## wpiea2019100

## Source details

**Canonical URL:** [wpiea2019100](https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019100.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2019/wpiea2019100.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2019/wpiea2019100.pdf.json)

---

### 3.1 Transaction Costs — Measurement of transaction costs
- Transaction costs measured by the effective spread (pips):
  - Spreadτ = dτ × (fτ − mτ) × 10^4,(3.1)
    - fτ = contractual forward rate; mτ = contemporaneous mid-price; dτ = 1 for client long positions in EUR/USD and −1 for short positions.
  - Example: client buys euro at 1.0500 while mid-price is 1.0450 → spread = 50 pips.
- Mid-price construction:
  - Indicative dealer quotes from Thomson Reuters Tick History (TRTH) at nine standard maturities: 1 day, 1 week, 2 weeks, 3 weeks, 1 month, 2 months, 3 months, 6 months, and 1 year.
  - Mid-price per maturity computed from best inside quotes of participating dealers; quotes valid for maximum of 30 seconds.
  - Mid-prices for non-standard tenors linearly interpolated across nine standard maturities (e.g., 10-day forward uses weights 3/7 and 4/7 on 1-week and 2-week mid-prices).

### 3.1 Transaction Costs — Explanatory variables used in analysis
- Client Sophistication (five measures; composite):
  - Log#Counterparties: natural logarithm of number of dealers a client trades with during one year.
  - HHI: Herfindahl-Hirschman index based on share of a client’s trades with each dealer.
  - LogTotalNotional: log of total notional (in euros) of all EUR/USD forwards traded by a client in the one year sample period.
  - Log#TradesFX: log of the number of EUR/USD forwards traded by a client in the one year sample period.
  - Log#TradesNonFX: log of one plus total number of a client’s outstanding positions in interest rate, credit, and commodity derivatives at start of sample (April 1, 2016).
  - Sophistication: demeaned first principal component of above five variables.
- RFQ Platform Use:
  - RFQPlatform: dummy = 1 for trades on major multi-dealer RFQ platforms (360t, FXall, Bloomberg, Currenex), 0 otherwise.
- Dealer-Client Relationships:
  - Relationship: dummy = 1 when client has pre-existing credit relationship with dealer (main lending bank identified from Orbis “banker” variable), 0 otherwise.
- Information Rents from Asymmetric Price Adjustment:
  - |∆m−dτ| = |∆mτ| if sign(dτ) ≠ sign(∆mτ), zero otherwise.
  - |∆m+dτ| = |∆mτ| if sign(dτ) = sign(∆mτ), zero otherwise.
  - Hypothesis: coefficient on |∆m−dτ| expected positive; coefficient on |∆m+dτ| expected zero if dealers update quotes quickly.
- Contract Characteristics (controls):
  - Notional (in e mn): notional amount (logs). Expect negative association with spreads.
  - Tenor: original maturity (days). Expect higher spreads for longer maturities.
  - Customization: difference in days between contract tenor and nearest standard tenor. Expect higher spreads for customized contracts.
  - Volatility: realized volatility of FX spot over preceding 30 minutes (one-minute intervals). Expect higher spreads when volatile.
  - Buy: dummy = 1 when client forward-buys euro against dollar.

### 3.1 Transaction Costs — Descriptive statistics (clients and transactions)
- Sample coverage:
  - 10,087 clients trading with 204 dealers in EUR/USD forwards between April 1, 2016 and March 31, 2017.
  - 548,298 EUR/USD forward contracts at transaction level.
- Client-level spreads:
  - AvClientSpread Mean = 18.12; St.Dev 6.5; p10 −2.9; p25 2.1; p50 14.3; p75 33.9; p90 52.4.
  - AvClientSpread average = 18.1 pips, standard deviation = 26.6 pips (text summary).
- Client counterparties and concentration:
  - More than half of clients trade with just one dealer.
  - #Counterparties Mean 1.8; St.Dev 2.0; p10 1; p25 1; p50 1; p75 2; p90 3.
  - HHI Mean = 0.8; St.Dev 0.3.
  - Top 1% clients: between 11 and 29 dealers and HHI < 0.006.
- Client notional and activity:
  - TotalNotional (in e mn) Mean 515; St.Dev 739; p10 60.1; p25 0.4; p50 1.8; p75 11.4; p90 114.
  - Median client trades 8 times; mean trade count = 54; 90th percentile trades 86 times; 25th percentile trades 3 times.
  - More than three quarters of clients never trade any non-FX derivatives.
  - Sophistication Mean 0; St.Dev 1.8; p10 −1.7; p25 −1.2; p50 −0.5; p75 0.7; p90 2.4.
  - Relationship (share of forwards with relationship bank(s)): Observations 6,638; Mean 0.6; St.Dev 0.5.
- Transaction-level spreads and features:
  - Spread (in pips) Mean 6.9; St.Dev 19.4; p10 −4.9; p25 −1.1; p50 2.0; p75 11.3; p90 31.0.
  - Transaction-level average spread = 6.9 pips vs client-level average 18.1 pips.
  - Notional (in e mn) Mean 9.55; St.Dev 53.6; p10 0.0; p25 0.02; p50 0.06; p75 0.2; p90 1.814.
  - Customization Mean 10.6; St.Dev 16.7; p10 1; p25 2; p50 3; p75 12; p90 33.
  - Tenor (days) Mean 69; St.Dev 80; p10 29; p25 35; p50 90; p75 180; p90 188.
  - Volatility Mean 0.007; St.Dev 0.004; p10 0.004; p25 0.004; p50 0.005; p75 0.006; p90 0.008; max 0.01.
  - Buy Mean 0.4; St.Dev 0.5.
  - RFQPlatform Mean 0.4; St.Dev 0.5; RFQ platform use ≈ just under 40% of all trades.
  - |∆m−dτ| and |∆m+dτ| average 0.5 pips over 30 seconds preceding a trade; large mid-price changes rare.
- Additional descriptive findings:
  - Negative correlation between transaction costs and sophistication.
  - Platform use associated with lower spreads.
  - Client industries: majority in trade or production; firms primarily domiciled in export-oriented economies.

### 3.1 Transaction Costs — Econometric specification
- Baseline linear model for 548,298 trades:
  - Spreadi,d,τ = Xi β1 + Zτ β2 + δd + γt + γm + ετ,(5.1)
    - Xi: proxies for client sophistication.
    - Zτ: contract characteristic controls.
    - δd: dealer fixed effects.
    - γt: date fixed effects.
    - γm: minute-of-day fixed effects.
  - Fixed effects compare spreads a dealer charges to different clients controlling for time-varying market conditions and time-invariant dealer characteristics.

### 3.1 Transaction Costs — Main empirical findings
- Client sophistication and spreads:
  - All five individual sophistication measures have expected directional effects and are statistically significant at the 1% level:
    - Log#Counterparties coefficient: −3.872*** (std err 0.225).
    - HHI coefficient: 8.798*** (0.681).
    - LogTotalNotional coefficient: −1.556*** (0.074).
    - Log#TradesFX coefficient: −1.783*** (0.099).
    - Log#TradesNonFX coefficient: −1.011*** (0.105).
  - Composite Sophistication (first PC) coefficient: −1.518*** (0.079).
    - An increase in client sophistication of one standard deviation is associated with a decrease in spreads of 2.7 pips (based on estimated coefficient).
  - Comparison to EURO STOXX 50 benchmark:
    - 38 index members identified in sample coverage (10 index members excluded).
    - Average Sophistication of these firms = 6.65.
    - Median Sophistication in full sample = −0.5.
    - Median firm pays an excess spread of 10.9 pips relative to average EURO STOXX 50 firm (based on estimated coefficients).
  - Aggregate dealer rents from price discrimination:
    - TotalRent = ∑τ β̂1 (Sophisticationτ − 6.65) × Notionalτ,(5.2)
    - For clients with sophistication below 6.65, price discrimination generates aggregate dealer rents from corporate clients of e638 million annually in the EUR/USD cross alone (using β̂1 from Table 2, Column (6)).
- Control variables and other effects:
  - LogNotional: larger notional commands lower spreads (examples: −0.620***; −0.481***; −0.303***; −1.101***; −0.789***; −0.608*** across specifications).
  - LogTenor: longer maturity associated with larger spreads (e.g., 1.144***; 1.193***; 0.947***; 1.142***; 1.224***; 1.089***).
  - Volatility: coefficient positive but statistically insignificant in many specs (date and minute-of-day fixed effects capture much variation).
  - Buy dummy: coefficients around −6.500*** to −6.187*** across columns (indicating dealers demand a premium for providing funding in USD consistent with covered interest parity deviations).
  - Customization: increase in customization by one standard deviation associated with a spread increase of approximately 1 pip.

### 5.2 RFQ Platforms — Mechanism and usage
- RFQ platforms allow clients to query multiple dealers simultaneously, curbing dealers' ability to exert market power.
- Around 39% of all trades are executed through RFQ platforms.
- RFQ trades executed by 1,218 clients (i.e. 12.1% of clients).
- Majority of clients never use an RFQ platform to trade FX forwards.

### 5.2 RFQ Platforms — Empirical findings on spreads and client sophistication
- Platform users obtain lower average spreads than similarly sophisticated firms that trade only bilaterally (Figure evidence, 10,087 clients).
- Negative relationship between transaction costs and client sophistication holds only for non-users; platform users obtain competitive prices irrespective of sophistication.
- OLS regression results (Table 3):
  - RFQPlatform dummy coefficient: −7.290*** (0.472) in Column (1) — platform trading associated with average spread reduction of 7.3 pips.
  - Effect diminishes to −3.815*** (0.433) when controlling for Sophistication (Column (2)), but remains statistically significant (3.8 pips).
  - Interaction Sophistication × RFQPlatform: 1.951*** (0.139) (Column (3)), implying larger platform benefits for less sophisticated firms and that RFQ trading eliminates discriminatory pricing based on Sophistication.
- Client fixed effects (Table 3, Columns (4) and (5)):
  - Attenuation consistent with selection effects, but baseline and interaction effects remain economically and statistically significant.
  - Column (4): platform trading implies average spread reduction ≈ 1.5 pips.
  - Column (5): median firm (Sophistication = −0.5) saves around 4.8 pips when trading through a platform.
  - Average EURO STOXX 50 firm saves around 1 pip (statistically insignificant) when trading through a platform.

### 5.2 RFQ Platforms — Robustness, signaling, and economic magnitude
- Subsample of clients that trade with only one dealer (6,816 clients): RFQ use reduces transaction costs even if client executes all trades with same dealer.
  - Explanation: dealers know requester identity but not number of other dealers simultaneously requested — platform use signals outside trading options.
- Platform-induced spread compression described as "impressive" in economic magnitude.
- Non-anonymity necessary because trades are not centrally cleared and carry counterparty credit risk.
- Despite non-anonymity and potential for discriminatory pricing, RFQ platforms substantially improve execution quality by forcing competition among dealers.

### 5.2 RFQ Platforms — Key quantitative points (preserved exactly)
- 39% of all trades executed through RFQ platforms.
- 1,218 clients (i.e. 12.1%) execute RFQ trades.
- Analysis sample reference: 10,087 clients in Figure 5.
- RFQPlatform dummy average spread reduction: 7.3 pips (Column (1)); 3.8 pips controlling for Sophistication (Column (2)).
- Interaction Sophistication × RFQPlatform: 1.95 (Column (3)).
- Client fixed effects: platform trading ~1.5 pips reduction (Column (4)).
- Median firm Sophistication = −0.5 saves around 4.8 pips (Column (5)).
- Subsample size: 6,816 clients trade with only one dealer.
- Client subset with both relationship and non-relationship trades: 895 clients; average Sophistication = 2.38 (roughly the 90th percentile).

### Policy-relevant implications (synthesized from findings)
- RFQ platforms effectively reduce dealers' market power and eliminate price discrimination based on client Sophistication.
- Broader adoption of RFQ platforms could:
  - Benefit less sophisticated clients the most.
  - Possibly induce additional firms with latent exchange rate exposure to participate in the market.
- Complementary policy: enhanced post-trade transparency could raise client awareness about discriminatory OTC pricing and spur RFQ platform adoption.

*Source: wpiea2019100 - 3.1 Transaction Costs and 5.2 RFQ Platforms (PDF chapter).*

### 3.1  Transaction Costs

### 3.1 Transaction Costs

### Measurement of transaction costs
- Transaction costs are measured by the effective spread (henceforth “spread”), expressed in pips:
  - Spreadτ = dτ × (fτ − mτ) × 10^4,(3.1)
    - where fτ is the contractual forward rate, mτ the contemporaneous mid-price, and dτ is a trade direction indicator (dτ = 1 for client long positions in EUR/USD and dτ = −1 for short positions).
  - Example: if a client buys euro at 1.0500, but the prevailing mid-price is 1.0450, the spread paid by the client is 50 pips.
- Mid-price construction:
  - Indicative dealer quotes from Thomson Reuters Tick History (TRTH) at nine standard maturities: 1 day, 1 week, 2 weeks, 3 weeks, 1 month, 2 months, 3 months, 6 months, and 1 year.
  - Mid-price per maturity is computed from best inside quotes of participating dealers.
  - Quotes assumed valid for a maximum of 30 seconds to avoid staleness.
  - Mid-prices for non-standard tenors are linearly interpolated across the nine standard maturities (e.g., 10-day forward uses weights 3/7 and 4/7 on 1-week and 2-week mid-prices).

### Explanatory variables used in analysis
- Client Sophistication (five measures; also collapsed into a composite):
  - Log#Counterparties: natural logarithm of the number of dealers a client trades with during the one year sample period (captures parameter ρ in Duffie et al. (2005)).
  - HHI: Herfindahl-Hirschman index based on share of a client’s trades with each dealer (inversely related to Log#Counterparties).
  - LogTotalNotional: log of total notional (in euros) of all EUR/USD forwards traded by a client in the one year sample period.
  - Log#TradesFX: log of the number of EUR/USD forwards traded by a client in the one year sample period.
  - Log#TradesNonFX: log of one plus the total number of a client’s outstanding positions in interest rate, credit, and commodity derivatives at the start of the sample period on April 1, 2016.
  - Sophistication: demeaned first principal component of the five variables.
- RFQ Platform Use:
  - RFQPlatform: dummy equal to one for trades on major multi-dealer RFQ platforms (360t, FXall, Bloomberg, Currenex), zero otherwise.
- Dealer-Client Relationships:
  - Relationship: dummy equal to one when the client has a pre-existing credit relationship with the dealer (main lending bank identified from Orbis “banker” variable), zero otherwise.
- Information Rents from Asymmetric Price Adjustment:
  - |∆m−dτ| = |∆mτ| if sign(dτ) ≠ sign(∆mτ), zero otherwise. (Absolute mid-market forward rate change over preceding 30 seconds when price change was in the opposite direction as the client order.)
  - |∆m+dτ| = |∆mτ| if sign(dτ) = sign(∆mτ), zero otherwise. (Absolute mid-market forward rate change over preceding 30 seconds when price change was in the same direction as the client order.)
  - Hypothesis 3: coefficient of |∆m−dτ| on transaction spread expected positive; coefficient of |∆m+dτ| expected zero if dealers update quotes quickly.
- Contract Characteristics (controls):
  - Notional (in e mn): notional amount of the forward contract (in logs). Expect negative association with spreads.
  - Tenor: trade’s original maturity (in days). Expect higher spreads for longer maturities.
  - Customization: difference in days between contract tenor and nearest standard tenor (0, 1, 7, 30, 60, 90, 180, 270, 360 days). Expect higher spreads for customized contracts.
  - Volatility: realized volatility of the FX spot rate over the preceding 30 minutes (one-minute intervals). Expect higher spreads when volatile.
  - Buy: dummy equal to one when a client forward-buys euro against dollar, zero otherwise. May affect spreads if buy/sell order imbalance exists.

### Descriptive statistics (clients and transactions)
- Sample coverage:
  - 10,087 clients trading with 204 dealers in EUR/USD forwards between April 1, 2016 and March 31, 2017.
  - 548,298 EUR/USD forward contracts at transaction level.
- Client-level spreads:
  - AvClientSpread average = 18.1 pips, standard deviation = 26.6 pips.
  - Percentiles: 75th = 33.9 pips; 50th = 14.3 pips; 25th = 2.1 pips.
  - Distribution positively skewed; evidence of substantial price discrimination.
- Client counterparties:
  - More than half of clients trade with just one dealer.
  - 75th percentile of clients has just two counterparties.
  - HHI average = 0.8 (close to perfect concentration).
  - Top 1% clients: between 11 and 29 dealers and HHI of less than 0.006.
- Client notional and activity:
  - Average total notional per client over one year = e515 mn.
  - 10th percentile ≈ e100,000; 90th percentile ≈ e114 mn.
  - Median client trades 8 times; mean trade count = 54.
  - 90th percentile trades 86 times; 25th percentile trades 3 times.
  - More than three quarters of clients never trade any non-FX derivatives.
  - Sophistication: nearly two-thirds of clients display negative value (positive skewness).
  - Relationship dummy: one third of clients never trade with relationship bank(s); just over one half exclusively trade with relationship bank(s).
- Transaction-level spreads and contract features:
  - Average spread over all trades = 6.9 pips (transaction-level), versus 18.1 pips (client-level).
  - Transaction-level 90th percentile spread = 31 pips; client-level 90th percentile = 52 pips.
  - Transaction-level 25th percentile spread = −1.1 pips (negative spreads occur in other OTC markets; may reflect dealer inventory rebalancing).
  - Most contracts have underlying notional < e1 million; just under 10% have notional > e15 million.
  - Half of transactions have original maturity < 35 days.
  - Clients enter long positions in around 40% of trades.
  - RFQ platform use ≈ just under 40% of all trades.
  - |∆m−dτ| and |∆m+dτ| average 0.5 pips over the 30 seconds preceding a trade; large mid-price changes are rare.
- Additional sorts (Online Appendix Table A.1 and A.2):
  - Negative correlation between transaction costs and sophistication.
  - Platform use associated with lower spreads.
  - Client industries: majority involved in trade or production; firms primarily domiciled in export-oriented economies (e.g., Germany).

### Econometric specification
- Baseline linear model for 548,298 trades:
  - Spreadi,d,τ = Xi β1 + Zτ β2 + δd + γt + γm + ετ,(5.1)
    - Xi: proxies for client sophistication.
    - Zτ: vector of control variables describing contract characteristics.
    - δd: dealer fixed effects.
    - γt: date fixed effects.
    - γm: minute-of-day fixed effects.
  - Fixed effects allow comparison of spreads that a dealer charges to different clients, controlling for time-varying market conditions and time-invariant dealer characteristics.

### Main empirical findings
- Client sophistication and spreads:
  - All five individual sophistication measures have expected directional effects and are statistically significant at the 1% level.
    - More counterparties → lower spreads (Column (1)).
    - More concentrated counterparties → higher spreads (Column (2)).
    - More active clients by trades or notional → lower spreads (Columns (3) and (4)).
    - More outstanding derivatives in other asset classes → lower spreads (Column (5)).
  - Composite measure (Sophistication) result:
    - Estimated coefficient = −1.518 (Table 2, Column (6)), statistically significant at 1% level.
    - An increase in client sophistication of one standard deviation is associated with a decrease in spreads of 2.7 pips.
  - Comparison to highly sophisticated benchmark (EURO STOXX 50):
    - 38 index members identified in sample coverage (10 index members are banks/insurance and excluded).
    - Average Sophistication of these firms = 6.65 (above 99th percentile).
    - Median Sophistication in full sample = −0.5.
    - Median firm pays an excess spread of 10.9 pips due to lower sophistication relative to average EURO STOXX 50 firm (based on estimated coefficients).
  - Aggregate dealer rents from price discrimination by sophistication:
    - TotalRent = ∑τ β̂1 (Sophisticationτ − 6.65) × Notionalτ,(5.2)
      - Using β̂1 from Table 2, Column (6) and summing excess spreads relative to EURO STOXX 50 average.
    - For clients with sophistication below 6.65, price discrimination generates aggregate dealer rents from corporate clients of e638 million annually in the EUR/USD cross alone.
- Control variables and other effects:
  - Notional: larger notional commands lower spreads (consistent with corporate bond market evidence); reflects a size discount even after controlling for client sophistication.
  - LogTenor: longer contract maturity associated with larger spreads.
  - Volatility: coefficient positive as expected but statistically insignificant (date and minute-of-day fixed effects capture much time-series variation).
  - Buy dummy: statistically significant coefficient, suggesting dealers demand a premium for providing funding in USD (consistent with covered interest parity deviations).
  - Customization: trades with tenor differing from standard maturity command higher transaction costs; an increase in customization by one standard deviation associated with a spread increase of approximately 1 pip.

*Source: wpiea2019100 - 3.1 Transaction Costs (PDF chapter).*

### 5.2  RFQ Platforms

### 5.2  RFQ Platforms

### Mechanism and usage
- RFQ platforms allow clients to query multiple dealers simultaneously, curbing dealers' ability to exert market power.
- Around 39% of all trades are executed through RFQ platforms.
- RFQ trades are executed by 1,218 clients (i.e. 12.1% of clients).
- The majority of clients therefore never use an RFQ platform to trade FX forwards.

### Empirical findings on spreads and client sophistication
- Figure evidence (10,087 clients): clients that execute at least one trade through a platform obtain lower average spreads than similarly sophisticated firms that trade only bilaterally.
- The negative relationship between transaction costs and client sophistication holds only for non-users; platform users obtain competitive prices irrespective of their level of sophistication.
- OLS regression results (Table 3):
  - RFQPlatformdummy coefficient: platform trading is associated with an average spread reduction of 7.3 pips (Column (1)).
  - The effect diminishes to 3.8 pips when controlling for Sophistication (Column (2)), but remains statistically significant.
  - Interaction Sophistication × RFQPlatform yields a positive coefficient of 1.95 (Column (3)), implying larger platform benefits for less sophisticated firms and that RFQ trading eliminates discriminatory pricing based on Sophistication.
- Client fixed effects (Table 3, Columns (4) and (5)):
  - Some attenuation consistent with selection effects, but baseline and interaction effects remain economically and statistically significant.
  - Column (4): platform trading implies an average spread reduction of approximately 1.5 pips.
  - Column (5): the median firm (with Sophistication = −0.5) saves around 4.8 pips when trading through a platform.
  - The average EURO STOXX 50 firm saves around 1 pip (statistically insignificant) when trading through a platform.

### Robustness and signaling
- Subsample of clients that trade with only one dealer (6,816 clients, Online Appendix): RFQ platform use reduces transaction costs even if a client executes all trades with the same dealer.
- Explanation: while dealers know the identity of the requester (non-anonymity), they do not know the number of other dealers simultaneously requested on the platform; clients can signal outside trading options through the platform, giving platform use signaling power regardless of Sophistication.

### Economic magnitude and market features
- Platform-induced spread compression is described as "impressive" in economic magnitude.
- Non-anonymity of counterparties is necessary in these systems because trades are not centrally cleared and thus carry counterparty credit risk.
- Despite the possibility of discriminatory pricing (because counterparties are non-anonymous), RFQ platforms substantially improve execution quality by forcing competition among dealers.

### Key quantitative points (preserved exactly)
- 39% of all trades executed through RFQ platforms.
- 1,218 clients (i.e. 12.1%) execute RFQ trades.
- Analysis sample reference: 10,087 clients in Figure 5.
- RFQPlatformdummy average spread reduction: 7.3 pips (Column (1)); 3.8 pips controlling for Sophistication (Column (2)).
- Interaction coefficient Sophistication × RFQPlatform: 1.95 (Column (3)).
- Client fixed effects: platform trading ~1.5 pips reduction (Column (4)).
- Median firm Sophistication = −0.5 saves around 4.8 pips (Column (5)).
- Subsample size: 6,816 clients trade with only one dealer.
- Client subset with both relationship and non-relationship trades: 895 clients; average Sophistication = 2.38 (roughly the 90th percentile).

### Policy-relevant implications
- RFQ platforms effectively reduce dealers' market power and eliminate price discrimination based on client Sophistication.
- Because almost 90% of clients (the majority) do not use platforms and more than half of trades are bilateral, broader adoption of RFQ platforms could:
  - Benefit less sophisticated clients the most.
  - Possibly induce additional firms with latent exchange rate exposure to participate in the market.
- Alternative or complementary policy: enhanced post-trade transparency could raise client awareness about discriminatory OTC pricing and spur RFQ platform adoption.

*Source: Excerpt from "5.2  RFQ Platforms" (wpiea2019100).*

### References

### wpiea2019100 - References

### Literature cited
- Empirical and theoretical studies on OTC markets, dealer intermediation, transparency, and liquidity, including:
  - Abad et al. (2016). Shedding light on dark markets: First insights from the new EU-wide OTC derivatives dataset. Occasional Paper 11, European Systemic Risk Board.
  - Amihud & Mendelson (1980). Dealership market: Market-making with inventory. Journal of Financial Economics, 8(1), 31–53.
  - Bessembinder, Maxwell, & Venkataraman (2006). Market transparency, liquidity externalities, and institutional trading costs in corporate bonds. Journal of Financial Economics, 82(2), 251–288.
  - Duffie (2012). Dark markets: Asset pricing and information transmission in over-the-counter markets. Princeton University Press.
  - Duffie, Gârleanu, & Pedersen (2005, 2007). Over-the-counter markets; Valuation in over-the-counter markets. Econometrica, Review of Financial Studies.
  - Hendershott et al. (2017). Relationship trading in OTC markets. Discussion Paper 12472, CEPR.
  - Loon & Zhong (2014, 2016). Impact of central clearing on counterparty risk, liquidity, and trading; Dodd-Frank effects on OTC transaction costs and liquidity. Journal of Financial Economics.
  - Multiple policy progress and review reports: BIS (2016, 2017); Financial Stability Board (2017a, 2017b).

### Summary statistics (Table 1: Client and Transaction Level)
- Panel A: Client Data (10,087 non-financial clients trading at least one EUR/USD forward contract between April 2016 and March 2017)
  - Observations: 10,087
  - AvClientSpread: Mean 18.12; St.Dev 6.5; p10 −2.9; p25 2.1; p50 14.3; p75 33.9; p90 52.4
  - #Counterparties: Mean 1.8; St.Dev 2.0; p10 1; p25 1; p50 1; p75 2; p90 3
  - HHI: Mean 0.8; St.Dev 0.3; p10 0.1; p25 0.6; p50 1; p75 1; p90 1
  - TotalNotional (in e mn): Mean 515; St.Dev 739; p10 60.1; p25 0.4; p50 1.8; p75 11.4; p90 114
  - #TradesFX: Mean 544; St.Dev 1713; p10 8; p25 2; p50 4; p75 8; p90 6
  - #TradesNonFX: Mean 15; St.Dev 23; p10 2; p25 3; p50 0; p75 0; p90 3
  - Sophistication: Mean 0; St.Dev 1.8; p10 −1.7; p25 −1.2; p50 −0.5; p75 0.7; p90 2.4
  - Relationship (share of forwards with relationship bank(s)): Observations 6,638; Mean 0.6; St.Dev 0.5; p10 0; p25 0; p50 1; p75 1; p90 1
- Panel B: Transaction Data (548,298 EUR/USD individual trades)
  - Observations: 548,298
  - Spread (in pips): Mean 6.9; St.Dev 19.4; p10 −4.9; p25 −1.1; p50 2.0; p75 11.3; p90 31.0
  - Notional (in e mn): Mean 9.55; St.Dev 53.6; p10 0.0; p25 0.02; p50 0.06; p75 0.2; p90 1.814
  - Customization: Mean 10.6; St.Dev 16.7; p10 1; p25 2; p50 3; p75 12; p90 33
  - Tenor (in days): Mean 69; St.Dev 80; p10 29; p25 35; p50 90; p75 180; p90 188
  - Volatility: Mean 0.007; St.Dev 0.004; p10 0.004; p25 0.004; p50 0.005; p75 0.006; p90 0.008; max 0.01
  - Buy (dummy): Mean 0.4; St.Dev 0.5; p10 0; p25 0; p50 0; p75 1; p90 1
  - RFQPlatform (dummy): Mean 0.4; St.Dev 0.5; p10 0; p25 0; p50 0; p75 1; p90 1
  - |∆m−dτ|: Mean 0.5; St.Dev 1.0; p10 0; p25 0; p50 0; p75 11.5
  - |∆m+dτ|: Mean 0.5; St.Dev 0.9; p10 0; p25 0; p50 0; p75 11.5

### Key empirical findings (regression summaries)
- Table 2: Spreads and Client Sophistication (Hypothesis 1)
  - Sophistication measures (separate specifications):
    - Log#Counterparties coefficient: −3.872*** (std err 0.225)
    - HHI coefficient: 8.798*** (0.681)
    - LogTotalNotional coefficient: −1.556*** (0.074)
    - Log#TradesFX coefficient: −1.783*** (0.099)
    - Log#TradesNonFX coefficient: −1.011*** (0.105)
    - Sophistication (first PC) coefficient: −1.518*** (0.079)
  - Contract controls (example coefficients across columns):
    - LogNotional: −0.620***; −0.481***; −0.303***; −1.101***; −0.789***; −0.608*** (std errs reported)
    - LogTenor: 1.144***; 1.193***; 0.947***; 1.142***; 1.224***; 1.089***
    - LogCustomization: 0.991***; 1.168***; 0.900***; 0.893***; 1.048***; 0.965***
    - Buy: −6.500***; −6.764***; −6.187***; −6.388***; −6.644***; −6.393***
  - R-squared range: 0.260 to 0.289; Observations: 548,298; Dealer, date, and minute-of-day fixed effects included.

- Table 3: Spreads and RFQ Platform Use (Hypothesis 2)
  - RFQPlatform main effects:
    - Coefficients in various specs: −7.290*** (0.472); −3.815*** (0.433); −13.11*** (0.634); −1.475*** (0.272); −4.530*** (0.923)
  - Sophistication effects: −1.202*** (0.089); −1.926*** (0.080)
  - Interaction: RFQPlatform × Sophistication: 1.951*** (0.139); 0.505*** (0.130)
  - R-squared up to 0.513 in client-FE specifications; Observations up to 548,298 (546,796 where client FE included).

- Table 4: Spreads and Dealer-Client Relationships (Hypothesis 3)
  - Relationship main effects:
    - Relationship coefficients: 2.939*** (0.656); 0.700 (0.606); 3.594*** (0.821); 3.439*** (0.805); 2.186*** (0.829)
  - Sophistication coefficient examples: −1.730***; −1.340***; −0.896***; −1.764***
  - Interaction Relationship × Sophistication: −1.097*** (0.204); −1.070*** (0.201); −0.666*** (0.243)
  - RFQPlatform and interactions included in later columns:
    - RFQPlatform: −4.925*** (0.514); −15.000*** (1.019)
    - Relationship × RFQPlatform: 0.473 (1.025)
    - Sophistication × RFQPlatform: 2.253*** (0.185)
  - R-squared range: 0.285 to 0.328; Observations: 278,491; dealer, date, minute FE and contract controls included.

- Table 5: Information Rents from Asymmetric Price Adjustment (Hypothesis 4)
  - Price-staleness measures:
    - |∆m−dτ| coefficients: 0.409*** (0.050); 0.410*** (0.054); 0.658*** (0.075); 0.660*** (0.075); 0.647*** (0.074)
    - |∆m+dτ| coefficients: −0.243*** (0.052); −0.236*** (0.052); −0.100 (0.083); −0.093 (0.083); −0.142* (0.084)
  - Interactions with Sophistication:
    - |∆m−dτ| × Sophistication: −0.062*** (0.016); −0.065*** (0.015); −0.012 (0.016)
    - |∆m+dτ| × Sophistication: −0.035** (0.016); −0.037** (0.015); −0.051** (0.021)
  - RFQPlatform and interactions:
    - RFQPlatform: −3.810*** (0.434); −12.960*** (0.623)
    - |∆m−dτ| × RFQPlatform: −0.530*** (0.084)
    - |∆m+dτ| × RFQPlatform: 0.246** (0.104)
    - Sophistication × RFQPlatform: 1.952*** (0.139)
  - R-squared range: 0.246 to 0.300; Observations: 548,298; dealer, date, minute FE and contract controls included.

### Figures and notes
- Figure 1 (single-day illustration, 28 December 2016):
  - Plots contractual 30-day EUR/USD forward rates versus intraday mid-price constructed from Thomson Reuters interdealer quote data.
  - Mid-price shown as solid black line; contracts with original maturity between 25 and 35 days depicted.
  - Client long positions: blue dots; client short positions: red crosses. Blue dots above (or red crosses below) the mid-price imply a positive spread.
- Figure 2:
  - Plots cross-sectional distribution of average client spreads based on 548,298 EUR/USD forward transactions between 10,087 clients and 204 dealers; sample period April 1, 2016 to March 31, 2017.

*Italic: Source: References and tables extracted from the content unit "wpiea2019100 - References."*

### 2017.  Positive spreads are costly to the client and advantageous to the dealer.

### 2017.  Positive spreads are costly to the client and advantageous to the dealer.

### Trade distribution by tenor
- Sample: 548,298 EUR/USD forwards traded between dealers and clients over April 1, 2016 to March 31, 2017.
- Forward contract tenor (in days) distribution plotted across 0–360 days, distinguishing:
  - Standard tenors: 7, 14, 21, 30, 60, 90, 180, and 360 days (blue bars).
  - Non-standard tenors (red bars).
- Number of trades reported in thousands on the vertical axis (scale shown from 0 to 40 thousands).

### Average client spread by number of dealer counterparties (Figure 4)
- Average client spread (in pips) by number of counterparties (labels show percentage of clients in each group; marker size ∝ aggregate notional).
- Key average spreads (in pips) by counterparty count groups (marker values shown):
  - 0 counterparties: 67.57
  - 1: 16.32
  - 2: 6.67
  - 3: 3.61
  - 4: 1.77
  - 5: 0.94
  - 6: 0.63
  - 7: 0.52
  - 8: 0.43
  - 9: 0.38
  - 10: 0.28
  - 11: 0.17
  - 12: 0.10
  - 13: 0.14
  - 14: 0.13
  - 15: 0.07
  - 16: 0.05
  - 17: 0.23
- For readability, the 18 counterparty group aggregates all clients with 18 or more counterparties.

### Average client spread by sophistication and platform use (Figure 5)
- Client sophistication: first principal component of Log#Counterparties, HHI, LogTotalNotional, Log#TradesFX, and Log#TradesNonFX.
- Clients split by RFQ platform use:
  - Platform users (used RFQ at least once) colored red.
  - Non-users colored blue.
- Average client spread (in pips) plotted against Sophistication; Kernel-weighted local polynomial regressions plotted separately:
  - Solid black line: clients that never trade through a platform.
  - Dashed black line: clients that trade through a platform at least once.
- Vertical axis truncated at −10 pips for readability.

### Client and transaction characteristics by sophistication and platform use (Table A.1)
- Client-level (Panel A) — mean values:
  - AvClientSpread: Low soph. RFQ user 4.2; Low soph. non-user 5.6; Diff 1.4***.
  - #Counterparties: Low 1.0 vs 1.0; Diff 0.0.
  - HHI: Low 1.0 vs 1.0; Diff 0.0.
  - TotalNotional (in e mn): Low 1.9 vs 0.5; Diff 1.4***.
  - #TradesFX: Low 1.9 vs 2.7; Diff 0.8***.
  - #TradesNonFX: Low 0.11 vs 0.09; Diff 0.02.
  - Sophistication: Low −1.44 vs −1.48; Diff 0.04**.
  - Observations: 613,301 (Low), 1313 (Medium), 13,231 (High), 1,026, 2,337 (panel totals shown across columns).
- Transaction-level (Panel B) — mean values:
  - Spread: Low 4.0 vs 27.8; Diff 23.8*** in first column grouping.
  - Notional (in e mn): Low 1.0 vs 0.2; Diff 0.8***.
  - Tenor: 559 vs 136; Diff 423***.
  - Customization: 8.1 vs 13.8; Diff 5.7***.
  - Volatility: 0.0070 vs 0.0070; Diff 0.0000.
  - Buy: 0.5 vs 0.3; Diff 0.2***.
  - Observations (Panel B): 1179,029; 1,344; 54,411; 309,526; 173,871 (as reported across columns).
- Notes:
  - Clients sorted into low, medium, high sophistication (bottom/middle/top third).
  - RFQ platform user = used an RFQ platform at least once in sample.
  - AvClientSpread = average spread client pays on trades with dealers.
  - HHI = Herfindahl-Hirschman index of counterparty concentration.
  - TotalNotional (in e mn) is total notional traded by client during sample period.
  - #TradesFX = number of EUR/USD forwards traded by client.
  - #TradesNonFX = total number of client's outstanding interest rate, credit and commodity derivatives positions at sample start.
  - Sophistication = first principal component of Log#Counterparties, HHI, LogTotalNotional, Log#TradesFX, Log#TradesNonFX.
  - In Panel B: Spread = difference (in pips) between contractual forward rate and mid-price; Tenor = original maturity (in days); Customization = difference in days between tenor and nearest standard tenor; Volatility = realized volatility of FX spot rate over preceding 30 minutes (one-minute intervals); Buy = dummy = 1 when client forward-buys euro against dollar.

### Clients by location and sector (Table A.2)
- Panel A: Client location (top categories; number of clients, share %, total notional (in e mn), share %, Sophistication (mean), Spread (mean)):
  - Germany: 3,501 | 42.4 | 761,291 | 17.3 | −0.3 | 27.8
  - France: 941 | 11.4 | 999,971 | 22.7 | 0.1 | 8.6
  - Netherlands: 724 | 8.8 | 249,064 | 5.7 | −0.1 | 19.5
  - Spain: 538 | 6.5 | 56,985 | 1.3 | 0.1 | 1.4
  - Italy: 459 | 5.6 | 135,086 | 3.1 | −0.3 | 8.5
  - United States: 321 | 3.9 | 1,127,073 | 25.6 | 1.7 | 3.4
  - Belgium: 318 | 3.8 | 115,415 | 2.6 | −0.1 | 15.5
  - United Kingdom: 275 | 3.3 | 201,877 | 4.6 | 0.6 | 9.6
  - Austria: 158 | 1.9 | 33,821 | 0.8 | −0.1 | 22.1
  - Portugal: 129 | 1.6 | 850 | 0.0 | 0.0 | 13.6
  - All other locations: 899 | 10.9 | 723,763 | 16.4 | 0.4 | 15.1
- Panel B: Client sector (top categories; number of clients, share %, total notional (in e mn), share %, Sophistication (mean), Spread (mean)):
  - Wholesale trade: 3,324 | 40.2 | 196,281 | 4.5 | −0.3 | 21.9
  - Machinery and equipment: 408 | 4.9 | 414,578 | 9.4 | 0.1 | 15.8
  - Retail trade: 328 | 4.0 | 41,992 | 1.0 | −0.3 | 24.8
  - Head offices and consultancy: 317 | 3.8 | 176,961 | 4.0 | 0.6 | 12.0
  - Food products: 289 | 3.5 | 134,440 | 3.1 | 0.4 | 15.3
  - Computers, electronics, optics: 226 | 2.7 | 294,441 | 6.7 | 0.4 | 15.3
  - Financial service activities: 190 | 2.3 | 69,492 | 1.6 | 0.5 | 9.9
  - Metal products, except machinery: 190 | 2.3 | 18,543 | 0.4 | −0.3 | 22.1
  - Chemicals and chemical products: 188 | 2.3 | 246,268 | 5.6 | 0.6 | 15.2
  - Travel agencies: 170 | 2.1 | 11,800 | 0.3 | −0.9 | 32.7
  - All other sectors: 2,633 | 31.9 | 2,800,435 | 63.6 | 0.3 | 14.4
- Note: Categories are top 10 by population plus “other”; clients grouped by parent-level location and two-digit NACE Rev 2 sector.

### Spreads and RFQ platform use (Table A.3, alternative sample)
- Dependent variable: spreads. Sample restricted to 6,816 clients that trade with only one dealer.
- Coefficients (standard errors in parentheses):
  - RFQPlatform: −9.031*** (1.240) in Column (1); −8.261*** (1.284) in Column (2); −9.370*** (1.176) in Column (3); −9.551*** (4.049) in Column (4); −13.320*** (4.208) in Column (5).
  - Sophistication: −4.685*** (0.384) in Column (2); −4.853*** (0.415) in Column (3).
  - RFQPlatform×Sophistication: 1.957*** (1.025) in Column (4); 2.851*** (1.218) in Column (5).
- R-squared values: 0.319, 0.335, 0.335, 0.595, 0.596 across columns (1)–(5).
- Observations: 122,968 (Columns 1–3); 121,637 (Columns 4–5).
- Fixed effects and controls:
  - Dealer FE: Yes.
  - Client FE: added in Columns (4)–(5).
  - Date FE, Minute of day FE: Yes.
  - Contract characteristics controlled (LogNotional, LogTenor, LogCustomization, Volatility, Buy).
- Note: One, two and three asterisks denote statistical significance at 10%, 5% and 1% respectively. Standard errors clustered at client level.

### Spreads and dealer-client relationships (Table A.4, alternative definition)
- Relationship defined as notional traded between a client and dealer relative to the client’s total EUR/USD notional.
- Coefficients (standard errors in parentheses):
  - Relationship: 9.790*** (0.662) in Column (1); 2.724*** (0.884) in Column (2); 8.385*** (0.947) in Column (3); 7.763*** (0.948) in Column (4); 2.081* (1.196) in Column (5).
  - Sophistication: −1.232*** (0.118) in Column (1); −0.463*** (0.129) in Column (2); −0.218* (0.127) in Column (3); −1.485*** (0.150) in Column (4).
  - Relationship×Sophistication: −1.926*** (0.228) in Column (1); −1.942*** (0.233) in Column (2); −0.997*** (0.249) in Column (3).
  - RFQPlatform: −3.732*** (0.422) in Column (4); −14.590*** (1.453) in Column (5).
  - Relationship×RFQPlatform: 4.040** (1.753) in Column (5).
  - Sophistication×RFQPlatform: 1.933*** (0.207) in Column (5).
- R-squared values: 0.273, 0.283, 0.290, 0.295, 0.301 across columns (1)–(5).
- Observations: 548,298 in all columns.
- Controls: Dealer FE, Date FE, Minute of day FE, contract characteristics (LogNotional, LogTenor, LogCustomization, Volatility, Buy).
- Note: One, two and three asterisks denote statistical significance at 10%, 5% and 1% respectively. Standard errors clustered at client level.

### Spreads and client counterparty credit risk (Table A.5)
- Dependent variable: spread. Measures of client risk included (Altman Z-score, CashFlowVol).
- Coefficients (standard errors in parentheses):
  - Sophistication: −1.467*** (0.113) in Column (1); −1.468*** (0.110) in Column (2); −1.560*** (0.110) in Column (3); −1.557*** (0.110) in Column (4).
  - LogTenor: 0.936*** (0.095) in Column (1); −0.0950 (0.166) in Column (2); 0.932*** (0.094) in Column (3); 0.945*** (0.095) in Column (4).
  - ZScore: 0.043 (0.135) in Column (3); −1.521*** (0.285) in Column (4).
  - ZScore×LogTenor: 0.450*** (0.080) in Column (4).
  - CashFlowVol: −0.216 (0.212) in Column (3); −0.644* (0.340) in Column (4).
  - CashFlowVol×LogTenor: 0.151 (0.138) in Column (4).
- R-squared: 0.246, 0.250, 0.255, 0.256 across Columns (1)–(4).
- Observations: 331,388 (Columns 1–2); 359,443 (Columns 3–4).
- Controls: Dealer FE, Date FE, Minute of day FE, contract characteristics.
- Notes:
  - ZScore = client’s modified Altman Z-score (linear combination of working capital, retained earnings, profits, and sales).
  - CashFlowVol = standardized coefficient of variation of client cash flows.
  - LogTenor = natural logarithm of contract original maturity (in days).
  - One, two and three asterisks denote statistical significance at 10%, 5% and 1% respectively. Standard errors clustered at client level.

*Source: wpiea2019100 - 2017.  Positive spreads are costly to the client and advantageous to the dealer.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019100.pdf_