【作者】 刘璇；

【导师】 张朋柱；

【作者基本信息】 上海交通大学 ， 管理科学与工程， 2010， 博士

【摘要】 金融是现代经济的主要支柱之一,金融系统的稳定性更是世界各国金融监管部门的主要任务和核心职责之一。金融信息化改变了传统的金融体制和业务模式,随着金融机构信息化、网络化和全球化改变了金融机构的经营方式,金融网络中资金流动呈现速度快、数量大、创新多等特点,大量的资金在金融系统中流动着以满足正常的经济活动需求,但也夹杂着许多不正常、违规违法的资金流动,怀揣各种想法的资本所有者在金融网络中施展百般武艺,逃脱监管,达到资金流动的目的,给金融监管部门带来了空前的挑战。本研究拟建立科学系统的分析、识别可疑资金流动的智能化的金融监管模式、模型、方法,为金融机构和金融监管部门提供有效的决策支持工具。本文研究基于以下思路:基于约束理论的反洗钱业务流程瓶颈研究――>基于正常模板的资金流动异常识别方法研究――>基于扫描统计的单帐户资金流动异常识别研究――>基于序列匹配的单帐户资金流动异常识别研究――>基于社会关系网络的多帐户资金流动异常识别研究。本文主要工作和结论如下:(1)文章在约束瓶颈(TOC)理论的指导框架下,通过深入访谈反洗钱各相关部门,从整体上勾勒了目前反洗钱业务实施实际的整体框架图,理清了各个相关部门在开展发洗钱业务时的主要依据、数据的扭转方式、内部控制机制等;并进一步从物理约束、政策约束和行为约束三方面总结归纳了存在于金融机构和监管部门的主要反洗钱瓶颈。(2)基于正常模板的资金流动异常客户分析一章节,在总结分析影响企业经营活动的若干因素,如宏观经济环境、行业特征、企业规模和地区差异之后,利用统计模式识别的方法从资金流动交易金额分布、资金流动的时间特征、资金流动地区特点、资金流入、流出对比分析、资金流动间隔频度分布等多个纬度建立正常模式的相关模型。(3)基于扫描统计的单帐户资金流动异常识别研究一章,研究根据扫描统计相关原理,将资金流动过程中的异常识别问题转化为扫描统计研究问题。并结合上一章节的正常模式,设计利用扫描统计监测帐户资金流动异常的算法。实验结果表明,算法对甄别帐号交易过程中短期内的异常资金交易行为十分有效。能够大大降低监测的第一类错误,即降低漏报率,然而,在降低第二类错误(误报),提高算法敏感性方面仍然需要做进一步改进。(4)基于序列匹配的单帐户资金流动异常识别研究章节,研究充分利用金融机构在反洗钱识别中的主要信息源:客户信息、帐户信息和交易信息,试图完成资金识别的最终目的:分类正常与异常交易。算法基于序列匹配相关原理,建立识别问题对应的查询序列-高风险交易片段,参考序列-帐户的历史交易记录和同组帐号的交易记录,并建立相似度核-基于欧几里德距离和余弦法的相似度计算方法,最终根据分类阶段阈值识别异常识别。(5)最后,基于社会关系网络的多帐户资金流动异常识别研究,文章首先研究了基于社会关系网络相关理论描述多帐户间资金网络构建方法,然后从金融监管实际中隐藏网络的实际问题出发,建立基于隐藏网络分析的多帐户资金流动网络中隐藏网络的异常识别方法,并根据实验验证隐藏网络识别算法的可行性。文章的主要创新工作可以归纳为以下几个部分:第一,建立了基于约束理论的反洗钱业务流程瓶颈研究方法,为我国其他金融监管问题提供可参考的理论依据与研究方法。在对现有的金融监管问题的研究中,国内外学者多以‘点’切入,单就业务流程环节的某一单一环节的单一问题进行研究,提出监管建议。然而,这样的研究方法不利于从整体上把握影响监管效率的关键因素。本文通过引入供应链领域的‘约束理论’将反洗钱监管监管效率看作由监管领域上下游各环节间、各部门共同协作,共同决定的。通过从业务环节的各个环节之间的信息扭转、职责明确,以及各环节的主要工作重要和工作环境分析,从‘面’上描述反洗钱业务的全貌,同时又有针对性的从物理约束、政策约束和行为约束多方面总结归纳影响各关键环节的主要因素。第二,提出了基于正常模式识别异常的监测模式,为我国监管部门进一步完善大额可疑报告细则提供理论依据,也为金融机构抽取可疑资金报告提供新的思路。国家目前已经针对反洗钱金融犯罪制定了《金融机构反洗钱规定》和《金融机构大额交易和可疑交易报告管理办法》,然而在金融机构执行的可疑报告抽取的时候,多反映在实施过程按照管理办法中的条例“过于模糊”,“难以量化”,给实际监管工作带来困难。另一方面,目前我国的商业银行用于甄别洗钱活动的决策模型主要是基于固定规则的,其效率相对低下,存在大量的误判错判,并且犯罪分子可以通过简单的规避和反侦察手段逃过监管。本研究提出基于正常模式的异常识别监测模式,该模式可以根据行业、监测力度动态调整可疑标准,使得犯罪分析不能通过简单的规避手段逃过监管。另一方面,采用智能化的监测手段辅助可疑报告抽取工作,使得监测效率可以大幅提高,并在很大程度上将监管专家从人工识别的繁杂劳动中解放出来。第三,建立了基于社会关系网络分析的异常群体识别模型,使得利用客户之间的交易关系识别异常成为可能,为完善风险可控的监管平台提供新的监管思路。金融网络中对手方交易信息对于追溯资金的来源和识别资金的目的地发挥着重要的作用,传统的监管条例多单一地围绕客户的直接交易方,而犯罪团体可以在一定程度上通过借助空壳公司、离岸公司或者各种金融服务公司作为媒介,增加资金交易的路径,以掩盖其真实的目的,因此,迫切需要建立能够有效识别金融网络中隐藏团体的抽取模型。本研究利用社会关系网络和图论的相关理论的建立金融网络中隐藏团体的抽取模型,抽取客户间交易关系代表的社会关系,并试图通过智能化的监测手段,甄别隐藏在正常金融交易行为中的隐藏团体。第四,从网络层、客户层、交易层三个层次系统建立了适应于金融网络信息分布特征的监管体系,使得各个算法能够有针对性的服务于不同的监管主体和监管目的。根据《金融机构反洗钱规定》和《金融机构大额交易和可疑交易报告管理办法》,金融机构有义务向监管部门上报金融网络海量数据中的可疑交易和可疑客户。监管部门接收来自全国各地不同金融机构的报告,并进一步分析转移这些高度可疑的交易和客户,同时还有责任从宏观层次上分析当前最新的异常模式和资金流向规律,不同主体的不同监管职责和信息分布特征决定了我们需要有针对性的设计满足各自需要的监管模型。本研究根据不同主体的信息分布特征,设计满足于其自身需求的监测模型,同时又综合考虑模型之间的关联性和兼容性,使得从网络层、客户层、交易层三个层次建立的监管体系能够行之有效的服务于当前金融监管实际。更多还原

【Abstract】 Finance is one of the main pillars in modern economy, and the stability of financial system is the major tasks and core responsibilities for the world’s financial supervision authorities. Financial information technology has changed the traditional financial system and business model: the financial institutions become electronic connected, with more complexity and tend to be globalization; the capital flows within the financial institution are filled with large volume of transactions with high speed and showed as innovative forms. Large amount of money are flowing in the financial system to meet the needs of normal economic activities, but also few abnormal, illegal capital flows are carring a variety bad ideas and intent to escape regulation, which brought big challenges to supervision authorities. This study aims to establish a scientific and systematic analysis method to effectively provide intelligent decision support systems for financial institutions and financial regulators.This paper is organized as the following orders: first we study the main constrains and bottlenecks in anti-money laundering business process based on Theory of Constraints (TOC), then we conduct research to identify the abnormal customer in capital flow based on normal patterns; then we use Scan Statistics to identify the suspicious transactions from a single customer angle; and we establish a suspicious transaction identification method using sequence matching based algorithms; also we establish a social network based method to identify hidden groups in financial networks. In this paper, the main work are as follows:(1) We conduct a in-depth interviews under the framework of Theory of Constrains (TOC) to study the main constraints and bottlenecks in the AML organizational system, we surveyed those relevant agencies, figure out their AML responsibilities, their operations and business process, the information distribution they have for detecting suspicious reports, also the way transferring valuable data. Also we conclude the main AML bottlenecks existed in financial institutions and supervision agencies from three aspects: physical constrains, policy constraints and behavioral constraints.(2) We establish a suspicious customer identification method based on normal pattern. First we analyze the main factors which affect the capital flow of a company; those factors include the macroeconomic environment, the industry characteristics, the firm size, and the regional differences. Then we use statistical pattern recognition to analyze the distribution of capital flow from different latitude: transaction amount, time characteristics, regional characteristics, capital direction, and interval of the sequent transactions.(3) We also use scan statistic method to identify suspicious transaction activity in single account angle. Based on the principle of Scan Statistics, we transfer the identification of suspicious transaction into a scan statistics problem. In conjunction with the previous normal pattern model, we design a scan statistics based monitoring algorithm. The Experiment results show that the proposed algorithm can effectively detect the abnormal behavior in short observing period. Results also confirm that this intelligent algorithm can largely reduce the Type I error, which is also to say that it can effectively reduce the omitting rate. However, we need to find way to further reduce Type II errors (false positives) and improve the sensitivity results of the algorithm.(4) Based on sequence matching algorithm, we establish an algorithm to identify suspicious transaction using the main sources in financial institutions: customer information, account information, and transaction information. In order to achieve the final goals: classify normal transactions and suspicious transactions, we collaborate sequence matching algorithm, and establish the query sequences in this problem- high-risk transaction fragments, the reference sequence- the history of the query customer and also transactions of other customers in the same peer group. Also we use Euclidean distance based and cosine based similarity kernel to calculate the similarities between query sequences and reference sequences. The final stage of this algorithm label the query sequences and classify them into normal and abnormal based on given threshold.(5)Finally, we propose an algorithm to find hidden groups in financial networks based on social network theory. The paper first discuss the construction of a financial network using social network theory, and then bring up the practical issue of hidden group in financial supervision. Then we establish a model to identify hidden group in financial networks for supervision center. The experiment results show that this model is feasible in identifying small gangs or extracting related criminal hidden groups from a given nut.The main innovative ideas can be concluded as following:First, the method of identification of the main constraints and bottlenecks in AML organization system provide us the valuable theoretical basis to study other similar supervision problems.In previous research about current financial regulation issue, many scholars in and abroad solely study single point of the issue or focus on single link within business process chain. However, such research can hardly grasp the overall picture of a complex problem, or find the key factors which affect the supervision efficiency.This paper views the AML problem in an overall picture by surveying the upstream and downstream within the whole AML business process. Through analyzing the AML responsibilities for each AML related agencies, their operations and treatment process, the information distribution they have to detect suspicious reports, also the way to transferring valuable data, we draw the whole picture for anti-money laundering business process. We also summed up the key constraint from physical, policy and behavior angles to identify the most influential factors on AML efficiency.Secondly, we propose a pattern to identify suspicious behavior by comparing normal pattern, which can provide insight to improve Administrative rules for the reporting of large-value and suspicious activity reports, and also it explores new direction for suspicious detection.When reporting large amount or suspicious transactions according to Administrative rules for the reporting of large-value and suspicious activity reports, the financial institutions encountered problems of‘hard to quantify’, which brought difficulty to monitoring and detecting work. On the other hand, current extracting tool in financial institutions are mainly rule based system, with low efficiency and the criminals can easily escape the regulation by simply learn from the regulation rules.We proposed intelligent algorithm which can easily adjust the efficiency by changing different parameters, it makes the criteria difficult to escape. On the other hand, using intelligent tools, the type I and type II error can be reduced so that we can free the experts from huge labor.Thirdly, we incorporate the customer relationship information into detection, and propose algorithm to identify hidden groups in financial networks. This provides a new idea to regulation platform.The source and destination information of a transaction played important role in suspicious detection. However, traditional regulatory solely focus the direct transaction partner in suspicious detection. This encourages the criminals to use offshore company or variety financial services as medium to veil these illegal activities. Finding hidden groups within large financial networks is of urgency.In this study, we construct financial network using social network related theory, we extract networks which describe transaction relationship, and try to access the hidden groups in financial networks using intelligent monitoring tools.Fourth, we establish suspicious extracting tools from three layers: network layer, customer layer, and transaction layer for financial institutions and supervision agencies, according to their information distribution.Base on different responsibilities and information distribution of financial institutions and supervision agencies, we design intelligent tools which can exactly need their requirements, separately. The monitoring system based on algorithms from these three layers: network layer, customer layer, and transaction layer, can effectively serve the current financial regulation practice.更多还原