Skip to main content
Log in

Looking at ‘Crying Wolf’ from a Different Perspective: An Attempt at Detecting Banks Under- and Over-Reporting of Suspicious Transactions

  • Research Paper
  • Published:
Italian Economic Journal Aims and scope Submit manuscript

Abstract

This study aims to assess banks’ compliance in reporting suspicious transactions. It does so by estimating an econometric model on the suspicious transaction reports (STRs) filed by individual Italian banks from each of the provincial districts they operate in. Regressors include (1) indicators of banks’ operational activities, (2) measures of money laundering risk and (3) proxies of economic activity, all of which at local level. At an operational level, the model provides a tool that supervisory authorities can use to detect potentially under-reporting intermediaries, thus better targeting off-site controls and on-site inspections. More in general, the results provide some insights on banks’ reporting behavior at large. The main threat to the effectiveness of anti-money laundering systems is considered the asymmetry in the incentives: since sanctions apply only to omitted reports, banks have an incentive to over-report, thus potentially flooding the authorities with noise (‘crying wolf’ syndrome). Results show that the STR-filing strategies adopted by the banks being scrutinized may not necessarily give rise to this scenario.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Fig. 1
Fig. 2

Similar content being viewed by others

Notes

  1. According to most legislations and international standards, if a financial institution suspects that funds are the proceeds of a criminal activity, or are related to the financing of terrorism, is required to report promptly its suspicion to the national anti-money laundering authority.

  2. The study focuses on 2012 since it is the first year in which the UIF’s risk-rating system for STRs (see Sect. 2.1 below) was operating reliably. It was also the latest year for which, when the analysis was performed, data for all variables were available. A panel model was subsequently estimated adding data spanning from July 2013 to June 2015, obtaining substantially similar results.

  3. One exception is Masciandaro and Volpicella (2014), who, by setting up a political economy model, try to explain why after the 2001 September 11 events anti-money laundering systems have increasingly featured law-enforcement FIUs, as opposed to financial ones.

  4. Appendix 1 describes the theoretical framework which means to provide a simplified outline of the banks’ behavior underlying the empirical model.

  5. Full Italian name is Segnalazioni Anti-Riciclaggio Aggregate.

  6. The Italian anti-money laundering law (Legislative Decree 231/2007) requires banks and other intermediaries to record all single transactions equal to 15,000 euros or more in a specific archive (Single Electronic Archive). Each month intermediaries file these data to the UIF after aggregating individual records according to several criteria, which include the customer’s place of residence and economic sector, the intermediary’s branch where the transaction took place and the type of the transaction.

  7. See Europol (2015).

  8. See, for instance, ‘The Money-Laundering Cycle’, United Nations Office on Drugs and Crime (UNODC) (http://www.unodc.org/unodc/en/money-laundering/laundrycycle.html).

  9. That such conducts are widely spread among criminals and launderers is consistent with the findings of FATF (2015).

  10. Usually, when funding is received from a loan-shark, as collateral, the latter requires the borrower to issue several cheques with a future date on which they will be cashed. More often than not, reimbursement does not take place for lack of adequate funding, thus the cheques are recorded as unpaid. That further hampers the access to bank funding by the borrower (which was the original reason for referring to a loan-shark), thus forcing him or her to require additional funding to the usurer.

  11. The relevant data refer to 2010, the most recent year for which the information was available when the analysis was carried out.

  12. The vulnerability indicator factors in several variables. A first set refers to features of the credit market and to measures of potential bottlenecks in the local supply of banking and financial services; in particular, the variables included in the indicator are the total amount and number of lines of credit granted by banks and other financial intermediaries, and the total amount and number of non-performing loans. Additional components account for local economic conditions (i.e., the businesses birth-to-death ratio), households consumption patterns (i.e., the number of big distribution outlets, new vehicles registrations), and the extent of the underground economy (i.e., per-capita energy consumption). Provinces are ranked against each and every variable according to a growing scale of risk. The average rank for each province provides the values of the final vulnerability indicator.

  13. The distribution of Italian banks in 2012 was highly skewed towards small-sized intermediaries, with few branches (there were several cases of banks with just one branch) operating in a limited number of provinces. On the other end of the spectrum, there were a few big lenders with branches covering the whole of the country. Given this outlook, it is not surprising (and certainly not by itself a signal of low compliance with AML reporting obligations) that a lot of banks may have failed to file even a single STR from a province in the whole year, whilst the overwhelming majority of STRs were concentrated in a few lenders.

  14. Sum of squared Pearson’s residuals (Cameron and Trivedi 2013).

  15. The so called ‘likelihood ratio index’, R2 = 1 − (Lfit/L0), where Lfit is the likelihood of the fitted model and L0 is the likelihood of the no parameters model (Cameron and Trivedi 2013).

  16. The count fitting is equal to \( \sum \frac{{({\text{n}}\bar{\text{p}}_{j} - {\text{n}}\hat{\text{p}}_{j} )^{2} }}{{{\text{n}}\hat{\text{p}}_{j} }} \)where \( {\hat{\text{p}}}_{\text{j}} \) is the expected observations (as a share of the total) with j STRs and \( {\bar{\text{p}}}_{\text{j}} \) is the corresponding observed value.

  17. For robustness checks, the selected NB models for both sub-samples were estimated with clusterized standard errors at regional level, in order to take into account potential interdependence between banks operating in the same area. Coefficients for most regressors remained significant, in spite of higher standard errors.

  18. This is consistent with Gara et al. (2019) which shows that, in 2012 and 2013 pursuant to an audit by the anti-money laundering supervisory authority, Italian banks responded by filing more STRs across all risk levels.

  19. This is presumably due, to a large extent, to the widespread centralization of the STR selection process which takes place in Italian banks.

  20. There is only one exception, with one bank appearing to under-report in four provinces and over-report in three provinces.

  21. It is worth noting that, since the authorities fix \( P_{aml} = finserv_{i}^{\prime } /p_{fs} \) so that (4) holds as an equality, the bank is indifferent between setting up business in area i or not, hence it may end up filing no STRs altogether.

References

  • Al Rashdan M (2012) An analytical study of the financial intelligence units’ enforcement mechanisms. J Money Laund Control 15(4):483–495

    Article  Google Scholar 

  • Barth JR, Caprio G, Levine R (2004) Bank regulation and supervision: what works best? J Financ Intermed 13(2):205–248. Elsevier

  • Beck T, Demirgüç-Kunt A, Levine R (2006) Bank supervision and corruption in lending. J Monet Econ 53(8):2131–2163. Elsevier

  • Cameron AC, Trivedi PK (2013) Regression analysis of count data, 2nd edn. Cambridge University Press, Cambridge

    Book  Google Scholar 

  • Dalla Pellegrina L, Masciandaro D (2009) The risk-based approach in the new european anti-money laundering legislation. A law and economics view. Rev Law Econ 5(2):931–952. De Gruyter

  • Europol (2015) Why is cash still king? A strategic report on the use of cash by criminal groups as a facilitator for money laundering. Europol, The Hague

  • FATF—Financial Action Task Force (2015) Money laundering through the physical transportation of cash. FATF/MENAFATF, Paris/Manama

    Google Scholar 

  • Gara M, Manaresi F, Marchetti DJ, Marinucci M (2019) The impact of anti-money laundering oversight on banks’ suspicious transaction reporting: evidence from Italy. UIF, Quaderni dell’antiriciclaggio, Collana Analisi e Studi, n. 12

  • Guo JQ, Trivedi PK (2002) Flexible parametric models for long-tailed patent count distributions. Oxf Bull Econ Stat 64(1):63–82. Wiley

  • KPMG (2014) Global Anti-money Laundering Survey 2014. https://home.kpmg.com/xx/en/home/insights/2014/01/global-anti-money-laundering-survey.html (Accessed 4 Jan 2016)

  • Lambert D (1992) Zero-inflated Poisson regression, with an application to defects in manufacturing. Technometrics 34(1):1–14. American Statistical Association

  • Masciandaro D (1999) Money laundering: the economics of regulation. Eur J Law Econ 7(3):225–240. Kluwer Academic Publishers

  • Masciandaro D, Filotto U (2001) Money laundering regulation and bank compliance costs: what do your customers know? Economics and the Italian experience. J Money Laund Control 5(2):133–145. Emerald

  • Masciandaro D, Volpicella A (2014) Economics and politics in designing supervision: the case of the FIUs against money laundering. Paolo Baffi Centre Research Paper No. 2014-156. Bocconi University

  • McCullagh P, Nalder J (1983) Generalized linear models. Chapman and Hall, London

    Book  Google Scholar 

  • Pok WC, Omar N, Sathye M (2014) An evaluation of the effectiveness of anti-money laundering and anti-terrorism financing legislation: perceptions of bank compliance officers in Malaysia. Aust Account Rev 24(4):394–401

    Article  Google Scholar 

  • Takàts E (2009) A theory of “crying wolf”: the economics of money laundering enforcement. J Law Econ Org 27(1):32–78. Oxford University Press

  • Unger B, van Waarden F (2009) How to dodge drowning in data? Rule and risk-based anti-money laundering policies compared. Rev Law Econ 5(2):973

    Google Scholar 

Download references

Acknowledgements

We thank for useful suggestions Lucia Dalla Pellegrina, Marco Lippi, Domenico J. Marchetti, Marco Marinucci, Paolo Naticchioni, participants at the 2014 ISLE Conference in Rome, the 2015 UIF-Bocconi Workshop on “Quantitative Methods and the Fight to Economic Crime” and two anonymous referees. All remaining mistakes are all the authors’ fault. The views and the opinions expressed in this paper are those of the authors and do not represent those of the institutions they are affiliated with.

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Mario Gara.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendices

Appendix 1: The Theoretical Framework

In a standard principal-agent setup, we have two players: the authority, with a regulatory role within the anti-money laundering system, and a bank, operating in different geographical areas.

In each area it operates the latter offers financial services at market prices and consequently files STRs to the authority. As in Dalla Pellegrina and Masciandaro (2009), these are only ‘good’ STRs, that is to say, well-grounded ones. Filing STRs in an area i requires that the bank produce a certain level of effort, which is an increasing (concave) function of the width of financial services on offer in the area and the crime rate:

$$ e_{Bi} = e \left( {finserv_{i} , \;crime_{i} .} \right) $$

The wider and more sophisticated the financial services the bank offers, the more difficult it is to detect true potential money laundering transactions to report, hence the higher the effort required to do it. Likewise, the higher the crime rate, the more efficient and powerful criminal organizations are; thus money laundering schemes are more difficult to detect and the bank is more likely to be subjected to intimidation and pressure of some kind: increasingly complex audit mechanisms need setting up in order to comply with reporting obligations. For extremely high levels of both crime and financial services supply, the effort required becomes infinite.

Since the number of STRs is a function of the bank’s effort, it can be expressed as an increasing (convex) function of both the crime rate and the provision of financial services:

$$ STR_{Bi} = f \left( {finserv_{i} ,\; crime_{i} } \right). $$
(2)

By filing STRs the bank enjoys a reputational reward, featuring decreasing marginal returns. Hence, its utility depends on the reputation gains compliance with the reporting obligation ensures net of the effort needed for filing STRs:

$$ U_{Bi} = rep (STRs_{i} ) - g(e_{Bi} ). $$
(3)

Since reputation is concave in \( finserv\; {\text{and}}\;crime \), whilst the effort is convex, the first derivative of the level of financial services (for a given level of the crime rate) is positive up to a point and negative thereinafter (see Fig. 3). It seems sensible to argue that the reputational dividends that can be reaped by filing an increasing number of STRs outweigh the costs associated to monitoring for low levels of financial services supply (and of the crime rate), whilst the opposite holds as that supply grows larger, for the reasons which have already been mentioned above.

Fig. 3
figure 3

Bank’s utility function

The bank’s budget constraint requires that the proceeds from the provision of financial services cover potential losses from penalties paid due to non-compliance:

$$ p_{fs} *finserv_{i} - P_{aml} \ge 0, $$
(4)

where \( p_{fs} \) is the market price for financial services (assumed to be the same in all areas) and \( P_{aml} \) are sanctions paid pursuant to violations to anti-money laundering regulation in that area: for the bank to operate in an area, the bank’s profits cannot drop below 0 and hence we must have \( finserv_{i} \ge P_{aml} /p_{fs} \).

The authorities’ utility depends negatively on an area’s crime rate and positively on the number of ‘good’ STRs originating from that area and the penalties applied to the bank in the same area:

$$ U_{Ai} = u_{Ai} \left( {crime_{i} ,\;STRs_{i} ,\; P_{aml} } \right)\quad \forall i. $$
(5)

In this framework, the crime rate is observed and both the bank and the authorities have it as given. As in Dalla Pellegrina and Masciandaro (2009), the authorities design the anti-money laundering regulation and the penalties associated to each level of STRs in each area. One can assume that penalties may also be negative, thus representing rewards for particularly high levels of STRs. Penalties, however, are the same for all areas, whilst all other parameters are area-specific, bar from the price of financial services.

The equilibrium level of STRs filed in an area and the financial services on offer in the same area is set by the intersection of the bank’s utility function (3) and the authorities’ (5). The shape of (5) depends on an area’s crime rate: the higher the latter, the lower the authorities utility for any combination of \( STRs_{i} \) and \( P_{aml} \). Hence, three cases can occur (see Fig. 4):

Fig. 4
figure 4

Bank’s possible equilibria

  1. 1.

    no intersection exists (for instance, because the crime rate is extremely low), in which case any level of STRs on the bank’s utility curve is inadequate for the authorities, which therefore are pushed to rise the penalties so that the bank offers no financial services in the area;

  2. 2.

    there is only one intersection, which could happen at the bank’s optimal utility level \( U_{Bi}^{*} \) only by fluke or should the authorities know the bank’s effort and its optimal level of financial services supply so as to set \( P_{aml} \) accordingly; failing that, the amount of STRs corresponding to the bank’s equilibrium utility is lower than the optimal oneFootnote 21;

  3. 3.

    there are two STRs equilibrium levels corresponding to \( U^{\prime}_{\;Bi} \) and \( U^{\prime\prime}\;_{Bi} \), the former clearly preferable to the latter; however, if authorities set penalties so that \( P_{aml} /p_{fs} > finserv_{i}^{'} \), the bank shifts to the worst equilibrium, thus filing a lower amount of ‘good’ STRs, which might well be Takàts’s undesirable ‘crying wolf’ equilibrium.

Enters the FIU, which, as in Dalla Pellegrina and Masciandaro (2009), discharge (also) supervisory function and, therefore, can measure the level of financial services a bank offers in each area it operates in. Thus the FIU no longer relies solely on the number of STRs in order to measure a bank’s level of compliance, but have a rough idea of the bank’s real effort, for instance, by estimating the amount of STRs that should correspond to the level of financial services the bank offers, given the crime rate, and comparing it to the actual flow of STRs being filed. This is exactly what is done by the model that is described in Sect. 2.

Appendix 2: Modelling Count Data

2.1 Basic Parametric Models

For time series displaying the features of count data variables, such as those listed in Sect. 3, Guo and Trivedi (2002) suggest the adoption of some semi-parametric and non-parametric models, which they recognize as significantly demanding in terms of computational implications. Hence, in line with Cameron and Trivedi (2013), we decide to resort to more traditional non-linear distributions which are typically referred to in these cases, such as the Poisson and the Negative Binomial regression models, which are discussed.

The Poisson distribution function belongs to the Linear Exponential Family (LEF), whilst the NB distribution function belongs to the Linear Exponential Family with Nuisance parameter (LEFN). In order to have consistent estimates for the βs, LEF functions require that the function of μ is correctly specified, whilst LEFN functions require that this condition be met for both functions of α and μ. Empirically, the quadratic specification for Var[Y] of NB models ensures that the data over-dispersion is adequately accounted for.

2.2 Zero-Inflated Models

Following Lambert (1992), Poisson (and hence NB) models can be adjusted so as to accommodate distributions featuring large number of zeroes, as the one being analysed here.

In order to do this, it is assumed that the distribution of Y, the observed variable, is some distribution of zeroes with probability p and a Poisson with parameter μ with probability (1  p). Formally we have:

$$ \begin{aligned} &{\text{Prob}}\;[{\text{Y}} = 0] = {\text{p}} + (1 - {\text{p}})e^{ -\mu } , \end{aligned} $$

whilst

$$\begin{aligned} \text{Prob}[Y = y_{i} ] \;{\text{is}}\;{\text{as}}\;{\text{in}}\;(1) \, \;{\text{for}}\;{\text{y}}_{\text{i}} \ne 0. \hfill \\ \end{aligned} $$

Normally it is assumed that μ and p are related to the regressors Xi and hence they are linked to each other according to some functional form. A typical specification of the relation between μ and p is the following:

$$ \begin{aligned} & \ln (\mu ) = X_{i} \beta \quad {\text{and}}\quad {\text{logit}}({\text{p}}) = - \tau {\text{X}}_{\text{i}} \beta , \end{aligned} $$

which implies that

$$ \begin{aligned} {\text{p}} = (1 + \mu^{\tau } )^{ - 1} . \\ \end{aligned} $$

As in the non zero-inflated Poisson, the parameters, and hence the βs, are estimated using a maximum likelihood approach, granting the same properties that apply to non zero-inflated Poisson and NB.

Appendix 3

See Table 6.

Table 6 Estimation results for low turnover banks—parameters of the zero-inflated model (logit)

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Gara, M., Pauselli, C. Looking at ‘Crying Wolf’ from a Different Perspective: An Attempt at Detecting Banks Under- and Over-Reporting of Suspicious Transactions. Ital Econ J 6, 299–324 (2020). https://doi.org/10.1007/s40797-020-00122-3

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s40797-020-00122-3

Keywords

JEL Classification

Navigation