Pre-analysis plan: the effect of anti-money laundering regulation

on global payment flows

Matthew Collin∗, Samantha Cook†and Kimmo Soramaki†

October 20, 2016

Abstract

This pre-analysis plan (PAP) describes proposed research using SWIFT MT103 pay-ments data to investigate the impact of anti-money laundering and counter-the-financial-of-terrorism (AML/CFT) risk-rating and regulation on cross-border banking links and trans-actions.

1 Introduction

1.1 Overview

The past fifteen years have seen an unprecedented level of attention on anti-money launderingand countering-the-finance-of-terrorism (AML/CFT) issues by financial regulators, law enforce-ment agencies and international standard-setters. Following the financial crisis and several highprofile cases of sanctions violations, the international banking sector in particular has been underscrutiny. The result has been a resurgence in the number and size of AML-related enforcementactions filed by regulators: in the US alone, the total value of fines levied by regulators exceeded$15 billion in 2014.1

While terrorism and illicit finance are real and present threats, there are growing concernsthat this increased regulatory activity is leading to a chilling effect on cross-border economicactivity as banks limit their exposure to high risk clients or jurisdictions, a process knownas ‘de-risking’ (Collin et al. 2015). Recent surveys of banks suggest that services offered topartner banks to facilitate international payments, known loosely as ‘correspondent banking’relationships, are declining along several corridors (SWIFT 2012; ECB 2014; World Bank 2015b;Erbenova et al. 2016; CPMI 2016), perhaps exclusively in ‘high risk’ jurisdictions (BBA 2014).One of the oft-cited reasons for this decline is a sharp increase in the costs of compliance resultingfrom AML/CFT enforcement. For example, a 2014 survey by the International Chamber ofCommerce of 298 banks in 127 countries found that over 31% claimed to have terminatedrelationships “due to the increasing cost of compliance (including more stringent AML andKYC)” (ICC 2014). Similarly a World Bank survey of large international banks conducted lastyear found that over 60% cited the cost of customer due diligence (CDD) as a driver for thedecline in correspondent banking relationships.

There are also concerns that AML/CFT activity is affecting a number of different typesof cross-border transactions, both directly by increasing the requirements for transactions to

∗Center for Global Development, 54 Wilton Road, London SW1V 1DE, UK; Email: mattcollin@gmail.com†Financial Network Analytics, Ltd.1See Figure 1 for the trend in enforcement actions over the past five years.

1

Figure 1: AML-related enforcement actions by US regulators (2000-2015)

(a) Number of fines

(b) Total value of all fines (USD billions)

Note: data compiled from from ACAMS.org reports on enforcement actions.

be processed or indirectly by creating incentives for banks to withdraw from whole industrieswhich specialise in these transactions. The ICC survey reported that over 40% of respondentsfelt that AML and know-your-customer (KYC) requirements were a significant impedimentto trade finance, with nearly 70% reporting they declined transactions that year (ICC 2014).The British Banker’s Association reporting on global de-risking recorded several case studieswhere banks have lost the ability to process import/export letters of credit (LCs) as a result of

2

losing their correspondent bank account. The global payments business is also under pressure,as reports suggest that a large number of money transfer companies in the US, the UK andAustralia have lost access to banking services as a result of banks’ desire to reduce their exposureto regulatory risk, potentially leading to a reduction to a decrease in formal remittances (BBA2014; Attridge 2015; World Bank 2015a) .

Most of these example should be driving declines in the aggregate flow of cross-border pay-ments, yet to date there have only been a few empirical studies of the effect of AML/CFTenforcement and regulation on these flows. Kudrle (2009) investigate the impact of financial‘greylisting’ by the Financial Action Task Force (FATF) on cross-border bank flows to 38 taxhavens, but found little effect. However, the study only considered a small number of juris-dictions during a period (the early 2000s) when AML regulatory enforcement was limited tothe early 2000s when AML regulatory enforcement was less biting (see Figure 1). In a studyof the global network of customer-to-customer payments over the SWIFT network, Cook andSoramaki (2014) report a decline in connections between countries listed as offshore financialcentres or countries subject to international sanctions. Most recently, Balakina et al. (2016) ex-amined how banking flows (measured using BIS data on total foreign claims) respond to FATFgreylisting, finding no consistent effects. However, their analysis was confined to annual dataon net flows, when in reality concerns over regulatory risk are likely to affect both paymentssent to high-risk locations and payments received from those locations.

1.2 Study and purpose of the pre-analysis plan

To contribute to the meagre evidence base on de-risking, we will use this document to outlinea proposed study of the reduced-form effect of financial ‘greylisting’ and risk-rating on cross-border payments. For this study, we will be using proprietary data provided to us by theglobal banking cooperate SWIFT on the number of monthly customer-to-customer cross-borderpayments made between two countries.

The main purpose of this pre-analysis plan is to transparently commit to publishing theresults of a number of hypotheses we wish to test in a context where our study must be approvedby several stakeholders before being released. The second purpose is to restrict researcherdegrees of freedom by committing to a minimum set of analyses to appear in the final paper(or online appendices). We reserve the right to include new analyses should we find sufficientreason to do so. In the even that new or more complete data are presented to us, we will issuean update of this pre-analysis plan.

At the time of this writing, we have been granted access to SWIFT’s data on cross-borderpayments, but as of the publication of this plan, we have only utilized an anonymized versionof the data (with all country-identifiers scrambled) and have not linked the original data to anyof our treatment measures. Two of the authors have used an earlier version of the data before,but also have not empirically linked it to changes in AML policy. Due to disclosure restrictionswe will not be presenting any of the SWIFT data here, but we describe its structure in Section3.1.

2 Theoretical framework and hypotheses

This study will examine is whether greater regulatory scrutiny (brought on by changes in eitherinternational AML risk-rating or regulatory sanctions) leads to a reduction in either the totalnumber or direction of payment flows to and from affected countries. In this section we sketch

3

out a basic framework for how risk rating and regulatory scrutiny might lead to de-risking and,in turn, lead to a reduction in payment flows.

2.1 How can greater regulatory risk lead to de-risking?

On the surface, banks face a relatively straight forward cost-benefit decision when they process-ing transactions or take on new clients and correspondents. All of these provide some revenuefor the bank, which has to be weighed against the costs associated with the relationship. Clientsand transactions carry with them an element of regulatory risk, which is the risk that a bankwill be sanctioned by a financial regulator. Regulatory scrutiny carries with it the possibilityof large formal fines, but the costs of litigation can also extend well into the future as banksare often forced to undertake costly compliance investments. Reputational damage can also besevere: Karpoff, Lee, and Martin (2008) estimate the costs of lower sales and higher contractingand financing costs to exceed the cost of a regulatory fine by over seven fold. There may evenbe a significant market penalty, as firms fined for financial misconduct tend to perform worsecompared to those not under litigation (ESRB 2015).

These expected costs can be mitigated if the bank invests heavily in AML compliance activ-ities such as customer due diligence (CDD) or enhanced due diligence (EDD).2 However, duediligence work is growing substantially more costly as global demand has increased (KPMG2014). Enhanced due diligence brings even greater costs: in a recent report commissioned bythe UK’s Financial Conduct Authority, a UK-based bank indicated that enhanced due dili-gence checks on a correspondent account could add between £7,000 to £20,000 onto the cost ofa relationship which typically costs no more than £2,400 to maintain (Artingstall et al. 2016).

These two costs are represented in Figure 2, which shows the relationship between expendi-ture on due diligence work, compliance costs and the expected cost of regulatory action. Bankswill choose an optimal amount of due diligence which balances compliance costs and a tolera-ble level of regulatory risk. If the regulatory risk associated with a given client or transactionincreases, then so will the optimal level of due diligence the bank must invest in to offset regu-latory risk. The higher this optimal level, the greater the chance that the client or transactionwill be unprofitable for the bank. As certain classes of clients or regions where clients operateare designated by regulators as being of inherently higher risk of abuse or requiring greaterdegrees of due diligence, banks may err towards ‘de-risking’ by closing accounts and denyingtransactions for these clients.

As discussed in the introduction, there is growing anecdotal and survey evidence that this ishappening (World Bank 2015b; World Bank 2015a) and that it is being driven by a combinationof regulatory-driven cost concerns and profitability (Durner and Shetret 2015).

2.2 How can de-risking lead to lower payment flow volumes?

The next step of the causal chain is the impact of de-risking on cross-border payments. Asdiscussed in the introduction, there is growing evidence that de-risking is affecting a number ofcross-border activities, including remittances, correspondent banking connections, trade financeand humanitarian operations run by NGOs (Collin et al. 2015). Banks are worried aboutboth outgoing and incoming payments; the former might be subsequently laundered or use tofinance overseas terrorism, while the latter might be already-laundered money or an attempt atsanctions-ducking. A decrease in outgoing flows such as remittances or humanitarian aid flows

2Examples of enhanced due diligence provided by FATF (2012) include approval by senior management,examining patterns of transactions, regular updates of the customer/beneficial owner’s data and obtaining moreinformation on the purpose of transactions

4

Figure 2: Banks balance the cost of due diligence and regulatory risk

Due diligencei

COSTi

Regulatory risk

Cost of due diligenceTotal cost

DD∗

C∗

to a given jurisdiction should manifest as a decrease in the aggregate amount of cross-borderpayments being sent there.

Correspondent banking is another animal altogether. The practice is seen as the backboneof the global payment system, as banks rely on these relationships to offer their customerscross-border services (including remittances) (CPMI 2016). If a bank does not have a directcorrespondent relationship with another bank in its target jurisdiction, then it must ‘completethe chain’ by using intermediary banks with which it does have a relationship. If these rela-tionships are severed, banks must rely on more complex, costly and indirect means of offeringcross-border services, which are likely to lead to a reduction in flows. In the short term, servicesmay be disrupted altogether.

To date, there is no public empirical evidence on the causal importance of correspondentbanking connections for cross-border flows. In reality, the empirical relationship between the twois likely to be deeply endogenous, as the creation and destruction correspondent relationships aretypically driven by the same market forces that drive cross-border flows. Recent evidence fromthe Bank of International Settlement’s Committee on Payments and Market Infrastructures(CPMI) suggests that, causality aside, growth in correspondent networks and payment flowsare closely related.

2.3 Hypotheses to be tested

In this paper, we will test three main hypotheses. The first is related to the number of paymentsleaving and entering countries for which the regulatory risk of doing business has gone up. Thesecond two are related to the direction of payments entering and leaving these countries.

Hypothesis 1: There will be a decline of total payments entering and leaving coun-tries who see a rise in their associated AML/CFT regulatory risk.

Given the anecdotal and survey evidence on de-risking to date, our prior is that when therelative AML/CFT regulatory risk of doing business with clients from a given country increases,then banks outside that territory will limit their exposure by severing correspondent bankingconnections, refusing individual payments or closing the accounts of entities that facilitatepayments. The net effect of these practices will be a decline of payments both in and out of

5

affected jurisdictions.This decline will be mitigated by any alternative arrangements that affected countries find,

such as correspondents in third party countries. If there is no affect on aggregate payments toand from affected countries, de-risking might still have occurred.

Hypothesis 2: The decline in payments will be greater between affected countriesand countries with more regulatory scrutiny/stronger AML/CFT institutions.

Banks in jurisdictions with more regulatory scrutiny (such as the United States or other membersof the FATF), are likely to face a higher level of regulatory risk associated with doing businesswith an affected country. These are likely to be the first countries for which de-risking willreduce payments to/from affected countries.

Hypothesis 3: Affected countries will increase their routing of payments to andthrough countries with less regulatory scrutiny/weaker AML/CFT institutions.There will be an increase in the number of payments from affected countries whichare handled by ‘third party’ countries but routed to countries with greater regula-tory scrutiny.

Regardless of de-risking, banks and customers in affected countries (or those wishing to sendpayments to affected countries) will wish to continue making payments in both directions. If noarrangements can be made to continue direct payments to/from their preferred country, theywill rely on other correspondents in third party countries to complete payments. In some cases,payments will be completely diverted to other countries. In others, payments will continueon their path to/from countries with strong AML/CFT institutions (such as the US or otherFATF members), via this new route. This would indicate that de-risking only diverts paymentsto/from affected countries, rather than eliminating them.

3 Data sources and description

3.1 SWIFT data

The data we will be using this analysis comprises monthly counts of cross-border paymentmessages being sent between banks, covering the period January 2004 to June 2014. This datahas been provided by the SWIFT Institute for the purpose of this analysis. The payment datacomprises the aggregate count of SWIFT MT103 messages being sent between a country A toanother country B, for any two countries connected to the SWIFT network. As described inCook and Soramaki (2014), the MT103 data forms a directed network, with a link indicatingthat a bank in country A has sent a message to a bank in country B during that month.

3.1.1 Message types

The MT103 message type, known as a ‘single customer credit transfer’, is used for all customer-to-customer payments across the SWIFT network. If, for example, a Kenyan IT firm wishesto buy computer processers from a firm based in Silicon Valley, it would instructs its bankin Nairobi to transfer the money to the vendor in California. The Nairobi bank would thensent an MT103 message, either directly to the Californian bank or along a chain of intermedi-aries, informing the receiving bank to credit its customer’s account. These messages make upthe plurality of SWIFT message traffic. While the MT103 covers most customer-to-customer

6

payments, it does not capture bank transfers not made on behalf of customers (such as bank-to-bank transfers), transactions related to trade finance, treasury/security market related transfersor travellers checks.

There are two types of MT103 messages. One is a serial payment, which is sent betweenevery bank in a payment chain. If the Nairobi bank needs to send its payment via a correspon-dent bank in London, one serial MT103 message will be sent between Kenya and the UnitedKingdom and a second one will be sent between the United Kingdom and the United States.The other possible type of message is a covered MT103, which is sent directly to the receivinginstitution, but is followed up with a cover message, known as an MT202 COV, which is sentalong the payment chain to manage settlement. In the above example, the Kenyan bank wouldsend an MT103 directly to the American Bank, and would send an MT202 cover message viaits correspondent in London. Figure 4 in the appendix illustrates how payments are structuredunder both systems. If only a single correspondent is involved in the transfer (e.g. the Kenyanbank and the American bank), then a single MT103 message between the two institutions suf-fices. While we are unable to directly discern between these two methods of sending MT103message in the data, it seems that the vast majority of customer-to-customer payments sentover the SWIFT network are sent using the serial method.3

The structure of the data means that we are unable to observe the ultimate destination ofmessages that move between more than two countries. In the above example a Kenyan banksending a serial MT103 message to the USA via the UK, we would observe two links reflectedthe flow of messages, rather than a single link between Kenya and the US. This has implicationsfor our interpretation of the results, which we will discuss below.

3.1.2 SWIFT payment message counts versus total value of messages

At present we only have access to the total number of MT103 messages sent between twocountries, not the actual value of those messages. In most circumstances we would expectAML/CFT regulatory risk to affect both of these outcomes in a similar fashion. For example, ifour Kenyan bank loses access to its correspondent account in the United States and no suitablereplacements are found (or replacements are more expensive), then we would expect either afall in the frequency of payments between Kenya and the United States or in the value of thosepayments, or both.

One potential challenge to this assumption would be if increased regulatory scrutiny or riskcreated an incentive for banks or customers to break up larger payment into smaller ones inorder, for instance, to avoid enhanced forms of due diligence by partner banks. An extremeversion of this is known as ‘smurfing,’ where money launderers intentionally break up cross-border transfers to fall underneath legal limits for compliance checks. This is unlikely to be aproblem for remittance payments, as the actual cross-border bank transfers which are used tosettle remittance payments are highly-aggregated in nature. Smurfing can be a real issue forother types of transactions coming from high risk countries, but because these locations areboth more likely to be the source of illicit money and are less likely to employ due diligencechecks, this is less likely to be a problem with payments being sent to those locations. In theinstances where these kinds of practices are enabled by enhanced due diligence checks, they willlikely bias any estimates of the impact of risk-rating on payment volumes downward.

3The CPMI (2016) report reveals that approximately 40 million MT 103 messages are sent every month,compared to only 5 million MT 202 COV messages. There can be, at the very least, two MT202 COV messagesfor every covered MT103, indicating that there are, at most, 2.5 million MT103 message sent via the coveredmethod in a given month, only 6% of the total.

7

Table 1: Correlates of SWIFT MT103 message counts to be tested

Outcome: Country-level measure Bilateral measure Sources

Trade flows Annual/quarterly nationalimport/export data

Annual/quarterly nationalimport/export data

COMTRADE

Remittances Bilateral remittance matrix Personal remittances re-ceived/sent, based on BOP

World Bank; IMF

ODA Bilateral aid commitments,annual

Annual ODA OECD-DAC,CRS

Economic growth GDP per capita - level andgrowth rate

World Bank

3.2 Data on ‘reduced form’ flows underlying SWIFT measurements

The purpose of the study is to examine the impact of fluctuations in AML regimes on the numberof payment messages sent between countries using the SWIFT network, but the underlyingmotivation is that these payment flows represent valuable transactions which are important toboth sending and receiving countries. An observed reduction in payment messages sent to acountry can’t - in isolation - be considered to be an inherently negative outcome, but it mighthint at one. To better motivate the use of SWIFT messages as an outcome measure, we willshow correlations between MT103 messages and other measures of economic activity at boththe national and bilateral levels. Table 1 shows the various outcomes we will show correlationsfor in the resulting paper.

In addition to calculating these correlations, when there is an overlap between data sourcesand our specification of choice (e.g. trade flows and our specification is annual), we will run thesame specification again, swapping out our MT103 outcome measure with the ‘reduced form’outcome measure discussed here.

There is one caveat to this approach. When the unit of observation is a single country,incoming and outgoing SWIFT messages should partially track these underlying flows. However,when the unit of observation is a bilateral pair, the two measures will diverge even further if alarge share of SWIFT messages are serial payments moving through a third (or fourth or fifth...)country.

3.3 Treatment measures

Our aim is to isolate a time-varying treatment measure that raises the regulatory risk associatedwith a given jurisdiction abruptly. Our main measure is taken from a list of countries withAML/CFT deficiencies which is maintained by the Financial Action Task Force (FATF).

3.4 FATF greylisting

Through a myriad of formal review processes, FATF routinely monitors each country’s com-pliance with its official recommendations on fighting money laundering and terrorist financing.Countries for which compliance with the recommendations is found to be a major issue aremonitored by FATF’s International Co-operation Review Group (ICRG). Three times a year,the ICRG places countries who continue to struggle with compliance on its High-risk and Non-Cooperative Jurisdictions list.4. This list comprises both territories which have AML regimedeficiencies but are making efforts to improve compliance to those who are not making anyattempt to improve their regime. FATF classifies countries on this list by three categories:

4Formerly known as the non-cooperative countries and territories (NCCT) list

8

1. Jurisdictions with an action plan for which FATF calls on its member countries to considerthe information it has presented on AML/CFT deficiencies (56 countries since Feb, 2010).

2. Jurisdictions for which FATF calls on all other countries to consider the risks arising fromAML/CFT deficiencies (22 countries since Feb, 2010).

3. Jurisdictions for which FATF calls on all other countries to apply counter-measures toprotect the international financial system from the money laundering/terrorisft financingrisks (2 countries since Feb, 2010).

While only in the third case does being added to FATF’s list include recommendations forother countries to enact explicit countermeasures, the FATF “grey and black list” is widely seenas being a direct signal that regulators consider a territory to be particularly risky. Regulatorsin turn offer guidance to financial institutions either implicitly or explicitly asking them totake on board FATF’s assessment. For example, following FATF’s publication of its list, theUS regulator FinCEN issues guidance advising banks to consider the risks associated withjurisdictions in category (1), enact enhanced due-diligence on jurisdictions in category (2) and toemploy explicit countermeasures5 to prevent jurisdictions in category (3) from having substantialaccess to the global financial system. Banks interpet signals from FATF and from their ownregulators as a call to perform enhanced due diligence on customers from these jurisdictions orelse face the risk of enforcement actions or fines from regulators (Hodge and Trofimov 2014).

We have gathered data on FATF ‘greylisting’ decisions over the past fifteen years, codingthem according to the severity of the designation.6 FATF makes these announcements threetimes a year in February, June and October. Using FATF announcements we have created aquarterly panel for every jurisdiction indicating its FATF designation during each quarter. Wecan see in Figure 3 that FATF greylisting spiked in the past three years, with over forty countriesbeing listed at one time. This will be our main treatment measure used in the analysis below.Because some FATF greylisting carries with it the requirement of counter-measures, we willpresent results after dropping countries affected by this particular requirement as a robustnesstest.

3.5 Other treatment measures

We will employ at least three other ‘treatment’ measures in this analysis. Two of these willbe used to investigate the extent to which banks respond to changes in actual or perceivedAML/CFT risk as opposed to the regulatory risk introduced by FATF greylisting. The thirdwill be used to bound the effects of extreme restrictions on cross-border transactions as a proof-of-concept test of the outcome measure. As each these outcomes may co-vary with FATFgreylisting, we will also include them as controls as a robustness test. We will use these in thefirst two specifications (described below), but not the third, which is explicitly designed withFATF greylisting in mind.

5While not fully prohibitive, the countermeasures suggested by FATF are extremely restrictive, in line withrestrictions imposed by most financial sanctions.

6This includes jurisdictions which “have strategic AML/CFT deficiencies for which they have developed anaction plan with the FATF,” jurisdictions which have the same deficiencies but have not made progress oragreed on a plan to deal with them and jurisdictions for which the FATF has called on its members to enactcountermeasures to prevent money from being laundered through these locations.

9

Figure 3: FATF grey/blacklisting (2000-2015)

Note: Graph shows the current number of countries publicly listed by the FATF as high risk or

non-cooperative.

3.5.1 FATF Mutual evaluation scores

FATF greylisting has the potential to raise both the regulatory risk and the perceived ML/TFrisk associated with a jurisdiction. In order to investigate FATF’s influence on the latter, wewill investigate the impact of FATF’s period publication of its ‘Mutual-Evaluation Reviews.’These are in-depth studies conducted by regional ‘FATF-style’ bodies which assess exactly howcompliant is with each of its 40 recommendations and 9 special recommendations. These reviewsare conducted every few years and are scheduled years in advance. We will construct a measureof country level compliance which is equal to the average score a country receives on the 49recommendations.7 We will update this score with its new value each time a mutual-evaluationreport is issued for a given jurisdiction.

3.5.2 INCSR ratings and designations

Every year the US State Department produces its International Narcotics Control Strategy Re-port (INCSR). In its second volume, it summarizes actions taken by governments to curb moneylaundering and designates jurisdictions which are“of primary concern” for money laundering.We will construct two indices using this information: the first is a measure of government ‘effort’against money laundering, which will be the percentage of positive actions a country has takenout of the total number of possible actions put forth by the INCSR. We will construct a secondindicator which will be equal to one if the country is listed as a jurisdiction of primary concernand zero otherwise.

7For countries reviewed under the new FATF methodology, we will use the average over both the old set oftechnical compliance indicators and the new set of effectiveness indicators, where available.

10

3.5.3 Economic Sanctions

While changes in FATF or INCSR designations should in theory lead to changes in the waysthat bank assess clients from certain jurisdictions, they do not mandate explicit restrictions onthe flow of money in and out of these locations.8 If we are unable to reject the hypothesis thatFATF greylisting does has no impact on SWIFT messages, it won’t be clear if this is because thetrue effect of greylisting on cross-border transactions is zero or if our outcome measure is unableto pick up the true effect. To account for this, we plan to investigate the impact of regulatorydecisions that place explicit restrictions on cross-border payments, which should show up inSWIFT data. To do this, we will construct an indicator which is equal to one if a country issubject to OFAC, EU or UN Sanctions and zero otherwise.

4 Specification issues and identification

S1: The impact of risk-rating on aggregate MT103 payment counts

To investigate the impact of a given treatment on MT103 messages, we will use the followingspecification:

Mit = θTit + Xitδ + µi + γt + tµit + εit (1)

Where Mit indicates the number of MT103 messages being sent from (or to) country i in agiven period t. We will estimate equation (1) separately for incoming and outgoing messages.Tit is a treatment indicator which is equal to one when a jurisdiction i is identified on a FATFgreylist during time period t.9 θ is our estimate of the impact of the treatment (greylisting,etc) on the count of incoming or outgoing messages. Xit is a vector of time-varying countrycharacteristics (described below). The parameters µi, γt and tµit control for country fixedeffects,10 period fixed effects and country time trends, respectively. In final analysis, we willintroduce these controls iteratively.

We will estimate equation (1) using monthly, quarterly and annual periods.11 Because ofthe count nature of the MT103 message data, we will estimate equation (1) using a Poissonregression. We will also present results using ordinary least-squares (OLS) and a negativebinomial regression.12 For OLS results, we will present the outcome measure in both absolutelevels and in logs. To get a sense of how country i’s connection to the global payment networkchanges after it is added to a FATF greylist, we will also consider the number of countries ireceives (or sends) at least one payment message as an outcome measure Nit.

Approximately one half of all country-specific country series show signs of being non-stationary. Most of these follow a simple linear trend, so we expect this to be accountedby the country trends indicated by tµit. While we commit to reporting the results from the

8The only exception is when FATF indicates a country should be subject to countermeasures, which usuallyhappens concurrent with sanctions.

9For other treatment measures, Tij may be continuous.10Masciandaro (2005) finds that the probability a jurisdiction is reviewed by the FATF depends on a number

of country-level characteristics, ranging from GDP per capita, foreign deposits, and the presence of organizedcrime or terrorist activities.

11Some covariates in Xit will only be observed at quarterly or annual levels.12The SWIFT data shows signs of over-dispersion, making a negative binomial model more appropriate. How-

ever, given the concerns over implementing fixed effects in these models (Allison and Waterman 2002; Guimaraes2008), these results will only be reported in the appendix

11

above specification, we may employ further specifications as robustness checks to account forother time series characteristics of the data.

This specification is aimed as being a direct test of hypothesis (1). Our prior is that θ < 0for incoming and outcoming messages. We summarize the different outcomes/regression speci-fications in Table 3.

S2: The impact of risk-rating on the path of MT103 payment counts

To investigate the impact of a given treatment on the direction of MT103 payments in thenetwork, we will use the following specification:

Mijt = ρTjt + α(Tjt × Ei) + XitδX + YjtδY + ZijtδZ + µij + γt + tµij + εijt (2)

In this specification, Mijt indicates the number of MT103 messages sent from country i tocountry j (or, separately, country j to country i) in a given period t. The treatment indicatorTjt indicates when country j is subject to FATF greylisting (or other treatments). We alsointeract the treatment indicator with Ei, which is a fixed characteristic of country i which weuse to scale the treatment effect. In Subsection 4 below we discuss the characteristics we willconsider for inclusion in Ei to investigate which countries decrease or increase their exposureto the country that was greylisted.

The vectors Xit, Yjt and Zijt contain time-varying characteristics for country i, countryj and pairwise characteristics capturing the relationship between countries i and j. The pa-rameters µij , γt and tµij will respectively capture pairwise fixed effects, period fixed effectsand pairwise time trends. As before, we will estimate specification (2) using Poisson, OLS andnegative binomial regressions at the monthly, quarterly and yearly level.

Testing whether we can reject the null of α = 0 will inform the second hypothesis: whethercountries with more robust AML/CFT regimes will see a greater decline in payments to/fromgreylisted countries. Our prior is that α <= 0 and α + ρ (the combined effect for countriesdescribed by Ei is < 0. The parameter ρ in isolation isolates the change in payments fromcountries with less robust AML/CFT regimes. It is a partial test of our third hypothesis, andwe would expect to see ρ > 0, which would indicate an increase in payments via third partycountries.

The above specification is agnostic towards which countries i are likely to reduce theirpayments to and from country j. However, previous work by Cook and Soramaki (2014) suggeststhat the network of MT103 messages can be characterized by a core-periphery structure, wherecore countries are connected to each other and at least one periphery country and peripherycountries are not connected to each other. These classifications are an artefact of the underlyingstructure of payments: many countries send serial MT103 messages through correspondents in aset of ‘core’ countries. Thus payments from some (usually small) economies often travel throughlarger ‘hubs’ before ending up at another small economy. Because we only observe one edge ofa serial payment, a lot of payments between small economies go unobserved. Unsurprisingly,nearly 60% of all country pairs never send a message to each other across the entire panel.This suggests that effects on bilateral payments between two countries are more likely to beobserved between core-core pairs and core-periphery pairs. Thus, we will present the resultsfrom Equation 2, restricting countries in i to those classified as core countries using the methoddescribed in Craig and Von Peter (2014). We will also display the results using the full sample,for robustness.

12

Finally, similar to the previous specification, we will also consider an indicator = 1 if i hassent j any messages in the given period, as well as a separate indicator for vice versa. Wesummarize the different outcomes/regression specifications in Table 4.

Heterogeneity/diversion across countries

The goal of specification (2) is to determine if certain types of countries (i) are more or lesslikely to send or receive payment messages to greylisted countries (j). We will investigate this byrunning specification (2) multiple times, swapping out different characteristics for parameter Ei.Below we list these characteristics, their sources, and our theoretical justification for includingthem:

S3: The impact of risk-rating on use of third party jurisdictions to completepayments

Where the second specification aims to uncover whether certain ‘protected’ jurisdictions (suchas the US or FATF-members as a whole) reduce their exposure to payments from/to greylistedcountries, the aim of this specification is to investigate whether greylisted countries are divertpayments through ‘third party’ jurisdictions onward to ‘protected’ jurisdictions. This would bedriven by the use of third party correspondents to complete payments. MT103 messages containboth the origin and ultimate beneficiary of a transaction, so there is nothing illicit about sucha diversion. However, it would indicate that greylisted countries have to take more circuitousroutes to reach their destination.

We are largely limited by the data here, as we only observe on leg of an MT103 messages’sjourney (and are also unable to discern whether a given leg is the only leg or not). To accountfor this, we will investigate the correlation between payments between non-greylisted countriesand incoming payments from greylisted countries.

Let i ∈ I index countries which are likely to reduce their direct exposure to greylistedcountries (e.g. the US, FATF-members, countries with strong AML/CFT policies). Let k ∈ Kindex countries which are not in I which are greylisted at some point during the entire timeseries. Finally, let j ∈ J index countries which are neither I nor K and let K = I∪J . Considerthe following equation:

MIjt = πMjKt + β(MjKt × PKt) + ηMjKt + φ(MjKt × PKt) + Xitδ + µi + γt + tµit + εit (3)

In the above equation, MIjt indicates the aggregate number of messages sent to countriesin I that are sent from country j. The variable MjKt indicates the total number of messagessent from countries in K to country j in the same period. Thus, in isolation, π captures thepartial correlation between messages sent to j from ever-greylisted countries and messages sentto countries in I from j. An increase in π would suggest that more MT103 messages are ‘passedon’ as subsequent messages to ‘protected’ jurisdictions. We allow this partial correlation tovary across time by introducing the interaction term MjKt × PKt, where PKt is equal to theproportion of countries in K which are currently greylisted in time t. This allows the partialcorrelation to change as more countries in K are added to the greylist. When β > 0, this wouldindicate that more messages from greylisted countries are being passed on to countries in I viaj.

In addition to the standard set of controls, we also control for incoming messages from allother countries which are not in K (K) and let it vary by the number of countries in K which

13

Tab

le2:

Cou

ntr

yca

tego

riza

tion

for

spec

ifica

tion

(2)

Ei

=M

oti

vati

on

Con

stru

cti

on

Sou

rce

Th

eU

SA

Sou

rce

of

stro

nge

stA

ML

/CF

Tre

gula

tory

over

sight

Ei

=1

ifi

isth

eU

SA

n/a

Mem

ber

sof

FA

TF

(as

of200

4)

Mem

ber

sar

eex

pec

ted

tofo

llow

FA

TF

reco

mm

end

atio

ns;

hav

est

ron

gA

ML

/CF

Tin

stit

uti

ons;

=1

ifFA

TF

mem

ber

in20

04FA

TF

docu

men

tati

on

Cou

ntr

ies

wit

hst

ron

gA

ML

/C

FT

inst

itu

tion

s

Ban

ks

inth

ese

cou

ntr

ies

are

mor

eli

kely

tob

esu

bje

ctto

stri

nge

nt

regu

lati

on,

CD

Dex

pec

atio

ns

=1

ifco

untr

ysc

ored

ator

abov

e75

per

centi

lein

INC

SR

rati

ngs

in200

4IN

CS

Rre

por

ts

Hom

eco

untr

ies

ofb

an

ks

sub

ject

toU

Sre

gu

lati

on

Ban

ks

are

exp

osed

toU

S,

thu

sw

ill

lim

itow

nex

pos

ure

tori

sky

cust

omer

s

=1

ifco

untr

yh

asb

anks

list

edas

hav

ing

abra

nch

bas

edin

the

US

A

Fed

eral

Res

erve

Boar

dli

stof

fore

ign

ban

ks

Cou

ntr

ies

hig

hly

dep

end

ent

on

SW

IFT

acc

ess

toU

S.

Cou

ntr

ies

dep

end

ent

onfl

ows

to/f

rom

US

wil

lw

ant

top

rote

ctth

eli

nk.

=1

if50

%of

tota

lS

WIF

Tp

aym

ents

in/ou

tof

the

cou

ntr

yin

2004

cam

e/w

ent

toth

eU

SA

SW

IFT

data

14

are greylisted. This is to account for increase in flows from other third parties that may resultfrom greylisting (e.g. greylisted Venezuela sends payments to Chile, then Argentina (j) thenthe US).

If β > 0, we take this as evidence that greylisted countries are more likely to send paymentsto ‘protected’ countries via third parties, a sign that greylisting only diverts money throughother channels, and further support of hypothesis 3.

As an alternative specification to 3, we will consider a disaggregated version which takesinto account j and k’s pairing separately, where:

MIjt = πMjkt+β(Mjkt×Tkt)+ηMjKt+φ(MjKt×Tkt)+XjtδX+YktδY +ZjktδZ+µjk+γt+tµjk+εjkt(4)

Where Tkt is an indicator variable equal to one when k is greylisted.For I we will use the following categories: the United States, members of FATF. In addition,

we commit to using any category for which there is significant evidence that α is negative inspecifications (3)-(4). As before we will estimate this using Poisson, OLS and negative binomialat the monthly, quarterly and yearly level.

4.1 Controls

In all three specifications, we will report results will be present both with and without availablecountry controls, as well as country-pair controls for specifications (3) and (2).

4.1.1 Country-level controls

When available for the period chosen, we will include: GDP per capita (World Bank), a country’sdemocracy score (POLITY IV), other indicators of institutional quality (World Bank CPIA)and a dummy variable indicating whether or not there is an ongoing conflict. When we useFATF greylisting as our treatment indicator we will also check, for robustness, whether or notthe results are robust to including the other treatment indicators described in Section 3.5.

4.1.2 Country-pair controls

When available for the period chosen, we will include controls for the prevailing nominal ex-change rate and bilateral migration flows.

4.2 Inference

All country-level regressions (specification (1)) will be clustered at the country level. Allbilateral-pair specifications (2 and 4) will be clustered both at the i country level and at the jcountry level (i.e. two-ways for each of the pairs) following Cameron et al. (2012).

4.3 Heterogeneity across time

As discussed in the introduction, there has been a significant rise regulatory pressure andAML/CFT enforcement in the past few years. Fines have increased drastically in the pastfive years, and what few anecdotes and limited survey evidence exists points to a sharp rise inincidents of de-risking over this time period. It is plausible, then, that increases in regulatoryrisk have only been binding for banks since 2010, when fine levels began to rise sharply.

15

Also, in addition to the fact that several of the country-level series show signs of being non-stationary, others sign of a significant structural break around the time of the 2008 financialcrisis. To both account for this

To investigate this, for all three specifications we will examine heterogeneity in effect sizesacross time, by also presenting separate estimates for the 2004-2009 time period and the 2010-2014 time period. For quarterly specifications, we will divide the series at Q3-2009 and formonthly specifications we will divide the series in July, 2009.

References

Allison, P. D. and R. P. Waterman (2002). Fixed–effects negative binomial regression models.Sociological methodology 32 (1), 247–265.

Artingstall, D., N. Dove, J. Howell, and M. Levi (2016). Drivers and impacts of de-risking. Technical report, Available online at https://www.fca.org.uk/publication/

research/drivers-impacts-of-derisking.pdf.

Attridge, S. (2015). Walking a tightrope: Aml/ctf regulation, financial inclusion and remit-tances. Technical report, Annual Commonwealth & Francophonie Dialogue with the G20Discussion Paper 3.0.

Balakina, O., A. D’Andrea, and D. Masciandaro (2016). Bank secrecy in offshore centres andcapital flows: Does blacklisting matter? BAFFI CAREFIN Centre Research Paper (2016-20).

BBA (2014). De-risking: Global impact and unintended consequences for exclusion and sta-bility. Technical report, British Banker’s Association & The Wolfsberg Group DiscussionPaper Number 1.

Cameron, A. C., J. B. Gelbach, and D. L. Miller (2012). Robust inference with multiwayclustering. Journal of Business & Economic Statistics.

Collin, M., L. De Koker, M. Juden, J. Myers, V. Ramachandran, A. Sharma, and G. Tata(2015). Unintended consequences of anti–money laundering policies for poor countries.

Cook, S. and K. Soramaki (2014). The global network of payment flows. Technical report,SWIFT Institute Working Paper No. 2012-006.

CPMI (2016). Correspondent banking. Technical report, Available online at http://www.

bis.org/cpmi/publ/d147.htm.

Craig, B. and G. Von Peter (2014). Interbank tiering and money center banks. Journal ofFinancial Intermediation 23 (3), 322–347.

Durner, T. and L. Shetret (2015, November). Understanding bank de-risking and its effectson financial inclusion: An exploratory study.

ECB (2014). Ninth survey on correspondent banking in euro. Technical report, EuropeanCentral Bank.

Erbenova, M. M., M. Y. Liu, M. N. Kyriakos-Saad, A. L. Mejia, J. G. Gasha, M. E. Mathias,M. M. Norat, M. F. Fernando, and M. Y. Almeida (2016). The withdrawal of correspondentbanking relationships: a case for policy action. International Monetary Fund.

ESRB (2015). Report on misconduct risk in the banking sector. Technical report, EuropeanSystemic Risk Board.

16

FATF (2012). International standards on combating money laundering and the financing ofterrorism & proliferation. Technical report, The Financial Action Task Force.

Guimaraes, P. (2008). The fixed effects negative binomial model revisited. Economics Let-ters 99 (1), 63–66.

Hodge, N. and Y. Trofimov (June 11, 2014). Afghanistan faces financial blacklisting: Anti-money-laundering law required to avoid possible cutoff from international financial system.Wall Street Journal .

ICC (2014). Rethinking trade & finance: an icc private sector development perspective.Technical report, International Chamber of Commerce.

Karpoff, J. M., D. S. Lee, and G. S. Martin (2008). The cost to firms of cooking the books.Journal of Financial and Quantitative Analysis 43 (03), 581–611.

KPMG (2014). Global anti-money laundering survey 2014. Technical report, Availableonline at https://www.kpmg.com/KY/en/IssuesAndInsights/ArticlesPublications/PublishingImages/global-anti-money-laundering-survey-v3.pdf.

Kudrle, R. T. (2009). Did blacklisting hurt the tax havens? Journal of Money LaunderingControl 12 (1), 33–49.

Masciandaro, D. (2005). False and reluctant friends? national money laundering regulation,international compliance and non-cooperative countries. European Journal of Law andEconomics 20 (1), 17–30.

SWIFT (2012). Correspondent banking 3.0 the compelling need to evolve towards a customer-centric experience banking model. Technical report, SWIFT White paper series.

World Bank (2015a). Report on the G20 survey in de-risking activities in the remittancemarket. Technical report, The World Bank Group.

World Bank (2015b). Withdraw from correspondent banking: where, why, and what to doabout it. Technical report, The World Bank Group.

5 Appendix:

17

Table 3: Specification 1 - results to be reported

Outcome(s): Regression specification(s) Periods

# messages sent

# message received

(i) OLS: Pair FE + Pair trends+ Period Fixed Effects(ii) Poisson: Pair FE + Pair trends+ Period Fixed Effects(iii) Negative Binomial: Pair FE+ Pair trends + Period Fixed Effects

Monthly,Quarterly,Annual

Ln(# messages sent)

Ln(# messages received)

(i) OLS: Pair FE + Pair trends+ Period Fixed Effects

M,Q,A

Number of countries sendinga least one message to i

Number of countries ia least one message to.

(i) OLS: Pair FE + Pair trends,+ Period Fixed Effects(ii) Poisson: Pair FE + Pair trends,+ Period Fixed Effects

M,Q,A

18

Tab

le4:

Sp

ecifi

cati

on2

-re

sult

sto

be

rep

orte

d

Ou

tcom

e(s

):R

egre

ssio

nsp

ecifi

cati

on

(s)

Peri

od

sS

am

ple

#m

essa

ges

sent

fromi

toj

#m

essa

gere

ceiv

edbyi

fromj

(i)

OL

S:

Pai

rF

E+

Pai

rtr

end

s+

Per

iod

Fix

edE

ffec

ts(i

i)P

oiss

on:

Pai

rF

E+

Pai

rtr

end

s+

Per

iod

Fix

edE

ffec

ts(i

ii)

Neg

ativ

eB

inom

ial:

Pai

rF

E+

Pai

rtr

end

s+

Per

iod

Fix

edE

ffec

ts

Mon

thly

,Q

uar

terl

y,A

nnu

al

(A)i

cou

ntr

ies

rest

rict

edto

‘cor

e’co

untr

ies

as

defi

ned

in200

4(B

)al

lco

untr

ies

Ln

(#m

essa

ges

sent)

Ln

(#m

essa

ges

rece

ived

)

(i)

OL

S:

Pai

rF

E+

Pai

rtr

end

s+

Per

iod

Fix

edE

ffec

tsM

,Q,A

(A)

and

(B)

Ind

icat

or=

1if

ther

eare

any

mes

sage

sse

nt

fromi

toj

Ind

icat

or=

1if

ther

eare

any

mes

sage

sse

nt

fromj

toi

(i)

OL

S:

Pai

rF

E+

Pai

rtr

end

s+

Per

iod

Fix

edE

ffec

ts

(iv)

Pai

rF

E+

Pai

rtr

end

s+

Per

iod

Fix

edE

ffec

ts

M,Q

,A(A

)an

d(B

)

19

Tab

le5:

Sp

ecifi

cati

on3

-re

sult

sto

be

rep

orte

d

Ou

tcom

e(s

):R

egre

ssio

nsp

ecifi

cati

on

(s)

Peri

od

sS

am

ple

#m

essa

ges

sent

bet

wee

nI

an

dj

#m

essa

ge

rece

ived

bet

wee

nI

an

dj

Agg

regate

spec

(3):

(i)

OL

S:

Pair

FE

+P

air

tren

ds,

+P

erio

dF

ixed

Eff

ects

(ii)

Pois

son

:P

air

FE

+P

air

tren

ds,

+P

erio

dF

ixed

Eff

ects

(iii

)N

egati

veB

inom

ial:

Pai

rF

E,+

Pai

rtr

end

s+

Per

iod

Fix

edE

ffec

ts

Bil

ater

alsp

ec(4

):(i

v)

OL

S:

Cou

ntr

yF

E+

Cou

ntr

ytr

end

s,+

Per

iod

Fix

edE

ffec

ts(v

)P

ois

son

:C

ountr

yF

E+

Cou

ntr

ytr

end

s,+

Per

iod

Fix

edE

ffec

ts(i

ii)

Neg

ati

veB

inom

ial:

Cou

ntr

yF

E+

Cou

ntr

ytr

end

s,+

Per

iod

Fix

edE

ffec

ts

Mon

thly

,Q

uart

erly

,A

nnu

al

(A)i

rest

rict

edto

:co

untr

ies

inE

i

jre

stri

cted

toco

untr

ies

outs

ideE

i

bu

tn

ever

grey

list

ed.

Bot

hre

stri

cted

to‘c

ore’

cou

ntr

ies

and

full

sam

ple

.

20

Figure 4: Routing of MT103 messages when intermediaries/correspondents are involved

(a) Serial method

(b) Cover method

Note: both figures are reproduced from CPMI (2016)

21