of the market themselves, justifying the continuing publication of a benchmark in this market.
In addition, the framework by Duffie et al. (2017) involves the evolution of a market without a benchmark into one with a benchmark.In our case, a benchmark has always existed; however, the old version was largely akin to the type envisioned in an OTC market by Duffie et al. (2017), where a panel of dealers set and publish a benchmark price in a procedure that is "separate" from the market.The reformed benchmark is market based, driving consilience and transparency by aggregating price and volume data from multiple inter-dealer platforms to calculate a market-wide benchmark price.
Our empirical setting still exhibits several commonalities with the theoretical model presented in Duffie et al. (2017).The old assessment was opaque, subjective, and unregulated.Hence, the transition to a new transparent, objective, and regulated benchmark corresponds to the beneficial effects of the benchmarks outlined in Duffie et al. (2017).The reformed benchmark's improved reflection of market fundamentals can help to correct the prevailing sub-optimal price transparency and dealer competition, and therefore strengthen liquidity and lower transaction costs in the underlying market.
In support of this reasoning, Aquilina and Pirrone (2020) propose a stylized model that illustrates the effects of an increase in the "quality" of a so far unregulated benchmark.The authors propose that such quality improvements could be achieved by an appropriate regulatory intervention, which encourages an increase in the precision of the benchmark by inducing a reduction in the pricing noise during the benchmark fixing process.As a consequence, the quality of the underlying market, in terms of liquidity, participation, and pricing noise, improves too.Increased benchmark precision is likely achieved via interventions that increase the transparency and robustness of the benchmark via methodological changes and regulatory oversight and controls, such as the FCA-induced transition from the unregulated panel-based ISDAFIX benchmark to the regulated market-based ICE Swap Rate.Hence, the benchmark regime change provides a suitable natural experiment to test the predictions by Duffie et al. (2017) and Aquilina and Pirrone (2020) on the liquidity of fixed-for-floating IRS.

The swap market
Fixed-for-floating IRS (henceforth also simply referred to as swaps) are predominately traded on regulated trading venues, where buyers and sellers meet to exchange cash flows based on a notional amount, with one party paying the fixed rate and receiving the floating rate and vice versa.Each payment series of a swap is defined as a fixed or floating leg.Given the prominence of USD IRS, with a notional amount outstanding of $139 trillion, 6 we focus on the USD segment only.The data used for the USD ICE Swap Rate benchmark assessment, which determines the fixed leg price, are sourced from the order books of participating swap execution facilities (SEFs). 7SEFs were introduced by the Dodd-Frank Wall Street Reform and Consumer Protection Act enacted in the U.S. in 2010 (the Dodd-Frank Act), which stipulates the mandatory trading of certain traditional OTC derivatives, such as swaps, on regulated venues to promote competition and enhance transparency.Recent statistics studying data between January 1, 2013 and September 15, 2014 indicate that more than two-thirds of USD fixed-for-floating IRS trading takes place on-SEF [see (Benos et al., 2020) for more details].As such, SEFs are electronic trading platforms that post and execute bids and offers to trade swaps from multiple participants.Under the mandatory trade execution requirement, swaps made available to trade (MAT) 8 are required to be traded on SEFs over the full length of our sample period.Table 1 provides a list of the USD IRS maturities captured by the MAT mandate.A benchmark rate is assessed for all tenors covered by the MAT requirement, except for the 12Y swap-a peculiarity that we use to our advantage in the difference-in-differences analysis in this study.
Rules further require that registered SEFs must, as a minimum, operate limit order books (LOBs) for all listed swaps.The platforms can also offer request for quote (RFQ) or voice-based functionalities in conjunction with the LOB, and therefore often run a hybrid model, pairing electronic and voice broking.

The regulation of benchmarks
Given the economic significance of the IRS market and its high degree of interconnectedness with the fixed income and money markets, the need for a reference price in the form of a standardized benchmark rate for valuing and settling contracts was recognized early on.The ICE Swap Rate, formerly known as the ISDAFIX rate, is of crucial importance to these markets because it is used in the valuation of, for example, early-terminated IRS, cash-settled swaptions, interest rate indexes, and many others.It is also often used by pension funds to hedge interest rate risk or in the U.S. Federal Reserve's statistical release. 9he International Swaps and Derivatives Association (ISDA) established the leading benchmark for fixed rates on swaps in 1998.The benchmark rates were assessed based on submissions made by a panel of 16 banks representing the mid-market rates at which they were willing to trade a standard market size (SMS) swap in the current market environment.The SMS differs across tenors and is $50 million for the 10-year (10Y) USD contract, which is the most actively traded tenor in our sample.For USD swaps, the panel submission polling window ran from 11:00:00 to 11:15:00 ET and the ISDAFIX rates were published at 11:30:00 ET (see Panel A of Fig. 1). 10 In 6 See statistics produced by the Bank of International Settlement (http://stats.bis.org/statx/srs/table/d5.1) for more details. 7The four electronic trading venues are Trad-X (Tradition), BGC Trader (BGC Partners), i-Swap (ICAP), and tpSWAPDEAL (Tullett Prebon, which merged with ICAP), which are authorized multilateral trading facilities (MTFs) in the U.K. and also operate SEFs under U.S. legislation.For the EUR and GBP benchmark assessments, the data are sourced from the MTF order books.For the USD benchmark assessment, data are sourced from the respective SEF order books.
8 MAT is a procedure used to determine whether a swap that is required to be cleared is subject to the trade execution requirement and must be traded on a SEF from the effective date of February 2014 onwards, using one of the minimum execution methods.As such, a SEF establishes whether a swap is MAT based on predefined criteria such as availability of buyers and sellers, and trading frequency and volume, and submits the determination to the Commodity Futures Trading Commission (CFTC) for approval.Once certified by the CFTC, the MAT swap needs to be traded per the trade execution requirement on all SEFs. 9See https://www.cftc.gov/PressRoom/PressReleases/pr7505-order to establish the daily benchmark rates, ISDA computed a trimmed mean of the submitted rates, which depended on the number of bank participants.
The number of submitting banks for USD rates declined from 15 in 2012 to eight in 2014. 11The reasons cited for the decline are numerous, including increased regulatory, compliance, and operational costs.However, it highlighted the need for an alternative, robust, objective, and representative benchmark regime.Furthermore, considering that a benchmark such as LIBOR, with a larger number of contributors, was subjected to manipulation for an extended period, this stark reduction in the number of participants directly threatens the viability and integrity of the ISDAFIX.Indeed, the LIBOR manipulation scandal led to the assessment of other financial benchmarks, including ISDAFIX.It was reported in April 2013 that the CFTC was investigating instances of possible manipulation of the ISDAFIX rates.By August 2013, U.S. investigators had obtained evidence of bank manipulation at the expense of firms and pension funds, 12 leading to fines of over $570 million levied on banks and brokers by the CFTC. 13These events underscored the need to overhaul the ISDAFIX benchmark.
On August 1, 2014, the ICE Benchmark Administration (IBA) took over full responsibility from ISDA for the USD, EUR, and GBP assessments.IBA maintained the old submission-based methodology until March 30, 2015 (inclusive).The change of benchmark administrators was part of a wider attempt to enhance the integrity and robustness of benchmarks after investigations by regulators around the world into claims of misconduct and manipulation of them. 14n March 31, 2015, the IBA transitioned from the submission-based assessment system to an automated and market-based methodology, thus, for the first time assessing the benchmark rates by relying on tradable quotes from regulated electronic trading venues.The benchmark was renamed the ICE Swap Rate, which went into effect on April 1, 2015. 15The methodological change went hand-in-hand with the introduction of regulatory supervision by the FCA, which also started on April 1, 2015.Panel B of Fig. 1 presents a timeline of events.
The ICE Swap Rate is the principal global benchmark setting the fixed leg price for IRS at a particular time of day, and is assessed for tenors ranging from 1 to 30 years.By means of example, the USD ICE Swap Rate, assessed during the morning run, represents the midprice for the execution of an SMS 16 trade.The rate is based on the best available prices across trading venues, collected from 10:58:00 to 11:00:00 ET, and is published at 11:15:00 ET (see Panel C of Fig. 1).
The benchmark is computed by IBA as follows: the 2-min data collection window is divided into 24 blocks of 5 s, and a random snapshot is taken from the order book of each trading venue during each of the blocks.At each snapshot time, the benchmark administrator creates a synthetic order book from the snapshots collected from all venues by ranking the quotes by price.The order book is then used to calculate the volume weighted bid, offer, and average mid-price to execute an SMS order.This process is repeated for each snapshot time, and after discarding illiquid and outlier snapshots, the remaining snapshots are quality-weighted 17 to calculate the ICE Swap Rate.

Order book data and descriptive statistics
For the USD ICE Swap Rate assessments, IBA collects data from three trading venues: Trad-X (Tradition), BGC Trader (BGC Partners), and i-Swap (ICAP). 18We obtain the full proprietary order book data of the Trad-X SEF for London Clearing House (LCH) cleared swaps from Tradition (UK) Ltd. 19 Receiving and processing further data from alternative platforms was not practical due to the sheer size of the order book, which contains over 30 million messages for each day of every tenor and currency.Furthermore, such an endeavor is unlikely to offer further insights given that, for the period of our investigation, Tradition was the market leader in the interdealer brokers (IDB) segment, accounting for a market share of roughly 38% for all USD IRS tenors by volume traded, as outlined in Fig. 2. Across the three benchmark-participating platforms, the single 10Y maturity, which is the main focus of our analysis, accounts for a market share of approximately 50%. 20In addition, depending on the swap tenor, FCA supervisory intelligence suggests that, for the assessment of USD rates in the period of our analysis, around 50%-75% of tradable quotes contributing to the ICE Swap Rate assessment originated from the Trad-X LOB. 21ll usual order book variables and USD tenors, ranging from 1 to 50 years, are recorded in the data.The sample period extends from August 1, 2014, when IBA took over the benchmark assessment, to December 30, 2015-a total of 331 trading days. 22We employ an event study methodology, where March 31, 2015, the effective date of the new benchmark regime, is the event day, d 0 .The ISDAFIX regime, referred to as pre-BRC, encompasses 160 trading days In the data, the 10Y USD swap is the most actively traded on Trad-X, and therefore the target of our analysis.We reconstruct the aggregated 10-level full order book at the end of each second, t, during the normal trading hours of the major U.S. exchanges, from 9:30 a.m. to 4:00 p.m. Eastern Time (ET).
Messages consist of three action types-new order submissions, order changes, and order cancellations-and are timestamped in Greenwich Mean Time (GMT) to the nearest millisecond (ms).Given the USD emphasis, we convert all our time references to local ET.Each message is labeled with a unique order identifier, allowing us to follow its life cycle.A message cancellation is recorded following an active cancellation or after a transaction has been concluded.All messages are indexed by a sequence number, providing an audit trail of unfolding events.We do not observe trader identifiers associated with the quoting and trading activity.
Firm and executable quotes, both outright and implied, are recorded and contribute to the ICE Swap Rate assessment.The division between outright and implied orders is commonly employed in the swap market.An outright order is a direct price submission by a trader, for instance in an individual swap contract.An implied order is generated from the price differential between two existing contracts or spreads.As these are executable, implied orders are crucial for the functioning of the swap market, but do not exist in other markets, such as equities or fixed income that are more commonly studied in the literature.There are two types of implied quotes: implied ins and implied outs.Implied ins are generated by the price differential between two contracts: the difference in prices between two legs (e.g., the 5Y and 10Y swaps) generates a price for the spread.Implied outs generate the price of one leg, on the basis of the price of another leg and a spread.As an example, two orders in the 5Y and 10Y tenors can generate an implied in order in the spread contract between them.Assume a bid in the 5Y tenor at 100 and an ask in the 10Y tenor a 105; a trader could sell the 5Y swap at 100 and buy the 10Y swap at 105, meaning that there is an implied bid in the 5Y-10Y spread contract at − 5.If a trader hits the − 5 bid in the spread contract, the matching engine would fill three separate orders.The sale of the 5Y-10Y spread at − 5 for the trader that hit the bid, the bid at 100 in the 5Y tenor and the ask at 105 in the 10Y tenor for the two traders that were resting on the order books in the 5Y and 10Y tenors.If, rather than knowing the bid and ask prices in the 5Y and 10Y tenors, we have information on the spread and one of the legs, we can calculate an executable implied out message for the missing leg using a similar logic.Traders benefit greatly from these executable implied quotes as they enhance liquidity across the various tenors and spread contracts.The stylized example above only calculates one price (either the spread or the missing leg) on the basis of two others, but similar implied quotes can be calculated for all the tenors and spreads in the market.One can also calculate executable implied quotes from other implied quotes generating second or third generation implied quotes.The calculation of the implied orders therefore links the entire swap curve via spreads and butterflies, allowing traders to transact more cheaply.Calculating these implied quotes on a continuous basis and for all available contracts is computationally complex.However, the Trad-X platform includes an implied engine, explaining the large number of executable implied orders along the swap curve. 23oice trading and RFQ data are not included in our sample due to issues with timestamps that arise by trying to merge highfrequency LOB data with lower-frequency voice-managed orders and transactions.Moreover, upon receiving an RFQ, the SEF must provide the requester with both the quotes received from responding dealers and the firm resting bid and offer prices on the order book [see (Benos et al., 2020) for more details].The requester decides which quote to accept, consequently LOB and RFQ prices are expected to accurately reflect market conditions. 24In addition to the three IDB platforms from which the data for the benchmark assessment are sourced, there are other dealer-to-client venues, such as Bloomberg and Tradeweb.However, according to IBA, the reason that they decided to source data from IDBs was based on the fact that prices on these platforms are firm, while some dealer-to-client platforms operate last-look functionalities and, for this reason, their prices are not considered to be firm.Hence, the dataset obtained from Tradition, the IDB market leader, and consisting of electronically traded swaps, is representative of the market based on which the benchmark is assessed.
Our sample period is characterized by price volatility, arguably driven by macroeconomic and political events.Fig. 4 depicts the midpoint price (where the price of a swap is a percentage rate) of the 10Y USD IRS.The average quoted mid-price for a 10Y USD swap before March 31, 2015 is 2.33 and the average daily price volatility, measured as the standard deviation of the mid-price, during the pre-period amounts to 0.24.After March 31, 2015 (inclusive), the average price and volatility are lower, with values of 2.21 and 0.15 respectively.
Tables 2 and 3 provide descriptive statistics of quotes and transactions.The average pre-BRC best bid and offer (BBO) quote size is $50.66 million, and the post-BRC quote size is $45.18 million.There is less variability in the submitted BBO quotes after the event date ($40.52 million versus $37.14 million).For the 10Y USD swap contract, on an average day, a total of 30.27 million messages are recorded, of which 103,000 are outright orders, while the remaining 30.17 million are implied orders accounting for more than 99% of total message flow.Half of the daily messages are new order submissions, while the other half correspond to their respective cancellations.There are very few order changes (an average of two change messages daily), because canceling and replacing a message is faster, and given the paucity of transactions, time priority is less relevant.Total daily messages, as well as daily implied messages, increased jointly by 33%, from 25.89 pre-BRC to 34.37 million in the post-BRC period.Daily outright order submissions also increased, by 26%, from 91,000 pre-BRC to 115,000 post-BRC.
Given the large number of messages, trading on regulated SEFs is characterized by a low trade-to-quote ratio largely driven by the dynamic swap curves generated by Trad-X's implied engine explained above.In particular, transactions can either be directly executed in the individual swap legs, such as the 10Y IRS, or produced via a "packaged" trade.Packaged transactions, such as swap spreads, curve spreads or butterflies, technically correspond to simultaneous individual transactions in the respective swap legs and are the most frequent.As such, during the full sample period, there were only 165 direct 10Y USD swap trades, averaging less than one transaction per day.However, the daily average combined number of direct and packaged transactions in the 10Y USD swap leg contract on the platform is 21.The overall number of transactions in the 10Y tenor amounts to 6835.Note that the low trade-to-quote ratio, the elevated usage of packaged trades, and the perceived unassuming number of transactions is not a peculiarity of the trading venue that we study but a market reality.The IDB swap market is characterized by a relatively small number of transactions, however, the notional value traded is economically substantial.The average dollar trade size per transaction is a considerable $54.16 million, leading to a non-negligible daily trading value of $1.14 billion.Overall, between August 2014 and December 2015, a total volume of $370.19 billion was traded electronically in 10Y USD swaps on Trad-X alone.For the rest of this paper, we will consider all transactions in the 10Y IRS, direct executions as well as executions in the leg, as part of packaged trades.Post-BRC, daily transactions increased by 7%, from 20.29 to 21.80, while the average trade size remained stable (negligible change from $54.23 million to $54.10 million), and total transactions grew by 14%, from 3190 to 3650.The total volume traded likewise expanded from $172.94 billion pre-BRC to $197.25 billion post-BRC, a gain of 14%.
In summary, executable implied orders dominate, although most of them are canceled without being traded upon, and electronic trades are infrequent but considerable in terms of value.Nevertheless, the firm nature of the quotes ensures their reliability by holding participants accountable for submitted prices.The price discovery process of the market can therefore be compared to the "tâtonnement" process described in Biais et al. (1995, 1999), where the order flow in itself is informative, and the efficient price is discovered in a gradual learning process, even when no orders are executed.

The power of benchmarks: implications for market quality
To test the predicted beneficial effects of benchmarks, we examine the observed effects of the BRC on the quality of the swap market.Theory contends that an increase in benchmark accuracy through the transition to a transparent market-based assessment and regulatory oversight will have positive effects on the liquidity of the underlying market.

A more liquid market?
We focus on five metrics to measure market liquidity.The first is the quoted dollar spread (QS) and it is defined as the difference between the BBO prices, computed for each second t: (1) Secondly, we compute the relative quoted spread (RQS), defined as the ratio of the quoted spread and the quoted mid-price (M t ).The relative spread is sensitive to movements in the market price, which in our case is volatile and on average lower during the post-BRC period (see Table 2).Hence, we only use this measure to corroborate our results, since a lower price should lead to a larger relative spread if quoted spreads remain constant.
The third and fourth liquidity proxies employ market depth data.Specifically, quoted depth (QD) and 10-level quoted depth (QD10), are defined as the sum of the offer volume (V A t ) and the bid volume (V B t ) at the second t at the best level and the best ten levels (l = 1, ..., 10) of the order book respectively: Finally, we also develop an additional measure of the spread, which we call the "fill spread".The measure is useful in markets characterized by a LOB with active quoting but very few transactions, such as the one we examine, and approximates the effective spread. 25Typically, the effective spread is computed as 2 × DIR t × (P t − M t ), where DIR t is a directional parameter accounting for buyer-initiated and seller-initiated transactions and P t is the transaction price.A trader could either buy or sell an SMS swap of $50 million.Since we simulate the filling of both a buy (F A t ) and a sell (F B t ) SMS order for each second against existing orders on the book,  DIR t is immaterial.The hypothetical fill spread can thus be written as: (5) As the comparisons to the mid-price in equation ( 5) cancel out, however, this can be written as the difference between F A t and F B t as in equation ( 6).
In other words, the fill spread measures the roundtrip costs of completing a buy transaction and a sell transaction, approximating the liquidity on both sides of the order book at second t.Since quote sizes at the best level (and beyond) vary and commonly account for less than the $50 million standard trade size (see Table 2), the fill spread is an aggregate measure accounting for both the prevailing spread and depth of the order book.Hence, in our opinion this is the best proxy to measure changes in liquidity: For robustness, we also compute the relative fill spread (RFS), defined as: All measures are time-weighted (TW), as shown in the following equation, where LM t represents one of the above-described liquidity measures.t is the second timestamp of the i = 1, ..., N intraday quote update on day d.T is the length of the trading day.
By time-weighting over the length of the trading day, we report average daily liquidity measures that account for the full trading day and not just the most liquid trading periods.Still, in Fig. 3, we report the intraday quoting and trading evolution.In Panel A, the average hourly quoting activity for the 10Y swap is unsurprisingly most pronounced during normal trading hours (9:30 a.m. to 4:00 p. m.ET), which corresponds to the time frame of our analysis.On-platform messages peak around 9:00 a.m. and remain elevated until about 11:00 a.m. and then reduce gradually.In Panel B, the hourly sum of transactions over the sample period peaks during the hour (10:100 a.m.) ahead of the benchmark assessment and publication and remains elevated until mid-day.The results presented below are therefore conservative, since the applied measures cover the most and least active trading periods.
In Table 4, we report the long-term comparison of the liquidity measures by splitting the sample period into the periods before and after the exogenously determined event date. 26We report three spread measures and two market depth measures.Quoted spreads and relative quoted spreads are both significantly lower in the post-BRC period.The average daily time-weighted quoted spread (TWQS) decreases from 0.7 bps pre-BRC to 0.6 bps post-BRC, a reduction of 14%.Similarly, the average daily time-weighted relative quoted spread (TWRQS), which accounts for fluctuations in the price, narrows from 0.31 bps to 0.27 bps, a drop of 11%.The improvement in the time-weighted average spread measures is significant at the 1% level.Variations in the width of the spread measures reduce after the BRC, with the average daily standard deviation declining 34%-37%.Our results also hold if we use daily median values.
We complement the spread analysis with a study of market depth, both at the BBO level, as well as at the bids and offers across ten levels of the order book.Columns ( 5) and ( 6) of Table 4 report the results for the time-weighted quoted depth measures.On the one hand, the average daily quoted depth is lower during the post-BRC period ($100 million vs. $90 million), a deterioration of 10% at the 1% significance level.On the other hand, 10-level quoted depth increases marginally, from an average daily value of $3.39 billion pre-BRC to $3.52 billion post-BRC.However, this 4% increase in TWQD10 is not statistically significant.Again, the results are consistent when we use median values. 27n short, post-BRC spreads narrow and the order book at the first ten levels is marginally deeper, but depth at the top of the book is thinner.Traders, however, are interested in filling their order and the costs of trading.Consequently, in column (3) of Table 4, we report the results for the time-weighted fill spread (TWFS), our aggregate measure of the simultaneous impacts on spreads and depth.The average (median) daily fill spreads on the Trad-X platform in the post-BRC period narrow from 0.78 (0.74) bps to 0.7 (0.68) bps, a decrease of 11% (8%) at the 1% significance level.This result shows that it is cheaper to trade electronically under the ICE Swap Rate regime.Also, the total number of times that an SMS order cannot be completed (on a second-by-second basis) on the Trad-X platform on either side of the book due to missing liquidity decreases from 885 in the pre-BRC period to 326 in the post-BRC period.This corresponds to a drop of 63%.The finding is indicative of a more resilient order book, with traders confidently posting executable quotes.

Regulation as a driver?
Next, we employ a difference-in-differences (DiD) approach to link the observed improvements in market quality to the BRC.Specifically, we compare the changes in liquidity for tenors with a regulated benchmark assessment vis-à-vis those without a regulated 26 Please refer to the robustness section for the short-term liquidity effects. 27For robustness, we compute an alternative measure of order book depth by simulating the continuous filling of a large transaction (several multiples of the 10Y SMS).We find a highly significant improvement in execution costs for large and very large transactions too.The results are in Table B.1 of Appendix B. benchmark by estimating the following regression model: Treatment is a dummy taking the value 1 for tenors that are part of the treatment group described below, meaning a benchmark is assessed, and zero otherwise.The 10Y USD IRS is selected for the treatment group-the most actively traded MAT tenor for which a benchmark rate is assessed-and therefore expected to benefit from the increased benchmark precision following the change in regulatory regime.The 12Y USD IRS is chosen for the control group as it is the only MAT tenor for which no benchmark rate is assessed (see Table 1).For our identification strategy to be valid, these tenors need to exhibit parallel spread patterns ahead of the regime change.Figure B.1 in Appendix B shows that despite differing quoting and trading levels between the treatment and control tenors, we observe parallel trends.Hence, our DiD model is correctly specified.Moreover, we include several control variables in our model to further account for the different liquidity profiles.29X d is a vector of control variables that includes swap and debt market volatility, venue participation, quoting and trading behavior, and macroeconomic developments.β 1 captures any common effects that might have impacted all swap tenors following the BRC.β 2 absorbs any pre-existing differences in characteristics between the treatment and control groups.The coefficient of interest is β 3 , which captures the interaction of Event and Treatment and thus estimates any incremental effects of the BRC.Hence, β 3 reflects the change in liquidity for tenors that are part of the benchmark regime compared to the change in liquidity for tenors that are not.We use tenor fixed effects to estimate the model.Table 5 reports the estimation results.The DiD model is estimated under various specifications, excluding and including control variables (labeled as [1] and [2] respectively).The results show that there is little difference in the coefficients of interest between the two specifications.Overall, our control variables help to explain a significant proportion of the evolution of our liquidity measures, with an adjusted R 2 of 67% and 56% for the regression models in which the dependent variables are TWQS and TWFS, respectively.
First, with the BRC there is an improvement in TWQS for both groups of swap tenors (10Y and 12Y), as indicated by the negative and highly significant Event coefficient.Importantly, however, the significant Interaction term in Table 5 shows that the enhancement in TWQS for the 10Y tenor is greater than the improvement in the 12Y tenor.The Interaction coefficient for TWFS indicates that the execution costs for the 10Y USD IRS have also come down significantly, and crucially by more than those for the 12Y USD IRS, following the change in benchmark assessment methodology and the regulation by the FCA.The results are equally strong irrespective of whether we include multiple controls in the model specifications, suggesting that the liquidity improvement is over and above the other effects driving swap market liquidity.
There are no effects on either the quoted spread or the fill spread from changes in the IRS volatility (SRVIX) or the U.S. Treasury note volatility (TYVIX).An increase in quoting (MESS_10Y), however, appears to translate into significantly wider spreads and execution costs.This seems counterintuitive, but the observed reaction is driven by the fact that MESS_10Y is correlated with several other variables, which already capture narrowing spread effects: Event (40% Pearson correlation), PARTICIPANTS (52% Pearson correlation), MESS_12Y:10Y (34% Pearson correlation), and TRANS_10Y (29% Pearson correlation).Trading activity (TRANS_10Y) has a negligible effect on liquidity.The ratios of messages (MESS_12Y:10Y) and transactions (TRANS_12Y:10Y), proxying for a change in the liquidity pattern between the 12Y and 10Y tenors, do not affect our liquidity measures.The number of USD streamers (PAR-TICIPANTS), depicted in Fig. 5, has a strongly positive effect on the liquidity metrics.An increase in the number of participants on the trading venue around the event date leads to a sharp and highly significant reduction in quoted spreads and fill spreads.This aligns with the assertion that increased on-platform participation leads to a liquidity improvement, which is consistent with empirical market microstructure findings (e.g., Barclay and Hendershott, 2004).Unsurprisingly, macroeconomic announcement days (MACRO) are characterized by a significant widening of spreads and inflation of execution costs, which is in line with expectations, due to the increase in uncertainty on such days.Finally, a change in the ratio of outright to implied orders (O:I_10Y), for example due to a reduction in implied quotes and therefore a lower level of available liquidity translating into an increase in the ratio, leads to a widening of spreads.
Importantly, even after controlling for a multitude of potentially confounding effects, our findings show a significant incremental improvement in on-platform execution costs for benchmark-grade swaps.Taken together, our results suggest that the liquidity improvement is driven by the exogenous regulatory change and methodological evolution of the benchmark.
As a robustness check, in Subsection 5.1, we report the results of the analysis using multiple tenors, where the treatment group comprises the 2Y, 5Y, 10Y, and the 30Y tenors and the control group comprises a total of 17 tenors. 30The results are consistent when using this expanded specification and, if anything, the observed liquidity improvements are of bigger magnitude.Note that many of these tenors are not MAT swaps and hence not required to be traded on SEFs, unlike the 12Y tenor used in the main analysis.We report the multi-tenor results in Table 7, but base our conclusions on the model with the 10Y and 12Y tenors, where the treatment and control groups are directly comparable.

Table 4
Quoted liquidity under the ISDAFIX and ICE Swap Rate regimes.This table reports the results of a comparison of the liquidity variables before and after the BRC.TWQS reports the spread in absolute dollar terms.TWRQS reports the ratio of the quoted spread to the mid-price.TWFS reports the difference between the hypothetical execution price of an SMS trade on both sides of the book as per the methodology section.TWRFS reports the ratio of the fill spread to the mid-price.TWQD is the sum of the depth at the best bid and offer prices.TWQD10 is the sum of the depth at the bid and offer sides of the 10-levels of the order book.All liquidity measures are computed as daily averages (medians) and then averaged across the period of interest.The median captures the weighted median (by number of occurrence) of the liquidity measures.Standard deviation reports the average daily standard deviation of the liquidity measures.We also count the number of times that an SMS order cannot be completed (on a second-by-second basis) on either side of the book.TYVIX is the log return on the 10-year U.S. Treasury Note Volatility Index.MESS_10Y is the log daily count of the number of messages received by the platform operator for the 10Y IRS contract.MESS_12Y:10Y is the log ratio of messages for the 12Y contract relative to the 10Y contract.TRANS_10Y is the log daily number of transactions in the 10Y IRS contract.TRANS_12Y:10Y is the log ratio of the number of transactions in the 12Y contract relative to the 10Y contract.PARTICIPANTS represents the log number of USD streamers per trading day.MACRO is a dummy variable that takes the value 1 on days with macroeconomic announcements by the Federal Open Market Committee (FOMC) and the Governing Council of the ECB and 0 otherwise.O:I_10Y is the log ratio of outright to implied messages in the 10Y IRS contract.All spread measures are expressed in bps (1 bps = 0.01%).The models are estimated using tenor fixed effects.We use Driscoll and Kraay (1998)   The effects of the regulation are economically significant too.The costs savings, as measured by the total effect of the BRC on electronically executed 10Y USD swaps on the Trad-X platform alone, amount to between $3.33 million and $9.92 million. 31The marginal cost savings, computed on the basis of the incremental reduction in execution costs of the 10Y benchmark-grade swap tenor over the 12Y non-benchmark-grade tenor, range between $3.6 million and $6.7 million.Given that we only focus on one tenor and that the swaps can be traded on other venues too, the overall benefits are likely to be substantially larger.Conservative, parsimonious calculations extrapolate the cost savings to the total USD IRS notional traded on the three electronic trading venues that feed into the benchmark assessment, and put the overall cost savings in the region of $550 million.

When do the structural breaks in the time series occur?
Thus far we have relied on an exogenous determination of the event date to assess the implications of the BRC for liquidity, but changes in the microstructure of the underlying market could have occurred before or after the event date.Specifically, the measures were computed for the periods before and after the changes were introduced to the methodology and the benchmark was regulated.In this subsection, we discuss our endogenous determination of the statistical structural breaks in the liquidity measures.In doing so, we aim to provide a clearer picture of the liquidity's reaction to the BRC and during the period thereafter, as well as rule out the concept of "leakage", i.e., a preemptive adjustment ahead of the actual introduction of the changes.We follow the approach of Bai and Perron  (1998, 2003, BP), the application of which is described in Zeileis et al. (2003).The model set-up is based on a standard linear regression of the form: where y d and x d correspond to the values of the dependent and explanatory variables respectively on day d.β d is the regression coefficient, which can vary over time.The model tests the null hypothesis that the coefficient remains constant over time, versus the alternative of a change in the coefficient over time: In the method, we assume that there are m breakpoints in the time series at which points the mean of the coefficient moves from one long-term level to another.Hence, the set of breakpoints, which are unknown, must be endogenously estimated.m breakpoints imply m+1 segments with a constant coefficient.Based on Bai and Perron (2003), in order to date the structural changes, we use a dynamic programming algorithm to compare different combinations of m-partitions to achieve a minimum global residual sum of squares.We use this process to sequentially examine the partition of m+1 versus m breaks and compares which of the partitions provides the overall minimal residual sum of squares compared to one additional segment.
In our case, we apply a pure structural change model, and we test whether the mean of the liquidity measure in question changes over the course of our sample period.To do so, we fit a constant to the time series data of the dependent variable.We apply a trimming factor of 15%, as suggested by Bai and Perron (2003), allowing for a maximum of five breaks.The trimming factor determines the minimum number of observations in each segment.Since our sample consists of 331 trading days, the trimming value implies that each segment is required to have at least 49 observations.We determine the optimal number of breaks as in Zeileis et al. (2003).
Fig. 6 depicts the determined structural changes in the time series of the four different liquidity measures.The TWQS, TWFS, and TWQD10 experience two breaks each, while the TWQD shows three breaks.The common pattern that can be established is that, for each of the four liquidity measures, one break occurs very shortly before the BRC.For both spread measures, the multiple structural break models indicate a first break (upward) in the data on December 4, 2014.We identify two potential reasons for this change: (1) European Central Bank (ECB) president Mario Draghi announcing a potential quantitative easing intervention and (2) a drop in the number of USD streamers on the trading venue.On December 5, 2014, the number of dealers on the platform falls by roughly 45% (see Fig. 5), which could also be the cause of the observed widening of spreads.The number of dealers recovers to its previous level on the next day and stays relatively stable after that but, clearly, liquidity does not recover.However, participation over the following days is volatile, possibly explaining the wider spreads throughout the period from December to March.The second downward break occurs on March 26, 2015, three trading days before the BRC. 32Given the proximity to the event date of March 31, 2015 and the fact that structural breaks are usually modeled on less granular data (often monthly), we attribute this change in the long-term pattern to the imminent BRC and rule out an extended preemptive adjustment before the introduction of the ICE Swap Rate.There was also no major macroeconomic event around the break day.Duffie et al. (2017) suggest that improved price transparency generated by a benchmark encourages entry by traders and stimulates dealer competition on prices, which at the same time may lead to inefficient dealers exiting the market.In addition, in the spirit of Aquilina and Pirrone (2020), we argue that a more precise and regulated market-based benchmark encourages, among other things, greater participation, positively impacting market liquidity.The fact that on March 26, 2015 the Trad-X platform experiences a 10% increase in the number of participants is in line with our argument.Fig. 5 illustrates that the number of platform participants remains above its long-term average during the large majority of the post-BRC period.
The breaks we determine for the two depth measures are different.The quoted depth time series shows three breaks: December 18, 2014, March 26, 2015, and October 7, 2015.The first and third breaks are different than the breaks established for the spread measures, but importantly the second downward break immediately precedes the BRC and suggests a slight reduction in depth at the best order book level, which is consistent with our earlier findings in Section 4.1.For the 10-level depth time series, the BP multiple structural break model identifies two breaks: March 24, 2015 and October 7, 2015.The October break is identical to the quoted depth's October break, but the March break occurs five trading days before the BRC.The fact that all liquidity measures identify a break in the long-term time series imminently prior to the transition to the ICE Swap Rate regime supports our earlier findings in Sections 4.1 and 4.2.

Alternative specifications
In addition to estimating our model using one tenor in the treatment group and one in the control group as discussed in Subsection 4.2, we perform two additional sets of robustness checks on the model itself.The first is to replicate our analysis using relative measures of the spread and depth measures to check that our results do not depend on the level of prices.The second is to expand the number of tenors in the analysis to check that the results remain valid along the maturity dimension. 33To do so, we check that our results generalize along the maturity curve and run the analysis using the 2Y, 5Y, 10Y, and 30Y tenors as the treatment group and the 11Y, 12Y, 13Y, 14Y, 16Y, 17Y, 18Y, 19Y, 21Y, 22Y, 23Y, 24Y, 26Y, 27Y, 28Y, 29Y, 50Y as the control group. 34The control variables employed in the extended specification follow the same rationale as in the base specification.If a control variable exists for some tenors but is missing for others, this is due to either high levels of correlation between the independent variables or missing data points, for example in the case of transaction-related controls.
Table 6 and B.5 report the results of re-estimating the model in equation ( 9) using relative measures of the spread for the analysis of the 10Y versus the 12Y swaps and for the extended sample, respectively.In all cases, the interaction dummy remains negative and highly statistically significant, confirming the results obtained earlier in Table 5.Both the time-weighted relative quoted spreads and the relative fill spreads have significantly narrowed following the BRC.Liquidity in the swap tenors affected by the BRC has thus improved over the liquidity in those swap tenors that are unaffected. 35able 7 reports the results of including a much bigger number of tenors in the control and treatment groups.In these cases, the interaction dummies for TWQS and TWFS remain negative and significant.We therefore conclude that, within our empirical setup, the liquidity improvement for the swaps for which the benchmark is calculated is caused by the BRC.This result is robust to different ways of measuring liquidity, as well as different ways of constructing the treatment and control groups. 36o further corroborate our results, we next discuss additional evidence related to changes in liquidity in the tenors subject to the BRC.

What are the short-term liquidity effects around the event date?
We use the event study methodology employed in Hegde and McDermott (2003) to look at short-term liquidity behavior and assess the validity of the reported long-term results.We calculate our liquidity measures over different time intervals surrounding the event date of March 31, 2015 and compute a ratio comparing them to the long-term average of the estimation window d − 160 = August 1, 2014, d − 30 = February 13, 2015], which extends to 30 trading days before the regime change, and represents a period that is unlikely to have been affected by the BRC.If the ratio of the liquidity measure for some interval in Table 8 is greater (smaller) than unity, it indicates that the interval average is greater (smaller) than the estimation window average.We discuss the TWQS here since, for robustness purposes, we wish to measure on-platform liquidity directly at varying intervals around the event date.
The ratio obtained using the average (median) time-weighted quoted spread for the interval [0; 0], covering only the event date of March 31, 2015, is 0.87 (0.94), considerably below its long-term average.For the first five intervals ([− 1; +1], [− 2; +2], [− 3; +3], [− 4; +4], and [− 5; +5]) centered on the event date, the average daily TWQS ratio indicates that the spreads are significantly lower (at the 5%-1% level) than their long-term average.During the 11-day interval [− 5, +5] centered on the event date, the average as well as median spreads are significantly lower, with values of 0.92 and 0.96, at the 1% and 5% significance levels respectively.The results for longer time periods are insignificant.Importantly, the findings for the intervals [− 30; − 1] and [+1; +30] demonstrate that the narrowing of spreads is driven by a significant decrease in the post-BRC period rather than the pre-BRC period, as shown by the ratio of 0.89 at the 1% significance level versus the insignificant ratio of 1.08 for the post-BRC and pre-BRC periods respectively.
Since the earlier long-term results on depth were less clear-cut, the event study findings on TWQD and TWQD10 are of particular interest.The average time-weighted quoted depth at the best level is above its long-term average on the event date [0; 0] itself (1.03), although its median is below unity and further drops significantly below the estimation window reference value for the intervals [− 1; +1] and [− 2; +2].The interval [− 30; − 1] shows that TWQD is above its long-term average (1.09) at the 1% significance level ahead of the BRC.During the 30 days [+1; +30] after the BRC, the quoted depth (ratio of 0.99) is not significantly different from the reference value of the estimation window.Regarding the average 10-level quoted depth, the book is much deeper on the event date [0; 0] with a 33 Our preferred specification remains the one comparing the 10Y and the 12Y tenors which is the cleanest test given that the tenors share the same characteristics, including both being MAT tenors. 34The control group includes more tenors as the computational requirements to reconstruct the order book are much smaller than to reconstruct the order book for the treatment group given they are considerably less active. 35Table B.3 in Appendix B reports the results for the relative depth measures for the 10Y vs 12Y analysis.For the depth measures, the interaction coefficients and significance for the relative measures remain in the same direction and strength compared to the nominal measures, again confirming our main results.Depth at the best level has been unaffected by the BRC and the 10-level quoted depth has become significantly deeper. 36Tables B.4 and B.6 in Appendix B reports the results for the nominal and relative depth measures, respectively, for the expanded set of tenors.
The coefficients for TWQD and TWRQD turn significantly negative and TWQD10 and TWRQD10 become indistinguishable from zero suggesting that the depth of the order book became shallower at the best order book level.Importantly, however, our nominal and relative fill spread measures, which account for the overall effect on spreads and depth still show a highly significant improvement in liquidity.value of 1.27.The [− 30; − 1] interval shows that the 30 days before the regime change are characterized by a slightly thinner order book (median ratio of 0.97 at the 10% significance level).The [+1; +30] period, however, shows a deeper order book (highly significant average ratio of 1.11 and median ratio of 1.12).The event study results confirm our early findings, suggesting that market liquidity has reacted to the BRC, and done so positively.

A more precise benchmark?
In this subsection, we establish a closer link to the theory by contributing non-causal37 yet indicative evidence in support of the predicted increase in the precision of the benchmark assessment.We analyze whether the benchmark is more accurate, i.e., closer to market fundamentals, by comparing the benchmark rates under the ISDAFIX regime and the IBA regime to market prices available on regulated trading venues. 38o measure changes in the quality of the benchmark, we develop a simple measure termed the benchmark-to-market differential (BMD).The ISDAFIX ahead of March 31, 2015 represents the rate at which dealer banks are willing to buy and sell a swap of an SMS ($50 million for 10Y USD IRS) each day before the end of the polling period.The new ICE Swap Rate assessment methodology calculates the benchmark rate by continuously simulating the filling of an SMS order during a 2-min time window.Hence the benchmark rate should be indicative of market conditions and thus act as a representative price for the execution of an SMS trade, under both the ISDAFIX regime and the ICE Swap Rate Regime.
We define the BMD as: where R d is the assessed benchmark rate on day d and F t,d = is the estimated average of the buy and sell prices for an SMS order ) is the hypothetical execution price for an SMS buy (sell) order simulated for each second t, assuming that an aggressive buyer (seller) crosses the spread and consumes liquidity on the ask (bid) side of the order book. 39A small differential is interpreted as a benchmark rate that is indicative of market fundamentals as expressed by the wider market.
The pre-and post values in Table 9 report the average daily BMD during the ISDAFIX and ICE Swap Rate regimes respectively.The pre-BRC and post-BRC regimes differ both in terms of methodologies (panel-based vs. market-based) and in terms of assessment lengths (15 vs. 2 min).For reasons of comparability and robustness, we average the BMD t,d over multiple windows of different length (1 min, 10 min, 30 min, etc.), centered on the 11:00 a.m.ET assessment time and averaged across days within the pre-BRC and post-BRC periods.This allows us to provide a more comprehensive picture of the representativeness of the rate.
For the 1-min window [11:00:00; 11:00:59], the result indicates that an on-platform execution of an SMS order would have, on average, been executed closer to the benchmark rate under the old regime (0.11 bps vs. 0.15 bps differential).This difference, however, is likely driven by the differing assessment methodologies.Under the ISDAFIX regime, panel banks submitted point estimates that were concentrated at 11:00 a.m.ET and thus, by construction, the difference between the assessed rate and the market price at that point in time will have been small.The ICE Swap Rate, however, is essentially a 2-min average of the market price from 10:58:00 to 11:00:00,

Table 7
Difference-in-differences panel regressions for spread measures with multiple tenors.This table reports the results of the DiD panel regression model specified in equation ( 9) using the spread measures TWQS and TWFS as dependent variables.(1) presents the DiD model without controls while (2) presents the same specification with controls.Event is a dummy variable that takes the value 0 for the pre-BRC period [d − 160 = August 1, 2014, d − 1 = March 30, 2015] and 1 for the post-BRC period [d 0 = March 31, 2015, d 170 = December 30, 2015].Treatment is a dummy that takes the value 1 for benchmark-grade swaps (2Y, 5Y, 10Y, 30Y) and 0 otherwise (11Y, 12Y, 13Y, 14Y, 16Y, 17Y, 18Y, 19Y, 21Y, 22Y, 23Y, 24Y, 26Y, 27Y, 28Y, 29Y, 50Y).Interaction is a dummy variable computed as Event × Treatment.SRVIX is the log return on the Interest Rate Swap Volatility Index.TYVIX is the log return on the 10-year U.S. Treasury Note Volatility Index.MESS_10Y and MESS_13Y are the log daily count of the number of messages received by the platform operator for the 10Y and 13Y IRS contract respectively.MESS_12Y:10Y and MESS_13Y:2Yis the log ratio of messages for the 12Y or 13Y contract relative to the 10Y or 2Y contract.TRANS_2Y, TRANS_10Y, TRANS_11Y, TRANS_13Y, and TRANS_30Yare the log daily number of transactions in the 2Y, 10Y, 11Y, 13Y, and 30Y IRS contracts respectively.TRANS_12Y:10Y and TRANS_11Y:5Y are the log ratios of the number of transactions in the 12Y contract relative to the 10Y contract and the 11Y contract relative to the 5Y contract respectively.PARTICIPANTS represents the log number of USD streamers per trading day.MACRO is a dummy variable that takes the value 1 on days with macroeconomic announcements by the Federal Open Market Committee (FOMC) and the Governing Council of the ECB and 0 otherwise.O:I_2Y, O:I_10Y, O:I_11Y, and O:I_30Y are the log ratios of outright to implied messages in the 2Y, 10Y, 11Y, and 30Y IRS contracts respectively.All spread measures are expressed in bps (1 bps = 0.01%).The models are estimated using tenor fixed effects.We use Driscoll and Kraay (1998)   39 We use hypothetical execution prices because of the lack of enough direct swap trades per day in the 10Y USD IRS.As reported in the descriptive statistics section, over the full period only 165 direct 10Y USD IRS were executed electronically.As a check we also compute the BMD based on the few executed transactions and find a qualitatively similar result.
introducing stronger price sensitivity, and therefore a larger differential from the market price at 11:00 a.m.ET.Hence, we argue that a comparison of the benchmark rate to the estimated average execution price, for different time windows centered on 11:00 a.m.ET, is most meaningful.By extending the window length over which we compute the BMD measure, we find that post-BRC the benchmark rate is indicative of market prices for an extended period.For the 4 min, 10 min, 20 min, 30 min, and 60 min comparisons, the BMD is 3%-12% lower under the new regime than the old regime.Based on the 10-min window, we confirm that the BRC did positively affect the representativeness of the benchmark rate at the 5% significance level.Moreover, the BMDs at the respective assessment ends and publication times of the ISDAFIX and ICE Swap Rate regimes are substantially smaller under the new benchmark regime (reductions of 68% and 22% respectively)-results, which are statistically significant at the 1% level.The 10-min window immediately preceding the start of the benchmark assessments is supportive of a more accurate benchmark too.The market price should be indicative of how dealers value an SMS swap at the time of the assessment, and the quote submissions ahead of the

Table 8
Short-term liquidity reaction to the benchmark regime change.This table reports the short-term reaction of the liquidity variables around the BRC.
Interval represents the time period, in number of days d ε D, before and after the event date [d 0 = March 31, 2015], over which the liquidity measures are averaged.TWQS reports the spread in absolute dollar terms.TWQD is the sum of the depth at the best bid and offer prices.TWQD10 is the sum of the depth at the bid and offer sides of the 10-levels of the order book.All liquidity measures are computed as daily averages (medians) and then averaged across the intervals of interest.The ratios are computed relative to a reference value, which is the average of the same liquidity measure over the estimation window [d − 160 = August 1, 2014, d − 30 = February 13, 2015].All values are ratios.The t-value is the statistic of a one-sample t-test of μ = 1.*, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.'' is reported when the significance could not be assessed due to the small sample size of the interval.assessment start should thus reflect the upcoming benchmark rate.With a mean value of 0.35 bps versus 0.30 bps, the average daily BMD during the 10-min window [10:48:00; 10:58:00] is 15% smaller at the 5% significance level during the post-BRC period.
The results in Table 9 are, therefore, in line with theory, suggesting that stricter assessment instructions and controls increase the price accuracy of the benchmark.Specifically, we show that the ICE Swap Rate more precisely reflects market conditions at the end of the assessment and its publication.
In a further test of the improvement in price accuracy of the benchmark during the post-BRC period, we estimate a series of predictive regressions (e.g., Chordia et al., 2008; Chung and Hrazdil, 2010) as follows: mean(M) period,d = α + βR d + ε d (13)   and where period denotes the time period following the publication of the benchmark over which the average mid price (M) is calculated and tenors and d denotes days.R d is the assessed benchmark rate on day d.X d is the same vector of control variables used in equation ( 9).Specifically, we test the accuracy of the benchmark as a predictor of the prevailing market price pre-and post-BRC by regressing the average mid-price on the Trad-X platform, over different time windows following the publication of the benchmark, on the benchmark price assessed on that day.The windows employed are 5, 10, 30, and 60 min following the benchmark publication.Two versions of the model are estimated: one with controls and one without.The root mean square error (RMSE) and mean absolute error (MAE) estimates of the models pre-and post-BRC are in Table 10.Panel A shows the estimates obtained from the regression models without controls,

Table 10
Prediction accuracy.This table reports the accuracy with which the benchmark price is predicting the prevailing mid-market price before and after the BRC.We regress the average mid-price on the Trad-X platform, over different time windows following the publication of the benchmark, on the benchmark price assessed on that day.The different windows are 5, 10, 30, and 60 min following the benchmark publication respectively.We estimate the regression models without and with controls.The controls are the same as used in in equation ( 9  while Panel B presents the estimates obtained from the regressions with controls.The results of two statistical tests are also presented in both panels.The first test, based on the squared residuals of the equations yielding the error estimates, is in the spirit of Fama and  MacBeth (1973), and includes employing the standard errors of the squared residuals for statistical inference.Specifically, the standard errors of the squared residuals of the pre-and post-BRC models are used to test the null that there is no difference in the squared residuals for the pre-and post-BRC models generating the MAE and RMSE estimates.The second test is the Diebold and Mariano (1995)  test comparing the performance of the forecasting models pre-and post-BRC.This test is appropriate in this context because estimating the pre-and post-BRC models implies parameter instability; therefore, the pre-and post-BRC models can be viewed as distinct.
The estimates show a clear reduction in the dispersion of the regressions' residuals from the pre-to the post-BRC periods irrespective of the measure or windows used, or even whether or not we include controls in the regression models.For example, the RMSE estimates for the regressions without controls for the post-BRC periods are always lower for all the four time windows analyzed.The trend is consistent for the MAE estimates.Therefore, the estimates further show that the benchmark price during the ICE Swap Rate regime better reflects market conditions in comparison to the ISDAFIX regime.It can also be inferred from the results that the reductions observed in the RMSE and MAE estimates post-BRC are statistically significant.

A more efficient price?
Finally, the reduction of noise in the benchmark should also translate to less noise in the prices.Hence, to test the informational efficiency of market prices and provide some indicative evidence in support of this claim, we estimate "unbiasedness regressions" consistent with Biais et al. (1995, 1999).The level of price efficiency for the 10Y swap is computed for both the pre-BRC and post-BRC periods through separate estimations of equation ( 15) and the averaging of the slope coefficients across seconds t. r oc = α + βr ot + ε ot .(15)   where r oc is the open-to-close return for the time period of interest and r ot is the return from the opening of the chosen period to the second t.Since our interest lies in the benchmark assessment period, we define the open and close to be 10:58:00 and 11:30:00 respectively.According to Biais et al. (1995, 1999), β measures the signal-to-noise ratio.Since the observable return consists of the true return and some noise element (Barclay and Hendershott, 2003; Ibikunle, 2015), a coefficient close to one suggests informationally efficient prices, while a coefficient smaller than one is consistent with noisier prices.A coefficient bigger than one may be driven by stale prices.our estimation window, the returns from 10:58:00 to interval t do not explain the total 10:58:00-to-11:30:00 return well.Still, as time progresses, it becomes apparent that noise decreases rapidly and the swap price efficiency improves continuously.Under the ICE Swap Rate regime, informational efficiency is achieved faster than under the ISDAFIX regime, as the coefficient quickly converges to unity and remains at that level.Returns from the pre-BRC period indicate that swap prices are noisier (as indicated by coefficient values below one) and informational efficiency takes longer to achieve.By means of example, β reaches unity at approximately 11:04:00 in the post-BRC regime, while it only achieves similar levels roughly 15 min later in the pre-BRC regime.Overall, the results suggest that, in the post-BRC period, price efficiency is enhanced compared to in the pre-BRC period.

Conclusion
We present empirical evidence in support of the recent regulatory interventions in financial benchmark assessments.According to theory, by increasing the transparency, regulatory oversight, and monitoring of the benchmark fixing process, such interventions can have direct beneficial market outcomes in terms of efficiency and liquidity.As such, our study complements the theoretical work of Duffie et al. (2017), who show that benchmarks can increase social surplus and have positive welfare implications.
Our empirical analysis of the natural experiment provided by the transition on March 31, 2015 from the unregulated panel-based ISDAFIX regime to the regulated market-based ICE Swap Rate regime is linked to a measurable improvement in market liquidity.The regulatory intervention by the FCA led to an overhaul of the assessment methodology for the principal IRS benchmark, making it more transparent and tightly monitored, the effects of which should be comparable to the efficiency improvements in Duffie et al. (2017) and the reduction of noise in Aquilina and Pirrone (2020).The liquidity improvement translates into reduced execution costs for the participants in electronically traded swaps.The cost savings for electronic transactions in the 10Y USD IRS from April 2015 to December 2015, on the Trad-X platform alone, amount to between $4 and $7 million.A large part of the liquidity enhancement is already captured by an increase in the number of venue participants, which coincides with the regulatory intervention-our estimate is conservative since it is impossible to attribute the beneficial impact of the concurrent increase in participation to the benchmark transition.Nonetheless, applying a DiD technique, we can attribute the liquidity enhancement to the ISDAFIX-to-ICE Swap Rate transition induced by the regulatory intervention of the FCA.Specifically, the effect is stronger for tenors with a daily benchmark determination, which are impacted directly by the change in the benchmark regime, compared to tenors without reference rates.Hence, our results suggest that the influence of the regulatory regime is beyond the effect of other confounding events.We also find that the accuracy of the benchmark itself improved following the regulatory change, with the benchmark rates now more closely reflecting the market prices.
The results should be interpreted with some caution for two reasons.First, we only analyze the order book data of the main interdealer platform, Trad-X.While Trad-X is the principal inter-dealer platform, the contributions of the other platforms to the ICE Swap Rate benchmark assessments are not negligible.Moreover, developments in market quality on the other contributing venues and dealer-to-client platforms might look different to the observed reaction on Trad-X.However, given that these markets are traded electronically, we would expect participants to arbitrage out any meaningful differences across platforms.Secondly, this study only captures electronic trading, and is unable to account for voice broking.However, consolidating electronic and voice trading activity is not currently advisable given the inevitable timestamping issues that are bound to arise.As a whole, this study robustly demonstrates that transparent and appropriately regulated benchmarks can contribute to better financial markets.

Table B 3
Difference-in-differences panel regressions for relative depth measures.This table reports the results of the DiD panel regression model specified in equation ( 9) using the relative depth measures TWRQD and TWRQD10 as dependent variables.(1) presents the DiD model without controls while (2) presents the same specification with controls.Event is a dummy variable that takes the value 0 for the pre-BRC period [d − 160 = August 1, 2014, d − 1 = March 30, 2015] and 1 for the post-BRC period [d 0 = March 31, 2015, d 170 = December 30, 2015].Treatment is a dummy that takes the value 1 for benchmark-grade swaps (10Y) and 0 otherwise (12Y).Interaction is a dummy variable computed as Event × Treatment.SRVIX is the log return on the Interest Rate Swap Volatility Index.TYVIX is the log return on the 10-year U.S. Treasury Note Volatility Index.MESS_10Y is the log daily count of the number of messages received by the platform operator for the 10Y IRS contract.MESS_12Y:10Y is the log ratio of messages for the 12Y contract relative to the 10Y contract.TRANS_10Y is the log daily number of transactions in the 10Y IRS contract.TRANS_12Y:10Y is the log ratio of the number of transactions in the 12Y contract relative to the 10Y contract.PARTICIPANTS represents the log number of USD streamers per trading day.MACRO is a dummy variable that takes the value 1 on days with macroeconomic announcements by the FOMC and the Governing Council of the ECB and 0 otherwise.O:I_10Y is the log ratio of outright to implied messages in the 10Y IRS contract.All depth measures are converted to millions (m).The models are estimated using tenor fixed effects.We use Driscoll and Kraay (1998)   TYVIX is the log return on the 10-year U.S. Treasury Note Volatility Index.MESS_10Y and MESS_13Y are the log daily count of the number of messages received by the platform operator for the 10Y and 13Y IRS contract respectively.MESS_12Y:10Y and MESS_13Y:2Yis the log ratio of messages for the 12Y or 13Y contract relative to the 10Y or 2Y contract.TRANS_2Y, TRANS_10Y, TRANS_11Y, TRANS_13Y, and TRANS_30Yare the log daily number of transactions in the 2Y, 10Y, 11Y, 13Y, and 30Y IRS contracts respectively.TRANS_12Y:10Y and TRANS_11Y:5Y are the log ratios of the number of transactions in the 12Y contract relative to the 10Y contract and the 11Y contract relative to the 5Y contract respectively.PAR-TICIPANTS represents the log number of USD streamers per trading day.MACRO is a dummy variable that takes the value 1 on days with macroeconomic announcements by the Federal Open Market Committee (FOMC) and the Governing Council of the ECB and 0 otherwise.O:I_2Y, O:I_10Y, O: I_11Y, and O:I_30Y are the log ratios of outright to implied messages in the 2Y, 10Y, 11Y, and 30Y IRS contracts respectively.All spread measures are expressed in bps (1 bps = 0.01%).The models are estimated using tenor fixed effects.We use Driscoll and Kraay (1998)

Fig. 1 .
Fig. 1.Timeline of events.Panel A shows the polling and publication times under the old ISDAFIX regime.Panel B shows the timeline of events of the BRC.Our sample period starts on August 1, 2014 and ends on December 30, 2015.On March 31, 2015 (⬧) ICE Benchmark Administration successfully transitioned to the new assessment methodology.The FCA regulatory regime for the ICE Swap Rate started on April 1, 2015 (•).Panel C shows the assessment and publication times under the new ICE Swap Rate regime.

Fig. 2 .
Fig. 2. IDB market share by volume traded.This figure shows the market share across the inter-dealer broker (IDB) segment of the USD swap market from 2014 to 2015.The market share for the Trad-X SEF by Tradition comprises all USD swap tenors.The other IDB venues participating in the USD ICE Swap Rate assessment are: BGC and ICAP (here labeled as IGDL -ICAP Global Derivatives Limited).Our period of analysis ranges from August 1, 2014 to December 30, 2015.Source: SEFView, Clarus Financial Technology (https://sefview.clarusft.com/).
Fig.2.Across the three benchmark-participating platforms, the single 10Y maturity, which is the main focus of our analysis, accounts for a market share of approximately 50%.20  In addition, depending on the swap tenor, FCA supervisory intelligence suggests that, for the assessment of USD rates in the period of our analysis, around 50%-75% of tradable quotes contributing to the ICE Swap Rate assessment originated from the Trad-X LOB.21All usual order book variables and USD tenors, ranging from 1 to 50 years, are recorded in the data.The sample period extends from August 1, 2014, when IBA took over the benchmark assessment, to December 30, 2015-a total of 331 trading days.22We employ an event study methodology, where March 31, 2015, the effective date of the new benchmark regime, is the event day, d 0 .The ISDAFIX regime, referred to as pre-BRC, encompasses 160 trading days [d − 160 = August 1, 2014, d − 1 = March 30, 2015].The ICE Swap Rate regime, referred to as post-BRC, extends over 171 trading days [d 0 = March 31, 2015, d 170 = December 30, 2015].In the data, the 10Y USD swap is the most actively traded on Trad-X, and therefore the target of our analysis.We reconstruct the aggregated 10-level full order book at the end of each second, t, during the normal trading hours of the major U.S. exchanges, from 9:30 a.m. to 4:00 p.m. Eastern Time (ET).Messages consist of three action types-new order submissions, order changes, and order cancellations-and are timestamped in Greenwich Mean Time (GMT) to the nearest millisecond (ms).Given the USD emphasis, we convert all our time references to local ET.Each message is labeled with a unique order identifier, allowing us to follow its life cycle.A message cancellation is recorded following an active cancellation or after a transaction has been concluded.All messages are indexed by a sequence number, providing an audit trail of unfolding events.We do not observe trader identifiers associated with the quoting and trading activity.Firm and executable quotes, both outright and implied, are recorded and contribute to the ICE Swap Rate assessment.The division between outright and implied orders is commonly employed in the swap market.An outright order is a direct price submission by a trader, for instance in an individual swap contract.An implied order is generated from the price differential between two existing contracts or spreads.As these are executable, implied orders are crucial for the functioning of the swap market, but do not exist in other markets, such as equities or fixed income that are more commonly studied in the literature.There are two types of implied quotes: implied ins and implied outs.Implied ins are generated by the price differential between two contracts: the difference in prices between two legs (e.g., the 5Y and 10Y swaps) generates a price for the spread.Implied outs generate the price of one leg, on the basis of the price of another leg and a spread.As an example, two orders in the 5Y and 10Y tenors can generate an implied in order in the spread contract between them.Assume a bid in the 5Y tenor at 100 and an ask in the 10Y tenor a 105; a trader could sell the 5Y swap at 100 and buy the 10Y swap at 105, meaning that there is an implied bid in the 5Y-10Y spread contract at − 5.If a trader hits the − 5 bid in the spread contract, the matching engine would fill three separate orders.The sale of the 5Y-10Y spread at − 5 for the trader that hit the bid, the bid at 100 in the 5Y tenor and the ask at 105 in the 10Y tenor for the two traders that were resting on the order books in the 5Y and 10Y tenors.If, rather than knowing the bid and ask prices in the 5Y and 10Y tenors, we have information on the spread and one of the legs, we can calculate an executable implied out message for the missing leg using a similar logic.Traders benefit greatly from these executable implied quotes as they enhance liquidity across the various tenors and spread contracts.The stylized example above only calculates one price (either the spread or the missing leg) on the basis of two others, but similar implied quotes can be calculated for all the tenors and spreads in the market.One can also calculate executable implied quotes from other implied quotes generating second or third generation implied quotes.The calculation of the implied orders therefore links the entire swap curve via spreads and butterflies, allowing traders to transact more cheaply.Calculating these implied quotes on a continuous basis and for all available contracts is computationally complex.However, the Trad-X platform includes an implied engine, explaining the large number of executable implied orders along the swap curve.23   Voice trading and RFQ data are not included in our sample due to issues with timestamps that arise by trying to merge highfrequency LOB data with lower-frequency voice-managed orders and transactions.Moreover, upon receiving an RFQ, the SEF must provide the requester with both the quotes received from responding dealers and the firm resting bid and offer prices on the order book [see(Benos et al., 2020) for more details].The requester decides which quote to accept, consequently LOB and RFQ prices are expected to accurately reflect market conditions.24In addition to the three IDB platforms from which the data for the benchmark assessment are sourced, there are other dealer-to-client venues, such as Bloomberg and Tradeweb.However, according to IBA, the reason that they decided to source data from IDBs was based on the fact that prices on these platforms are firm, while some dealer-to-client platforms operate last-look functionalities and, for this reason, their prices are not considered to be firm.Hence, the dataset obtained from Tradition, the IDB market leader, and consisting of electronically traded swaps, is representative of the market based on which the

Fig. 3 .
Fig. 3. Intraday liquidity.Panel A shows the development of the average total number of messages and Panel B shows the sum of the number of transactions for the 10Y USD IRS during each hour over the course of the trading day.Sample period for the dotted pre-period line is the ISDAFIX regime [d − 160 = August 1, 2014, d − 1 = March 30, 2015].Sample period for the solide post-period line is the ICE Swap Rate regime [d 0 = March 31, 2015, d 170 = December 30, 2015].Times depicted are interval start times and show the corresponding average value for the total interval.Timestamps are in ET.

Fig. 4 .
Fig. 4. 10Y USD swap price development.This figure shows the mid-price development of the 10Y USD IRS over the full sample period from August 1, 2014 to December 30, 2015.The shaded area marks the period of the new benchmark regime from March 31, 2015 to December 30, 2015.

Fig. 5 .
Fig. 5. USD participants.This figure shows the development of the daily count of USD streamers on the Trad-X platform over the sample period.The numbers are normalized and presented in percentage terms (%).The solid black line depicts the long-term average of the time series.The vertical dotted line marks the event date [d 0 = March 31, 2015].Pre-BRC refers to the ISDAFIX regime [d − 160 = August 1, 2014, d − 1 = March 30, 2015].Post-BRC refers to the ICE Swap Rate regime [d 0 = March 31, 2015, d 170 = December 30, 2015].

Fig. 6 .
Fig. 6.Structural breaks.This figure shows the development of the TWQS, TWFS, TWQD, and TWQD10 for the 10Y USD IRS over the sample period.The dashed lines mark the break dates as determined by the BP model.The solid line depicts the long-term average of the time series, while the dot-dashed line shows the segment averages.The dotted line marks the event date [d 0 = March 31, 2015].Pre-BRC refers to the ISDAFIX regime [d − 160 = August 1, 2014, d − 1 = March 30, 2015].Post-BRC refers to the ICE Swap Rate regime [d 0 = March 31, 2015, d 170 = December 30, 2015].All spread values are expressed in bps (1 bps = 0.01%) and all depth values in dollars ($).
M.Aquilina et al.
) and described in Subsection 4.2.We measure the root mean square error (RMSE) and the Mean Absolute Error (MAE) of the regression models Pre-and Post-BRC.Pre-BRC the publication of the benchmark happened at 11:30:00.Post-BRC the publication of the benchmark took place at 11:15:00.Pre-BRC refers to the ISDAFIX regime [d − 160 = August 1, 2014, d − 1 = March 30, 2015].Post-BRC refers to the ICE Swap Rate regime [d 0 = March 31, 2015, d 170 = December 30, 2015].We employ the Diebold and Mariano (1995) (DM) test with a loss function power of two on the residuals of the models and a standard t-test on the squared residuals of the models to test for significance.The t-value is the statistic of a two-sample t-test of μ 1 − μ 2 = 0. *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.%-Diff reports the simple percentage difference between the two periods.All values are expressed in bps (1 bps = 0.01%).

Fig. 7 Fig. 7 .
Fig. 7. Price efficiency around the benchmark assessment.This figure shows the price efficiency of the 10Y USD IRS between 10:58:00 and 11:30:00.Timestamps are in ET.The dotted shows the price efficiency during the ISDAFIX regime [d − 160 = August 1, 2014, d − 1 = March 30, 2015].The solid line shows the price efficiency during the ICE Swap Rate regime [d 0 = March 31, 2015, d 170 = December 30, 2015].The coefficient β measures the signal-to-noise ratio.A coefficient close to one suggests informationally efficient prices.A coefficient smaller than one is consistent with noisier prices.A coefficient bigger than one may be driven by stale prices.

( a )
Time-weighted quoted spread trends.(b) Time-weighted fill spread trends.

Fig. B. 2 .
Fig. B.2. Benchmark differential.This figure shows the development of the daily differential between the 10Y benchmark rate and the on-platform mid-price for the 10Y USD IRS using a two-tiered approach (see below).The dashed lines mark the break dates as determined by the BP model.The solid line depicts the long-term average of the time series, while the dot-dashed line shows the segment averages.The dotted dotted line marks the event date [d 0 = March 31, 2015].Pre-BRC refers to the ISDAFIX regime [d − 160 = August 1, 2014, d − 1 = March 30, 2015].Post-BRC refers to the ICE Swap Rate regime [d 0 = March 31, 2015, d 170 = December 30, 2015].For the ISDAFIX period, the differential is calculated based on the benchmark rate and the point observation of the quoted mid-price at 11:00 a.m.ET.For the ICE Swap Rate period, the differential is computed based on the benchmark rate and the average quoted mid-price during the 2-min benchmark assessment.All values are expressed in bps (1 bps = 0.01%).

Figure B. 2
Figure B.2 illustrates the outcome of the BP multiple structural break test on the time series of the benchmark differential.The BP model establishes that breaks occur on December 1, 2014 and March 25, 2015.On 1 December 2014, the FCA published the Consultation Paper CP14/32, discussing the inclusion of additional benchmarks in the regulatory and supervisory regime.The break on March 25, 2015 arises four trading days before the effective date of the BRC.The benchmark differential dropped on this date and settled at a significantly lower level thereafter.It should be noted that, during the four days from March 25 to March 31, the benchmark rate was still relying on the panel-based assessment methodology.This finding suggests that a change in submission behavior might have occurred slightly before the introduction of the market-based benchmark assessment.Panel banks potentially geared the submitted rates more strongly towards the price quoted on regulated trading venues.

Fig. B. 3 .
Fig. B.3.Robustness test: identification of structural breaks.This figure shows the development of the TWQS for the 10Y USD IRS over the sample period.We use a trimmed time series in order to exclude extreme days such as macroeconomic outliers.The dashed lines mark the break dates as determined by the BP model.The solid line depicts the long-term average of the time series, while the dot-dashed line shows the segment averages.The dotted dotted line marks the event date [d 0 = March 31, 2015].Pre-BRC refers to the ISDAFIX regime [d − 160 = August 1, 2014, d − 1 = March 30, 2015].Post-BRC refers to the ICE Swap Rate regime [d 0 = March 31, 2015, d 170 = December 30, 2015].All values are expressed in bps (1 bps = 0.01%).

Table 1
Fixed-for-floating interest rate swaps.This table shows the tenors, which are captured by the MAT mandate, and those for which IBA is assessing the ICE Swap Rate benchmark.The USD MAT swaps relevant for our study have a 3month LIBOR interest rate basis, a semi-annual payment frequency, and a day count convention of 30/360, aligning with the characteristics of swaps feeding into the assessment by IBA.The MAT mandate for USD tenors was implemented in February 2014.Under the ICE Swap Rate regime, no benchmark rate is assessed for the 12Y USD tenor, which is relevant for later parts of this study.See http://www.cftc.gov/idc/groups/public/@otherif/documents/file/swapsmadeavailablechart.pdf and https://www.theice.com/iba/ice-swap-ratefor more information.

Table 2
Summary statistics: messages.This table reports simple descriptive statistics on electronic trading of the 10Y USD IRS on the Trad-X SEF.n D reports a count of the number of trading days.μ and σ report the arithmetic mean and standard deviation of the mid-price and quote size for orders at the best bid and offer respectively.n reports the average daily count of the total number of messages, new quote submissions, cancellations, changes, outright messages, and implied messages respectively.k and m refer to thousand and million respectively.Pre-BRC refers to the ISDAFIX regime [d − 160 = August 1, 2014, d − 1 = March 30, 2015].Post-BRC refers to the ICE Swap Rate regime [d 0 = March 31, 2015, d 170 = December 30, 2015].%-Diff reports the simple percentage difference between the two periods.

Table 3
Summary statistics: transactions.This table reports descriptive statistics on transactions that were executed electronically on the Trad-X platform.n TRANS reports the number of transactions.Vol TRANS reports the transaction volume.k, m, and b refer to thousand, million, and billion respectively.Pre-BRC refers to the ISDAFIX regime [d − 160 = August 1, 2014, d − 1 = March 30, 2015].Post-BRC refers to the ICE Swap Rate regime [d 0 = March 31, 2015, d 170 = December 30, 2015].%-Diff reports the simple percentage difference between the two periods.

Table 5
Difference-in-differences panel regressions for spread measures.This table reports the results of the DiD panel regression model specified in equation (9) using TWQS and TWFS as dependent variables.(1)presents the DiD model without controls while (2) presents the same specification with controls.Event is a dummy variable that takes the value 0 for the pre-BRC period [d − 160 = August 1, 2014, d − 1 = March 30, 2015] and 1 for the post-BRC period [d 0 = March 31, 2015, d 170 = December 30, 2015].Treatment is a dummy that takes the value 1 for benchmark-grade swaps (10Y) and 0 otherwise (12Y).Interaction is a dummy variable computed as Event × Treatment.SRVIX is the log return on the Interest Rate Swap Volatility Index.

Table 6
Difference-in-differences panel regressions for relative spread measures.This table reports the results of the DiD panel regression model specified in equation (9) using the relative spread measures TWRQS and TWRFS as dependent variables.(1) presents the DiD model without controls while (2) presents the same specification with controls.Event is a dummy variable that takes the value 0 for the pre-BRC period [d − 160 = August 1, 2014, d − 1 = March 30, 2015] and 1 for the post-BRC period [d 0 = March 31, 2015, d 170 = December 30, 2015].Treatment is a dummy that takes the value 1 for benchmark-grade swaps (10Y) and 0 otherwise (12Y).Interaction is a dummy variable computed as Event × Treatment.SRVIX is the log return on the Interest Rate Swap Volatility Index.TYVIX is the log return on the 10-year U.S. Treasury Note Volatility Index.MESS_10Y is the log daily count of the number of messages received by the platform operator for the 10Y IRS contract.MESS_12Y:10Y is the log ratio of messages for the 12Y contract relative to the 10Y contract.TRANS_10Y is the log daily number of transactions in the 10Y IRS contract.TRANS_12Y:10Y is the log ratio of the number of transactions in the 12Y contract relative to the 10Y contract.PARTICIPANTS represents the log number of USD streamers per trading day.MACRO is a dummy variable that takes the value 1 on days with macroeconomic announcements by the Federal Open Market Committee (FOMC) and the Governing Council of the ECB and 0 otherwise.O:I_10Y is the log ratio of outright to implied messages in the 10Y IRS contract.All spread measures are expressed in bps (1 bps = 0.01%).The models are estimated using tenor fixed effects.We use Driscoll and Kraay (1998) consistent standard errors.Robust t-statistics are shown below the estimates.*, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.Sample period is 08/01/2014-12/30/2015.

Table B 4
consistent standard errors.Robust t-statistics are shown below the estimates.*, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.Sample period is 08/01/2014-12/30/2015.Difference-in-differences panel regressions for depth measures with multiple tenors.This table reports the results of the DiD panel regression model specified in equation (9) using the depth measures TWQD and TWQD10 as dependent variables.(1) presents the DiD model without controls while (2) presents the same specification with controls.Event is a dummy variable that takes the value 0 for the pre-BRC period [d − 160 = August 1, 2014, d − 1 = Event × Treatment.SRVIX is the log return on the Interest Rate Swap Volatility Index.TYVIX is the log return on the 10-year U.S. Treasury Note Volatility Index.MESS_10Y and MESS_13Y are the log daily count of the number of messages received by the platform operator for the 10Y and 13Y IRS contract respectively.MESS_12Y:10Y and MESS_13Y:2Yis the log ratio of messages for the 12Y or 13Y contract relative to the 10Y or 2Y contract.TRANS_2Y, TRANS_10Y, TRANS_11Y, TRANS_13Y, and TRANS_30Yare the log daily number of transactions in the 2Y, 10Y, 11Y, 13Y, and 30Y IRS contracts respectively.TRANS_12Y:10Y and TRANS_11Y:5Y are the log ratios of the number of transactions in the 12Y contract relative to the 10Y contract and the 11Y contract relative to the 5Y contract respectively.PARTICIPANTS represents the log number of USD streamers per trading day.MACRO is a dummy variable that takes the value 1 on days with macroeconomic announcements by the Federal Open Market Committee (FOMC) and the Governing Council of the ECB and 0 otherwise.O:I_2Y, O:I_10Y, O:I_11Y, and O:I_30Y are the log ratios of outright to implied messages in the 2Y, 10Y, 11Y, and 30Y IRS contracts respectively.All spread measures are expressed in bps (1 bps = 0.01%).The models are estimated using tenor fixed effects.We use Driscoll and Kraay (1998) consistent standard errors.Robust t-statistics are shown below the estimates.*, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.Sample period is 08/01/2014-12/30/2015.

Table B 6
Difference-in-differences panel regressions for relative depth measures with multiple tenors.This table reports the results of the DiD panel regression model specified in equation (9) using the relative depth measures TWRQD and TWRQD10 as dependent variables.(1)presentstheDiDmodel without controls while (2) presents the same specification with controls.Event is a dummy variable that takes the value 0 for the pre-BRC period [d − 160 = August 1, 2014, d − 1 = March 30, 2015] and 1 for the post-BRC period [d 0= March 31, 2015, d 170 = December 30, 2015].Treatment is a dummy that takes the value 1 for benchmark-grade swaps (2Y, 5Y, 10Y, 30Y) and 0 otherwise (11Y, 12Y, 13Y, 14Y, 16Y, 17Y, 18Y, 19Y, 21Y, 22Y, 23Y, 24Y, 26Y, 27Y, 28Y, 29Y, 50Y).Interaction is a dummy variable computed as Event × Treatment.SRVIX is the log return on the Interest Rate Swap Volatility Index.TYVIX is the log return on the 10-year U.S. Treasury Note Volatility Index.MESS_10Y and MESS_13Y are the log daily count of the number of messages received by the platform operator for the 10Y and 13Y IRS contract respectively.MESS_12Y:10Y and MESS_13Y:2Yis the log ratio of messages for the 12Y or 13Y contract relative to the 10Y or 2Y contract.TRANS_2Y, TRANS_10Y, TRANS_11Y, TRANS_13Y, and TRANS_30Yare the log daily number of transactions in the 2Y, 10Y, 11Y, 13Y, and 30Y IRS contracts respectively.TRANS_12Y:10Y and TRANS_11Y:5Y are the log ratios of the number of transactions in the 12Y contract relative to the 10Y contract and the 11Y contract relative to the 5Y contract respectively.PAR-TICIPANTS represents the log number of USD streamers per trading day.MACRO is a dummy variable that takes the value 1 on days with macroeconomic announcements by the Federal Open Market Committee (FOMC) and the Governing Council of the ECB and 0 otherwise.O:I_2Y, O:I_10Y, O: I_11Y, and O:I_30Y are the log ratios of outright to implied messages in the 2Y, 10Y, 11Y, and 30Y IRS contracts respectively.All spread measures are expressed in bps (1 bps = 0.01%).The models are estimated using tenor fixed effects.We useDriscoll and Kraay (1998)consistent standard errors.Robust t-statistics are shown below the estimates.*, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.Sample period is 08/01/2014-12/30/2015.