Counterparty Credit Limits: The Impact of a Risk-Mitigation Measure on Everyday Trading

ABSTRACT

A counterparty credit limit (CCL) is a limit that is imposed by a financial institution to cap its maximum possible exposure to a specified counterparty. CCLs help institutions to mitigate counterparty credit risk via selective diversification of their exposures. In this paper, we analyse how CCLs impact the prices that institutions pay for their trades during everyday trading. We study a high-quality data set from a large electronic trading platform in the foreign exchange spot market that allows institutions to apply CCLs. We find empirically that CCLs had little impact on the vast majority of trades in this data set. We also study the impact of CCLs using a new model of trading. By simulating our model with different underlying CCL networks, we highlight that CCLs can have a major impact in some situations.

KEYWORDS:

• Counterparty credit limits
• counterparty credit risk
• foreign exchange
• price formation
• market design

Acknowledgments

We thank Franklin Allen, Bruno Biais, Julius Bonart, Jean-Philippe Bouchaud, Yann Braouezec, Damiano Brigo, Rama Cont, Jonathan Donier, Doyne Farmer, Ben Hambly, Charles-Albert Lehalle, Albert Menkveld, Stephen Roberts, Cosma Shalizi, Thaleia Zariphopoulou, and Ilija Zovko for helpful comments. MDG gratefully acknowledges support from the James S. McDonnell Foundation, the Oxford–Man Institute of Quantitative Finance, and the EPSRC (through Industrial CASE Award 08001834).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1. In February 2017, Hotspot FX was acquired by Cboe Global Markets and rebranded as ‘Cboe FX Markets’. At the time of writing, Cboe FX Markets remains a major electronic trading platform in the FX spot market (Cboe Global Markets 2017).

2. We use the same two classes of networks in our simulations of our model of trading in Section 5.

3. Note, however, that the use of CCLs does not exclude the subsequent clearing of trades via a CCP. We return to this discussion in Section 6.

4. The Reuters and Electronic Broking Services (EBS) platforms offer institutions an additional data feed that, in exchange for a fee, provides snapshots of the global LOB at regular time intervals.

5. A price for the currency pair XXX/YYY denotes how many units of the counter currency YYY are exchanged per unit of the base currency XXX.

6. Consequently, the number of seconds between successive samplings is larger in periods that have fewer trades. We repeated all of our calculations by sampling the same series at regularly spaced intervals in calendar time, and we obtain results that are qualitatively similar to those that we obtain with regularly spaced intervals in trade time.

7. When estimating the volatility of a price series, sampling with an interval length of 5 minutes is often regarded as a simple way to reduce the impact of microstructure noise (Hansen and Lunde 2006).

8. A possible refinement of our model is to include a common market factor $W_T$ in addition to the idiosyncratic noise terms. In that case, $\mathrm{d}{M}_t^i=\gamma M_t^i\left({\rho }_i\mathrm{d}{W}_t+\sqrt{1-{\rho }_i^2}\mathrm{d}{W}_t^{M^i}\right)$, where ${\rho }_i$ is a Pearson correlation coefficient. We also performed simulations of this more complicated model, but we found that this additional complication adds little to our model. Therefore, we restrict our discussion to our simpler model.

9. Recall from Section 3 that the bilateral CCL between ${\theta }_i$ and ${\theta }_j$ in real markets is given by $min\left(c_{(i,j)},c_{(j,i)}\right)$. Therefore, although it is necessary to use a directed network to model both $c_{(i,j)}$ and $c_{(j,i)}$ between individual institutions, the network of bilateral CCLs is, by definition, an undirected network.

10. For simplicity, we use an unweighted average. It is possible to extend our model by instead using a weighted average to reflect possible asymmetries between different institutions. For example, one can assign a larger weight to larger institutions.

11. It is possible to avoid this situation by implementing an algorithm to search for the exact time $t$ such that $B_t^i=A_t^j$. However, as we discuss later in this subsection, we choose a sufficiently small value of $\mathrm{\Delta }t$ to ensure that it is rare for multiple trades to occur within a given time step. Therefore, we regard our choice of using fixed-length time steps as a sensible heuristic.

12. We also conducted simulations with several different choices of $N$ between 100 and 1000 (with appropriately modified values of $\mathrm{\Delta }t$). Our results are qualitatively similar in each case.

13. When studying our model, we consider realized volatility that we compute based on event-time sampling. We did not repeat our calculations using calendar-time sampling, because our choice of time scale is arbitrary.

14. See Bech (2012) for estimated transaction volumes for several platforms during this period. Since then, the market share of Hotspot FX (which is now called ‘Cboe FX Markets’) has increased considerably.

15. Because each partial matching of a single market order is subject to the same CCLs, we regard it as inappropriate to study each such partial matching as a separate event, as doing so would produce long sequences of correlated data points from single market orders.