Introduction 02 - Deloitte

2y ago
12 Views
2 Downloads
1.60 MB
42 Pages
Last View : 5d ago
Last Download : 2m ago
Upload by : Mariam Herr
Transcription

The case for artificial intelligencein combating money launderingand terrorist financingA deep dive into the application ofmachine learning technology

ContentsIntroduction02The promise of machine learning in compliance04Uses and potential applications of machine learningin fighting money laundering08Navigating and adopting machine learning12Case Study: UOB24Conclusion34End notes37Contact us38

Glossary of TermsAI –Artificial IntelligenceAML –Anti-Money LaunderingCFT –Counter Terrorist FinancingFCC –Financial Crime ComplianceLIME –Local Interpretable ModelAgnostic ExplanationsMAS – Monetary Authority ofSingaporeML –Machine LearningPEP –Politically Exposed PersonPOC –Proof of ConceptPPP –Public-private PartnershipsSLA –Service Level AgreementSAR –Suspicious Activity ReportSTR –Suspicious Transaction ReportSHAP –SHapley Additive exPlanation

The case for artificial intelligence in combating money laundering and terrorist financing IntroductionIntroductionCombating money laundering is an enormous task, and it comes withsubstantial costs and risks, including but not limited to regulatory,reputational and financial crime risks.Managing these risks rest with the guardiansof the financial system. Moreover, criminalscontinue to evolve in their launderingtechniques, finding and exploiting loopholesin the system to move money.These criminal minds are also capable ofusing new technologies such as onlinebanking, electronic payments, andcryptocurrencies to move illicit funds acrossborders at breakneck speed. This createscomplex and layered transactions that areincreasingly real-time, making it difficultto monitor and to detect with traditionalapproaches.At the heart of criminal activity aresophisticated money launderers withthe ability to move illicit funds seamlesslythrough the formal financial system. Thesemoney launderers are sophisticated andpose a serious threat to financial institutionsacross the globe, and their activities have adevastating consequence for society as well.As a result, societal ills such as terrorism,drug and human trafficking challenge socialstructures and order, societal governance, aswell as open and fair commerce. For thesereasons, the importance of continuousimprovement of an organisation’s financialtransaction monitoring and name screeningeffectiveness has never been more critical inthe digital age.2Singapore, as a top-41 global financial centrehas a front row seat to these money launderingthreats. As a nation, Singapore is not immuneto new laundering threats that emergeexpediently. In fact, the country has taken thelead in addressing these evolving ‘threatscape’through innovative initiatives, solutions andforums, as seen in the continued run of theSingapore Fintech Festival by the MonetaryAuthority of Singapore (MAS).More than ever, there is a need for the industryand regulators to sharpen surveillance on anongoing basis, or risk being at the wrong endof the threatscape. With the potential of publicprivate partnerships (PPP), an ecosystemdriven strategy will be a key step forwardto combat money laundering and terroristfinancing risks in the future. In fact, bankssuch as UOB have taken steps to work withdifferent players in the ecosystem to combatmoney laundering, as seen in a case study ontheir journey of co-creating a machine learningsolution that is discussed in this Whitepaper.In the interim, forging closer links to realisingthe benefits and the full potential of PPP,innovation and new technologies are the bestbet to better manage regulatory risks.

“When financial institutions, regulators,enforcement agencies work together usingnew technologies and sharing intelligenceand information, the entire ecosystem standsto benefit. It is paramount that internationalcooperation is prioritised to anchor goals towardfighting financial crime and making an impactthat matters in the face of rapid and fast-shiftingcriminal typologies.”Radish Singh, SEA Financial Crime Compliance Leader and AMLPartner, Deloitte Financial Advisory, Forensic, Deloitte

The case for artificial intelligence in combating money laundering and terrorist financing The promise of machine learning in complianceThe promise of machinelearning in complianceToday, banks have invested and continue to invest billions ofdollars to prevent money laundering.However, traditional technologicalapproaches to combat these evolvingthreats are meeting with less successresulting in large numbers of “falsepositives” (952 per cent of false positivesin some organisations where 98 per centdo not result in a SAR or STR) and an armyof resources to tediously dispose of these.Undoubtedly, using limited resources toclose off non-material and unimportantalerts is manual and onerous.Furthermore, the ballooning costs of AntiMoney Laundering (“AML”) compliance (ofmore than US 25 billion3 in the UnitedStates alone) coupled with the high volumeof backlog alerts swamp complianceteams and potentially distract them from‘true’ high risk events and customercircumstances.Needless to say, this demands a moreefficient and effective approach tostrengthen AML efforts. Ultimately,compliance teams ought to befocused on higher value work such asissues resolution and also to ensurethat policies and procedures arecontinuously reviewed and updated toreflect the typologies detected acrossthe bank.4In response, banks need to embracethe opportunity to apply technologicalinnovations – these include robotics,cognitive automation, machine learning(“ML”), data analytics and artificialintelligence (“AI”) to their AML complianceframework. As a result, the banking andfinance industry has been exploringopportunities to use AI and ML to alleviatesome of the compliance burden.In fact, a report released by the WorldEconomic Forum and Deloitte in August2018 entitled “How AI is transformingthe financial ecosystem”4 showed thatthe continued development of AI willradically transform the front and backoffice operations of financial institutions.The report goes on to state that the AIexpansion will require adjustments to longstanding regulations and major changesto the current structure of global financialmarkets. This shift is an opportunity forcompliance teams to strategically invest innew technologies in order to enable banksto become more future ready.

The case for artificial intelligence in combating money laundering and terrorist financing The promise of machine learning in complianceTechnology companies and banks areactively designing AI solutions and toolsto better assess high risk jurisdictions,to identify potentially problematic orsuspicious funds movements, and to refinethe screening of Politically Exposed Persons(PEP) and sanctioned individuals and/or organisations. Regulators are also inagreement that such advanced technologiescan and should be leveraged by banks toimprove risk identification and mitigation.“As financial institutions and FinTechsincrease the experimentation and useof AI and data analytics to improvetheir services, government agenciesneed to ensure that our support,policies and regulations are attuned todevelopments and remain supportive ofthese new technologies .”5Dr David Hardoon, Chief Data Officer, Monetary Authority of SingaporeAs some of the main advancements intechnology and analytics are relativelyrecent, there is often confusion when itcomes to understanding what AI and MLactually entail and the differences betweenthe two. To be clear, ML is a subset of AI, andwithin AI, there exist further subsets suchas natural language processing, robotics,image recognition, speech recognition, deeplearning, and virtual agents.5

6

The case for artificial intelligence in combating money laundering and terrorist financing The promise of machine learning in complianceUntil very recently, banks have reliedon traditional, rules-based AMLtransaction monitoring and namescreening systems, which generate highnumbers of false positives due to rulesthresholds (this will be discussed in thenext section). Accordingly, ML has servedas the first port of call for many banksbeginning their journey to advance theircompliance innovation programmes.Innovation in compliance is needed bothto reduce false positives and to bringabout greater effectiveness in the mannerin which AML and/or Counter TerroristFinancing (“CFT”) risks are monitoredand addressed by banks. As Alan Turingfamously said, “What we want is a machinethat can learn from experience.” Turing’sthinking can be applied to banks andAML compliance as these institutionsface increasing threats and risks.As the industry forges ahead with ML, theseadvancements present opportunities forbanks to consider the strategic creation ofan AI ecosystem in this wave of innovation.The learnings from the traditional approachto transactions monitoring reflect thatoperating in a silo environment has itspitfalls of, amongst others, creatinginconsistencies across the industry.With an AI ecosystem, it will meanmore sharing and transparency ofstandards, which can be advantageousto the industry in achievinggreater expertise, effectiveness,and efficiencies when consideringthe adoption and integration ofmachines into the mainstream.AI involves machines thatcan perform tasks that arecharacteristic of humanintelligence – anything thatcan be described by a humanbeing can be mimickedthrough AI applications. MLis a branch or subset of AI,encompassing those actionswhere a machine learnsto understand patterns indata or tasks without havingpre-defined coding. MLpromises to be particularlyrelevant and impactful fortransaction monitoringplatforms within banks.7

The case for artificial intelligence in combating money laundering and terrorist financing Uses and potential applications of machine learningUses and potential applications ofmachine learning in fighting moneylaunderingThe unintended consequences of regulatoryexpectations on AML compliance hasspurred banks to ensure that they are onthe right side of the regulatory fence andconsider the use of machines to learn andto detect suspicious activity and behaviourmore critically and effectively.8Banks are therefore keen toleverage the rise in computingpower to analyse large volumesof data assets and to “learn”from the results.In the compliance realm, thereis a real sense of opportunitywhere ML can assist inenhancing effectiveness,efficiency and accuracy ofprocesses within a bank’s coremoney laundering and terroristfinancing risks detection andreporting system. For example:

The case for artificial intelligence in combating money laundering and terrorist financing Uses and potential applications of machine learning1. ML algorithms can be taught to detectand to recognise suspicious behaviour andrisk rate them accordingly. For instance,machines will learn and focus on “bigger” riskswhilst knowing when to omit non-anomaloustransactions that do not present any risks asdictated by customers’ profile and behavior.The greatest opportunity for application is inthe money laundering and terrorist financingtransaction monitoring process. Traditionalsystems detect very specific typologies that canbe circumvented. Furthermore, the results fromthese models contain more noise than ‘signals ofrisk’ as the net is often cast wide in order to notmiss a potentially suspicious activity.2. Combining outputs of existingsystems, ML models can be trained toidentify the behavioural characteristicsor indicators that highlight whenactivity is truly suspicious. MLtechniques such as anomaly detectioncan be used to identify previouslyundetected transactional patterns, dataanomalies and relationships amongstsuspicious individuals and entities.Such ML techniques no longer requirestatic rules, and are based on known andtrending patterns or threats that makeit harder for criminals to hide within thebank’s environment.By relaxing rule thresholds to capture suspicioustransactions that are closer to ‘normal’ activity,there will inevitably be larger numbers of alertsrequiring costly manual reviews to resolve.However, only a very small number of thesealerts will result in suspicious behavioursrequiring escalation.9

The case for artificial intelligence in combating money laundering and terrorist financing Uses and potential applications of machine learning3. ML can be applied to name screeningwhere systems are required to screen customernames against global lists of known criminalsand black-listed and sanctioned organisationsand individuals.The challenge faced by many banks is balancing‘fuzziness’ with accuracy. In other words,current text matching algorithms are not aneffective tool to track potential data capturenuances such as the order of names, titles,salutations, abbreviations, name variants,common misspellings, etc. In addition, the taskbecomes complicated further when dealingwith common names where it is difficult topinpoint the exact individual. The prevailingrules-based approach is both onerous andmanual, resulting in increased workloadfor compliance, as well as potential gaps insurveillance and monitoring.4. Applying ML to improve the matchingcriteria as well as predicting the likelihoodof a name match can lead to significantefficiency gains while also increasing efficacy byidentifying hidden links (conducting link analysisfrom available) or relationships.Enriching the data with more contextualinformation about the entity such asdemographic, network and behavioural data iswhere the true enhancements to the accuracyof screening processes lie.Some other areas that are gaining tractioninclude Fraud Detection, Automated Reporting,Enhanced Surveillance including voice, video,text, pattern based transaction monitoring.1010When determiningwhere to apply ML, it isimportant to understandthe opportunities in termsof the bank’s innovationstrategy, key priorities,unique financial crimecompliance risks, existingoperational challenges,and long-term feasibility.A smart approach to compliancewill also be of strategiccommercial value. Aside fromhaving better known risks thatcan be escalated and investigatedby compliance teams, optimisinghistorically compliance drivenmethods such as the “KnowYour-Customer” process offersanother opportunity for thebanks’ business portfolio ofclients to be enriched.In the long term, the bank willhave a fuller and more robustprofile of their clients that canbe used to enhance ongoingclient management to delightcustomers and to build loyalty.In this regard, ML models canmodernise compliance byremoving needless interruptionsto services while achievingdeeper and more customisedinsights to improve customerexperience.

The case for artificial intelligence in combating money laundering and terrorist financing section title goes here The case for artificial intelligence in combating money laundering and terrorist financing Uses and potential applications of machine learningIn terms of business efficiency, using ML techniques incompliance has immense potential to reduce manualprocesses, and even streamline repetitive tasks that oftenweigh compliance teams down. Such improvementscan also alleviate cost and make compliance a moremeaningful exercise.The first steps to success include: maturity assessment or modelvalidation of the existing technologyin place and identification ofopportunities for enhancement; understanding of the key risks,threats and complexities of thebusiness including the bank’scorrespondent banks, customersand known transaction risks; effective governance frameworkfor AI and ML in the bank – clearfocus on AML/CFT controls it will bedeployed to address data qualityissues, project management,stakeholder expectationmanagement and engagement; approach for operationalising anddocumenting the AI process withparticular focus on deploymentinto production and an in-depthunderstanding of the models andalgorithms used; appropriate structure for monitoringand validating the approach foroperationalisation as well as theoutcome of deployment that it meetsregulatory objectives and addressesrisks appropriately; and robust due-diligence of vendorsselected to provide the technologyknow-how and infrastructure. strategy, framework and intendedoutcome of deploying AI and ML;1111

The case for artificial intelligence in combating money laundering and terrorist financing Navigating and adopting machine learningNavigatingand adoptingmachinelearningEven with the potential to beharnessed from ML and itspromise of increased efficacy,there are considerations thatshould be addressed beforecommencing on this journey.A UOB case study presented inthis Whitepaper details manyof these considerations.The key considerations areas follows:12Ensuring model results areconsistent and reproducibleIn the case of AML/CFT transactions monitoring,ML models dealing with high risk processes, willbecome an integral part of the control frameworkand it is imperative that: there is the ability to reproduce a bank’s resultswithin the production settings for the purposesof, inter alia, providing assurance from aregulatory standpoint and for maintaining a goodquality audit trail; and the model is designed and trained to produce aspecific and consistent set of results by learningbehaviours and patterns within data sets.There is recognition that due to variable factorssuch as random initialisation of parameters,different chip architectures that performcalculations differently, changes to data andchanges to underlying statistical libraries, it isoften a challenge to reproduce a set of results in aconsistent fashion.To address this inherent challenge, banks mustput in place a robust and continual process toevaluate and to validate the performance of theirmodels. This includes a governing frameworkthat measures performance, documents thetraining process, and ensures that steps can bereplicated with the same results. The frameworkshould test the performance impact of anychanges to production prior to release, and shouldalso execute unit testing algorithms, in order tounderstand the impact of specific componentsand parameters on performance.

The case for artificial intelligence in combating money laundering and terrorist financing Navigating and adopting machine learningFlexibility and customisationMuch of the shift and speed in theadoption of ML has led to the increaseddemand for “off-the-shelf” models thatuse pre-built data sets that are easy toimplement. The flip side is inaccuracywhen using different data sets.Off-the-shelf models are trained usingdata that are not specific to the bank anddo not reflect learnings from underlyingdata and transactions, key customersegments and profiles, as well as productsand services offered by the bank. Themachines are also not trained with datathat contain other risk nuances, existingtrends or typologies such as high riskcross-border transactions or simply thosethat do not correspond to the customers’behavior.While limitations do exist, off-the-shelfML models may still be used, but onlywhen they are calibrated according tothe bank’s unique data sets, profile andrequirements.In addition, transfer learning (which is theability to customise a pre-trained modelusing new and relevant data), imagerecognition in name screening (an existingoff-the-shelf ML image recognitionmodel could be modified to performfacial recognition) or natural languageprocessing, as applicable, are additionalenhancements that could be looked into.In most cases, banks will increasinglyrequire scalable and deployable MLmodels, which may require customisationto suit a bank’s needs and one thathas the ability to scale with a sustainedimpact. In such circumstances, the longterm benefits and scalability should berecognised rather than the focus onits relatively higher initial cost due tocustomisation.13

The case for artificial intelligence in combating money laundering and terrorist financing Navigating and adopting machine learningSustainability, scalability and industrialisation of themachine learning model – from POC to ProductionThe process to move models from Proofsof Concepts (“POC”) into a live settingneeds to be industrialised to the pointthat all the necessary considerationssuch as sustainability, scalability andindustrialisation as well as controls areembedded.The model should be scalable to handle inproduction: new data volumes being pumpedthrough for prediction; use feedback from live data to adjustparameters and help the model re-learn; business requirements that impact theproduction design;and appropriate and resilient technical andperformance failovers14To industralise ML within a bank,there has to be proper governance toensure consistency in the design, useand maintenance of the models. Anenterprise-wide strategy, framework andplatform is critical for the deployment ofmultiple ML models (that could be a mixof both Off-the-shelf and bespoke MLmodels) to effectively use resources andmanage risks across the bank.Moving into production also requiresworking towards the creation of acentralised platform embedded with datacontrols to attain accuracy, completeness,privacy and regulatory compliancethat will go the distance to ensure datastandardisation and reliability acrossthe organisation. Cleansed data is a direneed for the success of ML model andwe will expand on this point further in asubsequent section.

The case for artificial intelligence in combating money laundering and terrorist financing Navigating and adopting machine learningRecognising and addressing risks associatedwith machine learningFigure 1: The associated risks with using ML modelsINPUTDATAALGORITHMDESIGNInput data is vulnerable to risks,such as biases in the data used fortraining; incomplete, outdated, orirrelevant data; insufficiently largeand diverse sample size;inappropriate data collectiontechniques; and a mismatchbetween the data used for trainingthe algorithm and the actual inputdata during operations.Algorithm design is vulnerable torisks, such as biased logic, flawedassumptions or judgments,inappropriate modelingtechniques, coding errors, andidentifying spurious patterns in thetraining data.OUTPUTDECISIONSOutput decisions are vulnerable torisks, such as incorrectinterpretation of the output,inappropriate use of the output,and disregard of the underlyingassumptions.UNDERLYING FACTORSHUMAN BIASESTECHNICAL FLAWSUSAGE FLAWSSECURITY FLAWSHuman biases: Cognitivebiases of model developersor users can result in flawedoutput. In addition, lack ofgovernance andmisalignment between theorganisation’s values andindividual employees’behaviour can yieldunintended outcomes.Technical flaws: Lack oftechnical rigor or conceptualsoundness in thedevelopment, training,testing, or validation of thealgorithm can lead to anincorrect output.Usage flaws: Flaws in theimplementation of analgorithm, its integration withoperations, or its use by endusers can lead toinappropriate decisionmaking.Security flaws: Internal orexternal threat actors cangain access to input data,algorithm design, or itsoutput and manipulate themto introduce deliberatelyflawed outcomes.Source: Managing algorithm risks: Safeguarding the use of complex algorithms and machine learning, Deloitte6To effectively manage the risks of cutting edge technology such as ML, banks will need toestablish a solid framework; to restructure and to modernise traditional risk-managementframework and capabilities. This goes back to the key success factors discussed under thesection on “Uses and potential applications of machine learning in fighting money laundering.”15

The case for artificial intelligence in combating money laundering and terrorist financing Navigating and adopting machine learningReadiness for a machine learning pilotDifferent organisations have different levels of readiness for integration and usage.Ahead of any ML pilot programmes, the following considerations should be weighed:Figure 2: Readiness for a machine learning pilotWhat is the problembeing solved?Is the problem statement clear?Have the internal stakeholdersbeen informed?BUSINESSPROBLEMAre there people internally whounderstand the value of ML?Can they work with the vendor onthat project to validate outcomes?Are there people to take overafter the POC?PEOPLEIs there sufficient dataquality that is easilyaccessible to create ameaningful model?PROCESSIs there a process inplace to select andon-board a newvendor?DATAREADINESSFOR AN MLPILOTHARDWAREIs there somewhere tohost the ML applicationthat is secure andpowerful enough?LEADERSHIPDoes your organisationtrust ML?Is there buy-in fromyour organisation withan allocated budget?16BROADENINGTHE BENEFITSWhat are the benefits ofthe applications beyondcompliance operations?

The case for artificial intelligence in combating money laundering and terrorist financing Navigating and adopting machine learningData managementThe ML model is only as good as the data it receives. It is simple to conclude that badquality data will produce bad results. As such, selecting data sets to train the modelmay cause unintended biases. To mitigate this risk, it is important to consider:Figure 3: Considerations for data management Data is not perfect in an organisation.Issues such as duplicated accounts,information not being captured,information not being maintained acrosssystems and lack of consistency ininformation capture lead to themodel identifying patternsthat do not exist orthat are biased toWhatspecificimpact woulddemographics. Are the issuespotentiallyhiding thepatterns?known dataquality issueshave on the MLmodel? If the model is trained to understandtransactions between 100 and 50,000 what happens when there isa transaction worth 1m? Similarly ifthe training data only hasinformation for Monday –Friday what happenswhen a transactionHowoccurs on areasonable isSaturday?it that the rangeof values used totrain the model isprevalent inproduction?DATAMANAGEMENT If featuresareaggregatedDoes the featureinto a moreencoding makeusefulsense?representation(e.g. High/Medium/Low), isthe process correctand will it scale for allscenarios the model willencounter? If one hot encoding is used (create a newfeature for every category that is either aone or zero), is the process sustainable?Will those values always be retrieved?Is the datasufficientlyrepresentative? Does thedatacontain allthe patternsthat it willneed to make adecision on? Does the data containsufficient information tomake defensible decisions? Are any specific patterns ordemographics overrepresented whichcould lead to a skew in modeldecisions?17

Having good and clean data sets is acritical component for any bank venturinginto designing ML models. The foundationof data management is acquisition,preparation and maintenance, uponwhich ML models are embedded. In turn,the benefits of a well-managed datainfrastructure will enable multiple MLmodels to leverage richer datasets anddeliver valuable insights and patterns tohelp with complex analyses, especiallyin countering money laundering andterrorist financing.

The case for artificial intelligence in combating money laundering and terrorist financing Navigating and adopting machine learningExplaining the inner workings ofmachine learning modelsAs models become increasingly complexto uplift performance outcomes, theinner workings of the algorithms becomesmore opaque. Banks are often faced withthe tricky issue of balancing betweendecoding the algorithms and maintainingaccuracy.With the adoption of ML technology,banks are expected to understandand defend the algorithms used by themachine that may bring about betterpredictive capabilities but are significantlymore complex. In order to understand ifthe model is picking up valid patterns andthat it is not overfitting the training data,the outputs must be transparent andauditable.While ML models may pick up associationswithin the data, it is not necessarily aproof of causation and this may result infalse hypotheses. When left undetectedand not remedied, the opposite effect willcause greater risks and lack of precisionin monitoring AML/CFT risk, resulting in alack of accurate regulatory assurance.Practitioners using the model will needto understand decisions made by themachine and be able to explain how therelated data points shape the outcomes.Decisions need to be transparent todetermine if they were fair, ethicaland in line with legal and regulatoryrequirements.Interpretability is thedegree to which a humancan understand the causeof a decision. This is criticalin understanding modelweaknesses as well asbeing able to defend anydecisions made to theregulator as well as toother stakeholders.19

The case for artificial intelligence in combating money laundering and terrorist financing Navigating and adopting machine learningOne way to understand why models make decisionsis to appreciate the important features that drivethe model. This allows for determination of the coredecision criteria for the model but does not help withunderstanding why specific predictions were made.Additionally, if predictions impact customers, theymay want the right to understand the manner inwhich their data is being used and reasoning behindthe decision that impact their banking experience.Ultimately, any technology deployed in the AML/CFT framework is a control in itself. Therefore, itsexplanability and transparency is critical in providingongoing regulatory assurance that the risk is managedadequately as with managing customer experience.There has been a lot of research in this space on howto make model outcomes more explainable. Someexamples of this are SHapley Additive exPlanation(SHAP) and Local Interpretable Model-AgnosticExplanations (LIME) which attempt to determine theimpact of specific features on localised predictions.SHAP is filling the gap to provide the link betweenaccuracy and human interpretation of models usinggame theory7. While LIME aids the predictions of MLclassificaters 8 . Research in th

A deep dive into the application of machine learning technology. Introduction 02 The promise of machine learning in compliance 04 Uses and potential applications of machine learning 08 in fighting money laundering Navigating and adopting machine learning 12 Case Study: UOB 24 Conclusion 34 End notes 37

Related Documents:

XaaS Models: Our Offerings @DeloitteTMT As used in this document, "Deloitte" means Deloitte & Touche LLP, Deloitte Tax LLP, Deloitte Consulting LLP, and Deloitte Financial Advisory Services LLP. These entities are separate subsidiaries of Deloitte LLP. Deloitte & Touche LLP will be responsible for the services and the other subsidiaries

Deloitte & Touche South Africa is referred to throughout this report as Deloitte South Africa, and Deloitte Pan African Trust is referred to throughout this report as Deloitte Africa. Deloitte Africa holds practice rights to provide professional services using the Deloitte name which it extends to Deloitte entities within its territory,

May 02, 2011 · Deloitte & Touche LLP Cleveland, Ohio 1 216 589 5717 tgriffiths@deloitte.com Theresa Cui . Engagement Consultant . Deloitte & Touche LLP . Cleveland, Ohio Cleveland, Ohio 1 216 589 5018 1 216 . tcui@deloitte.com . Kathie Schwerdtfeger Advisory Principal Deloitte & Touche LLP . Austin, Texas 1 512 691 2333 . kschwerdtfeger@deloitte.com .File Size: 720KB

Knabe, Andrea Consulting Deloitte Consulting LLP Chicago Kwan, Anne Consulting Deloitte Consulting LLP San Francisco . Miller, Christian L. Tax Deloitte Tax LLP Washington DC . Smith, Sandra Consulting Deloitte Consulting LLP Chicago Spangrud, Chad Audit & Assurance Deloitte & Touche LLP Costa Mesa Springs, Christanna R. Tax Deloitte Tax .

** Deloitte Risk Advisory, Löffelstrasse 42, D-70597 Stuttgart, Germany, anlanger@deloitte.de *** Deloitte Legal, Schwannstraße 6, 40476 Düsseldorf, Germany, fwesche@deloitte.de **** Deloitte Risk Advisory, Löffelstrasse 42, D-7059

Deloitte MCS Limited is a subsidiary of Deloitte LLP, the United Kingdom member firm of DTTL. . Jason Geller jgeller@deloitte.com. Deloitte Consulting LLP. Asia Pacific & China. Jungle Wong junglewong@deloitt

Nov 21, 2013 · Deloitte Deloitte & Touche LLP Ten Westport Road P.O. Box 820 Wilton, CT 06897-0820 Tel: 2013 761 3000 Fax: 1 203 834 2200 www.deloitte.com October 29, 2013 Elizabeth M. Murphy

As used in this document, "Deloitte" means Deloitte & Touche LLP, which provides audit, assurance and risk management related services, Deloitte Consulting LLP, which provides strategy, operations, technology, systems, outsourcing and human capital consulting services, Deloitte Tax LLP, which provides tax services, and .