We have recently witnessed many spectacular losses in the financial industry. Firms that had been performing well suddenly announced large losses due to credit exposures that turned sour, interest rate positions taken, or derivative exposures that may or may not have been assumed to hedge balance sheet risks.
Perhaps the most striking example is the collapse of Barings Brothers, a highly respected 230-year-old British merchant banking firm. Their failure is a direct result of the trading losses from unauthorized speculative trading activities of a single trader (Nick Leeson) in the firm’s Singapore office [12]. A report by Britain’s Board of Banking Supervision indicated no one within Barings management, nor any of their auditors nor regulators, were aware of the vast amount of risk undertaken by the transactions. In a similar incident, the 130-year-old Wall Street firm, Kidder, Peabody & Co. went out of business after its chief trader covered up losses amounting to $100 million. The problem underlying many such financial disasters was best summed up by economist Henry Kaufmann in Business Week: “… these (financial) problems reflect inadequate monitoring and supervision.”
We have seen a dramatic increase in today’s financial institutions to acquire and utilize financial monitoring systems such as the risk management systems marketed by vendors such as C*ATS, Algorithmics, and Infinity. Within current financial monitoring systems, the pressing need is for a common understanding regarding the financial information involved. Due to the complexity and the myriad amount of this information, a serious challenge has been posed for financial monitoring systems to integrate an accurate and timely date on every financial transaction executed by an institution.
It is our belief that intelligent agents are well suited to dealing with the problem of monitoring vast volumes of dynamic information in a distributed fashion. In this way, they are able to detect hidden financial problems, such as financial fraud, unhedged risks, and other inconsistencies. By utilizing a society of intelligent agents, each charged with carrying out a different function autonomously, financial monitoring systems will not only be able to analyze financial transactions qualitatively, but will also deduce useful information regarding the state of current investments. Although there is no need for centralized storage of all knowledge regarding an organization—its financial instruments and the transactions involved—there must be one consistent database of knowledge that enables the various agents to exchange knowledge regarding the entities involved. Subsequently, the various agents that make up the system can perform the monitoring and analysis involved in a collaborative manner.
We present a lightweight, distributed, intelligent agent-based financial monitoring system that monitors and reports on transactions within an organization. The intelligent agents in our prototype system are assisted by a formal conceptual model that makes up an unambiguous understanding of the institution, the transactions, instruments involved, and the business processes. Our prototype system adopts software agents to define open and flexible distributed architecture by utilizing a knowledge repository to maintain and integrate all information relevant to the monitoring and reporting tasks. A major objective of the study was to demonstrate both the technical feasibility and the appropriateness of intelligent agent techniques for global financial monitoring. Based on our practical experiences, we will delineate important concepts, including the model, the technologies, and the architectural considerations required for developing such a society of intelligent agents that together will assist in the monitoring of finances within a global organization.
With the rising tide of financial fraud and the unmonitored undertaking of financial risks, the reinforcement of monitoring and the supervision of financial transactions has become central to the global financial market [9]. Institutions are enforcing stricter rules to ensure the monitoring and the supervision of financial transactions, such as those described in [7], which aim to curb financial activities that will pose significant danger to the organization as a whole:
- Improve the financial reporting environment, and thus reduce the incidence of fraudulent financial activities.
- Improve auditing standards, the standard-setting process, and the system for ensuring audit quality that will increase the likelihood of detecting fraudulent financial activities.
- Enhance regulatory and law enforcement to strengthen deterrence.
- Enhance the education of future participants in the financial transaction process.
Our work consists of the natural integration of two fields of computing techniques. The field of conceptual modeling provides our research with a formal framework with which knowledge regarding the problem domain can be specified in a concise and unambiguous manner. On the other hand, the recent flurry of research into the application of intelligent agents has resulted in providing a distributed, intelligent society of agents capable of dealing with vast amounts of information collaboratively. Consequently, it is our view (and this is shared by an increasing number of intelligent agent researchers) that one of the significant building blocks of intelligent agents is the need for an unambiguous, conceptual model of the information involved.
The concept of “agent” has become important in artificial intelligence, computer science, and e-commerce [3, 10]. Agents are used to denote a software-based computer system that possesses the following properties: autonomy (agents operate without the direct intervention of humans); social ability (agents communicate with other agents); reactivity (agents perceive their environment and respond in a timely fashion to changes that occur in it); proactivity (agents do not simply act in response to their environment, they are able to exhibit goal-directed behavior by taking the initiative); mobility (agents are able to travel through computer networks. In general, an agent on one computer may create another agent on another computer for execution. Agents may also transport from computer to computer during execution and may carry accumulated knowledge and data with them).
There has been a recent accretion of approaches for building multiagent systems for Internet applications [4]. Multiagent systems are generally computational systems in which several semiautonomous agents interact to perform some set of tasks or satisfy a set of goals [6]. Researchers have proposed to design and develop numerous intelligent agent-based systems to support business processes. Lee [5] used an intelligent agent UNIK-AGENT in contract processes. The messages in this agent-based commerce adopted three layers: Agent Communication Language, Electronic Commerce, and Product Specification. Barr et al. proposed using an intelligent agent for business data monitoring [2]. In the prototype system, intelligent agents will intermediate on behalf of business analysts by being able to perform limitless, error-free routine calculations and interpretations rapidly to the precise requirements of business managers. Dugdale [3] proposed the development of a cooperative problem-solving system for the domain of investment management. He proposed seven functions that a cooperative system should provide during knowledge acquisition and an assumption-based truth maintenance system (ATMS) used to store the solutions, constraints, and underlying decisions in each agent’s portfolio. Allouche et al. [1] proposed a multiagent system for the supervision of dynamic systems by using a temporal scenario recognition approach. The supervision is performed by a society of agents where an agent is considered a watching process responsible for a subset of possible scenarios of the way the system functions.
The development of our conceptual model for the agent-based financial monitoring system has drawn on results from data modeling, business modeling, and process modeling. The basic problem of modeling is the development of an expressive presentation notation with which to represent knowledge [8]. The scheme used to model the problem domain of the financial monitoring system is based on the Unified Modeling Language. UML adopts a representational framework that includes structuring mechanisms analogous to those offered by object-oriented data models, namely classification (inverse instantiation), aggregation (inverse decomposition), and generalization (inverse specialization).
A Conceptual Model for Financial Institutions
Capabilities of our integrated modeling framework include metalevel facilities as well as a formal classification of domain knowledge. This framework provides the conceptual basis for thinking about such institutions and it provides a formal basis for tools and techniques used in development and usage. Our model contains various basic entities represented as objects with specific properties and relations. Objects are then, in turn, organized into taxonomies. Definitions of objects, attributes, and relations are specified in UML. An ontology is then defined as follows: The objects in the domain of discourse are first identified; classes and instances in UML will represent these objects.
In order to model financial institutions, it is essential to understand their information structure. Figure 1 contains a small subset of the entire conceptual model for the typical financial firm. Such a subset is sufficient to demonstrate the framework from which an ontology of the financial markets may be constructed.
There are three different types of conceptual links in our model. A triangle hollow arrow represents a link between a subclass and its subclass. For example, Management Staff and Operational Staff are two subclasses of Staff. A spear arrow represents a containing link, for example, Management Office, Back Office, and Front Office are contained in Branch. A solid line represents an associate link. For instance, Trader works at Front Office.
It is important to note our conceptual model represents a logical model for the classical financial institution. An actual financial institution will have objects that correspond to categories in the logical model eliminated in the figure.
Trading is a dynamic process exhibiting changes through time. There are four steps to a futures transaction:
Stage 1. Decision-making. The general manager makes a decision on which kind of contract he or she will buy or sell due to the change of the contract market, then he will ask a trader to do the actual trading at the front office.
Stage 2. Trading-implementing (front office). When the trader gets the request from the manager, he or she will implement the actual trading with other traders in the market, for example, on Jan. 23, 1995, trader George bought 6000 Nikkei 225 Index contracts at the price of ¥19000 for the account (shown in Figure 1). After the trade, he should report his trading situation to the manager and pass the trading data to the clearer.
Stage 3. Trading-clearing (back office). When the clearer gets the closed price of the market, he or she will calculate the position and profits and losses of the trading done within that day and pay for the invariant margins to the market center.
Stage 4. Trading-reporting. The clearer will report all trading records and data to manager and director after he or she finishes the calculation, and the manager should report the trading situation to the director and senior supervisors. The stages mentioned here are the regular steps for futures or optional transactions in this kind of financial institution (shown in Figure 2).
From the model of trading process, we understand a transaction operation is a series of processes executed by different offices. If we want to monitor the trading transaction, we need to monitor all activities during the evolution of the whole process, that is, a continuous real-time monitoring is required. Hence, fraud can be detected and recorded at any time and at any place, it is effective to avoid the failure or risk in advance. The sequence diagram of the real-time monitoring process is shown in Figure 3. The problem that arises from continuous monitoring is who will undertake the monitoring.
Assessing trading risks is a cumulative process encompassing individual and collective risk factors, some of which are the general state of the economy; the ethical and morality levels in society and in the overall organization; employee loyalty; turnover and general sense of job security; the past trend of fraud cases inflicting the entity and its industry; and the strength of the justice system in dealing with fraud cases. Due to these risk factors, it is difficult for human beings to monitor financial transactions continuously without making errors. When influenced by difficult working relationships, strict codes of conduct and money, employees can neglect their duties. For instance, if managers or traders carry out financial transactions according to the highest risk averse model, it may benefit them only in the short term. Thus, a lapse in monitoring ensues because the monitor or supervisor wants to gain maximum compensation without considering long-term results and benefits.
To deal with this problem, we extend the concept of an intelligent agent to a design monitoring system. The concept of agent has become increasingly important in both artificial intelligence and computer science based on its features and functions, namely proactivity, cooperation, intelligence, and reliability [8, 10].
System Architecture Design and Proof of Concept
To design the architecture of a multiagent application, there are a number of components to the overall multiagent architecture, including a class of agents, a set of agent’s autonomous behaviors, and a knowledge repository. The taxonomy of intelligent agents is shown in Figure 4.
Interface agents enable users to view the current state of the trading and monitoring processes and allow them to convey their own opinions and arguments to the rest of the institution. Agents also enable the corresponding users to issue requests. Naturally, there are a number of user interface agents, such as Manager interface agents, Trader interface agents, Clearer interface agents, among others.
Data collection agents enable the system to collect data internally and externally. Monitoring agents monitor the system’s operation. Such monitoring tasks involve fraud detection, credit risk monitoring, and position risk monitoring.
Generally, the design of an agent should consider an agent knowledge base, its operational facilities, and its external interface facility. It is essential to design a set of autonomous types of behavior for the agent class, including reactive, proactive, and cooperative behavior. The Reactive behavior of an agent allows the agent to have actions based on other agents’ requests. For example, a Manager interface agent may ask a Trader interface agent to report a particular trading activity and the Trader interface agent reports such activity “reactively.” The Proactive behavior enables an agent to act based on its goal. For instance, the goal of a Position risk agent is to evaluate the risk in the currently held positions due to changes in market factors (such as interest rate, currency exchanges, and commodity volatility). Subsequently, a Position risk agent may ask an External Data agent (which would behave reactively to the goal) to collect and summarize any movements in the U.S. Treasury Bond yield in order to effectively monitor any financial risk that may result from movements within the bond market. Cooperative behavior enables an agent to cooperate with other agents. An agent may use knowledge assertion to initiate a goal to other agents, or use knowledge query to query other agents.
A knowledge repository can provide a centralized representation. Consequently, knowledge relevant to the management, evolution, operation, and maintenance of agents can be shared in a transparent fashion. The central repository may provide management for knowledge manipulations. A knowledge repository is a database of specifications—it contains what is commonly referred to as “metaknowledge,” or knowledge about knowledge. In providing metaknowledge, the repository provides an opportunity for agents to deal with their interoperability problems.
In order to evaluate our architectural design, a small prototype has been implemented and evaluated. The prototype system carries out the analysis and monitoring using simulated financial data based on a small number of intelligent agents.
Within our prototype, a Monitoring agent performs continuous monitoring of the activities of trading agents automatically during the trading process by recording, detecting, and analyzing the trading data. The monitor agent not only obtains all relevant information from contract objects of different agents, but it links to other classes of Warning Rule to check which rule is being violated and to send a warning message to the supervisor. The monitor is the real agent in our system: it has all the features of an agent and can execute its operation automatically. It can communicate and cooperate with other agents and give different responses to environmental change. The Monitor agent is instructed by the financial supervisor during initialization to focus on a particular type of transaction. Subsequently, the agent performs its actions continuously and accurately until the supervisor asks it to stop or change its goal.
The common software used in the implementation is Java. However, each agent may integrate Java with other necessary tools. For instance, the Monitoring agent integrates Java with JESS; the repository is implemented on the top of the commercial OODBMS ObjectStore.
Utilizing our proof of this concept system, we were able to perform numerous simulations of financial monitoring cases based on significant historical events. The following example demonstrates one of our test cases. It simulates two months of main transactions by Nick Leeson in his Singapore branch. On February 15, 1995, manager Leeson wanted to buy 6,000 Nikkei contract at the delivery price of ¥18,000 and asked the trader, George, to process the trade. George did it based on Leeson’s request, and Risselle did the clearing. After her calculation, the total position for Nikkei was 43,320, and the loss was ¥14.65 billion. In this example, our system gave three warning messages: the first one is “unauthorized trade warning,” because the trading amount 6,000 is over the limit. The second one is “position warning,” because the contract 43,320 is over the limit. The last one is “dangerous loss warning,” because the total loss (¥14.65 billion) is over ¥10 billion. Also, the supervisor agent sends a message to ask the manager to stop the trade.
From the preliminary results collected from those experiments, we were able to conclude the advantages of our intelligent agent-based financial monitoring architecture as follows:
- Our experiments demonstrated the benefits of a lightweight architecture, namely that such agents will provide an efficient and economical service, are compatible with existing societies of agents, and can immediately participate in the overall system of the operation. It will be possible to add and subtract agents from the society of agents based on the complexity of analysis being undertaken or on the consumption of the system’s resources.
- Our system provides the benefit of scalability in terms of being capable of dealing with the complexities and the sophistication of today’s financial institutions. Since the knowledge exchange between the various agents within our systems primarily consists of qualitative observations regarding the financial transactions, a society of agents of this kind can be programmed to handle the increasing size and complexity of the financial instruments and transactions monitored.
- By utilizing a formal conceptual model, the intelligent agents within our system are capable of carrying out unambiguous and concise communication as facilitated by a common knowledge repository. Common knowledge can be directly injected into the repository to enable more complex and meaningful forms of knowledge communication.
- We observed that by simply basing the conceptual model on a different financial instrument or institution, an entire intelligent agent can be reused for an alternative application domain. In fact, we have different courses of research under way that reuse our monitoring agents to monitor Internet security intrusions. Our agents were able to use data collections largely based on Internet traffic analysis rather than financial transactions as presented here.
Conclusion
Our research has introduced lightweight intelligent agents into financial monitoring systems. The intelligent agents will provide an effective means for systematic monitoring of financial transactions in the corporate world, to detect and report any abnormal financial transactions that may signify unhedged risks, fraud, and other financial inconsistencies. We have proposed a framework for constructing ontologies based on careful analysis and abstraction of those financial transactions using UML. Such ontologies provide a common vocabulary for all knowledge communication within the system. The application of ontologies can lead to unambiguous understanding of the concepts, and pinpoint the likely cause of confusion. Furthermore, such models provide a uniform framework with which different systems can be compared in terms of their individual strengths and weaknesses.
To demonstrate both the technical feasibility and validate our initial theoretical inclinations, we carried out a detailed architecture design from which a prototype proof of the concept system was developed. Throughout our design, we have generalized our experiences to outline some preliminary facets of intelligent agent architecture, that is, supported by the ontologies. The preliminary results of our proof of concept have been extremely promising, illustrating the extensibility and scalability of our intelligent agent-based architecture.
It is our hope that our work will be of value to practitioner and academic alike. Academic researchers for intelligent agents would benefit from our research since they would gain insight from an application of their techniques in a large distributed manner. Conceptual modeling researchers would benefit from having a clear and concise model based on their work to support our intelligent agents. In the near future, we aim to continue our research by extending our results into other application domains. We wish to validate our lightweight monitoring agents to operate in a large and complex financial domain that mirrors real-world financial institutions. It is our belief that intelligent agents will play a significant role in tomorrow’s organizations, and future businesses would benefit immensely from open, distributed agent-based, enterprise-computing architecture.
Join the Discussion (0)
Become a Member or Sign In to Post a Comment