STRUCTURING AND MODELING DATA FOR REPRESENTING THE BEHAVIOR OF AGENTS IN THE GOVERNANCE OF THE BRAZILIAN FINANCIAL SYSTEM

Abstract

This article proposes a method to collect and structure data in order to model the behavior of agents in an agent-based simulation model. This model aims to study the regulatory governance of the Brazilian financial system. In this article, the regulatory governance is understood as a phenomenon that results from several interactions and transactions among all actors that influence or are influenced by the activities of the regulation policies. The study focuses on the short-term interest rates and incorporates behavioral aspects and no-explicit interests of social and economical agents. The developed method integrates content analysis research and in-depth interviews to model the agent' behavior by means of fuzzy logic rules. It provides systematic collection and interpretation of data produced in textual form as well as knowledge from experts. The results of the model validation have shown that this approach contributes to the development of a methodology for the modeling and representation of agent behavior in social applied simulation models.

Regulatory governance of a financial system is defined, following the governance framework proposed by Kooiman (2003), as a phenomenon that results from the diverse interactions and transactions existing among all the agents impacting or being impacted by the regulatory activities of the financial sector. From this perspective, governance arises from the interaction among social, political, and economical agents (individuals, organizations, and institutions). According to Kooiman (2003), the concept of interaction lies at the core of the governance perspective. Thus, it is understood that the regulatory governance of a financial system (FS) can be described as a complex system, encompassing a large number of highly interdependent and interconnected entities, subject to regulatory policies in a market environment. From this perspective, governance can be studied using the notions of complexity and emergence phenomena concepts developed by the general systems theory.

With the aid of agent-based simulation, the emerging patterns of the regulatory governance of an FS can be explored over time. According to Edmonds (2005), in cases where the field of study is related with probabilistic human complex systems, it has been proven to be impractical or even impossible to rely only on mathematical models. Therefore, the construction of an agent-based model appears to be the most suitable way to assess the impact of social and political actors on the governance of an FS.

Figure 1 depicts a general view of the modeling process of an agent-based model. The focus of this process is on projecting the agents and their interactions in order to incorporate relevant aspects of the system under consideration. The conceptual model plays a fundamental role in the modeling process, since it is in this representation level that behaviors and intentions of the agents in different operational contexts are generally defined. Therefore, the conceptual model is a particular abstraction of the real world aimed at incorporating the assumptions made on actor' behavior and intentions. This article focuses on the process of building a conceptual model of a multi-agent system.

FIGURE 1 The modeling process.

The most relevant data for the development of a conceptual model are, with rare exceptions, of qualitative nature, as highlighted by Luna-Reyes and Andersen (2003). These authors discuss in detail the importance of qualitative data for the development of complex dynamic systems. While these models are, in general, represented mathematically, the authors highlight that, admittedly, most of the information available for the model developer is not numerical, but rather, qualitative. It is noted that, specifically in the development of agent-based models of complex social and economical systems, qualitative data are present in all stages of the modeling process.

The financial sector, in particular, can be characterized as a complex system comprising a large number of interdependent actors, with different types of behavior and interactions (Van den Berg et al. 1004). Each one of these actors experiences pressure from the environment, other actors, sectors, or groups that influence their behavior. Governance is a phenomenon that results from diverse interactions among all the sectors which impact or are impacted by the regulatory activities of the financial sector. Thus, it is of critical importance to identify the actors in the financial system, together with their interests, attitudes, and patterns of behavior including their interactions with others. While there are several mathematical models that attempt to represent the behavior of financial systems, little emphasis has been placed on the analysis of the governance of FS. Moreover, the data available for modeling the governance of an FS are found either in textual form or in the minds of financial market experts, that is, the data for the modeling are of an eminently qualitative nature.

Notwithstanding the importance of qualitative data for the development of agent-based models, no research work presenting formal procedures for the survey and organization of such data and the ensuing modeling of the agent's behavior was identified in the literature. The main objective of this research is, therefore, to introduce a method for the collection and structuring of qualitative data that can be incorporated into the development of agent-based models. The method developed makes use of content analysis and in-depth interviews as the research tools. The method was applied in the analysis of the regulatory governance of the Brazilian Financial System (BFS), in contexts relative to the exercise of interest rate policy. The outcomes obtained attest to the feasibility of applying these research techniques in designing an agent-based model, helping to define relevant agents and their behavior.

BACKGROUND

One of the major concerns in the process of developing agent-based models is the nature of the agents themselves and the definition of their behavior, that is, how they interact with other agents and with their environment (Edmonds 1998). The author argues that the purpose of modeling the agents is to reveal the emergent behavior of the system. In the literature, it is possible to identify diverse types of agent frameworks that have been conceived for various types of analyses. The agent model used in this study is based on the beliefs-desires-intentions (BDI) architecture. Beliefs-desires-intentions has been used for the modeling of different types of agent behavior, and adopted in numerous fields. Beliefs-desires-intentions agent architecture was introduced by the philosopher Bratman (1999), who proposed a framework for understanding ways of characterizing mental attitudes and rational actions in human beings, in terms of their intentions. The principles of Bratman's work have been fundamental for the theoretical formalization of computational agents with rational behavior, and for the development of formal agent architectures. In BDI, agents are described as a set of beliefs, desires, and intentions. The agent' decision-making process occurs during analysis of beliefs relative to their desires. Beliefs are items of information held by the agent about himself and about the environment in which he is active. They correspond to the informative component of the agent' status and may be subject to uncertainties and different perceptions regarding their believes about the same object or phenomena. Desires, in turn, are objectives the agent adopts and attempts to achieve. In terms of BDI architecture, an agent's desires are essentially the “options” or “possibilities” available to the agent (Wooldridge and Parsons 1998). The theoretical model of the BDI architecture also employs the concept of intentions, which represent courses of action chosen by the agents to achieve their goals (desires). The agent's actions are organized into plans. In the process of internal deliberation, after the selection of an intention, an agent's plan is chosen and initiated. Thus, intentions correspond to the agent's plans under execution.

Since the relevant BDI of agents in the regulatory governance of an FS dominion are of a subjective nature, the specification of the agents in this research employed a fuzzy-extension BDI agent model (Shen, O'Hare, and Collier 2004). Fuzzy logic provides a general method for handling uncertain and vague information, which are unfortunately unavoidable in many real-world, decision-making processes. Fuzzy logic is in this sense, a superb tool since the complexity of real-life situations is handled through “vague” variables and “vague” interactions, which better replicate the human mind in describing phenomena (Zadeh 1973). The basic idea behind the use of the fuzzy extension for modeling multi-agent systems is the specification and description of the agent behavior by means of fuzzy rules. The inference of these rules can be understood as the mapping between a set of inputs and a set of outputs. The practical reasoning of the agent consists of two principal activities (Shen et al. 2004; Schut, Wooldridge, and Parsons 2004): (i) deliberation, where the agent decides what to do (which intention to carry out); and (ii) planning, which is the decision of how to carry out the intention. Thus, the inference of these rules during simulation establishes the dynamic behavior of each agent in the system, and as a consequence, the behavior of the system as a whole.

In order to simplify the model proposed, the agent's deliberation (what to do) and planning (how to do it) processes have been combined into one process. In this case, the agent's practical reasoning mechanism consists of choosing a pair ⟨objective, plan⟩ for execution, that is, the intention the agent can adopt and the plan of actions for carrying out such intentions (intention selection module). This simplification was suggested in the work of Hsieh, Liu, Yu, and Hsu (2004). The intentions are the agent actions that are executed based on fuzzy rules (fuzzy-inference system module), after the beliefs and desires were processed by the agent. Figure 2 represents the internal model proposed for the agents and indicates its principal components (Streit 2006). These components can be described as follows:

	Perceptions: refers to the means by which the agent perceives the environment.
	Agent status: refers to the agent's current set of beliefs about its environment and by the intention it is currently pursuing.
	Database pair ⟨objective, plan⟩: data structure storing the possible space state of an agent's pair ⟨objective, plan⟩.
	Database “Beliefs”: stores the agent's beliefs about its environment.
	Components “revises beliefs” and “selects ⟨objective, plan⟩”: they are components that carry out the procedures for the selection of the agent's intentions and action plans. These components constitute the agent's decision-making process, along with the component “deliberation control.”
	Action: component that executes the actions for carrying out the current intention or the new intention selected by means of the fuzzy logic.
	Action outlet: refers to the means by which the agent transmits messages to the environment and to the other agents. It is the outlet for the outcome of the agent's inference process.

FIGURE 2 Agent's internal model.

As presented in Figure 2, the process of carrying out the intentions is based on fuzzy logic. This process is circumscribed by the component denominated “fuzzy-inference system.” In this case, the fuzzy rules execute the agent's actions following the BDI framework. The agent's beliefs are defined on the antecedent term of the fuzzy rules (IF side), while the term relative to the agent's deliberation is found on the consequent side (THEN side). For instance, the fuzzy rule “IF high inflation AND inflation variation increases THEN exert moderate pressure for reduction in the interest rate” indicates that if there is a belief on the part of an agent, that inflation is high and that there is a trend towards increasing inflation, this agent's deliberation may be to exert a moderate pressure on the agent “monetary authority” for an interest rate reduction. The value resulting from the pressure will depend on the degree of pertinence of the variables “inflation” and “inflation variation” to the fuzzy sets “high” and “increases,” respectively.

In addition to the work by Schut et al. (2004), other studies in the literature demonstrate the advantages of using fuzzy logic in the development of agent models (Bossomaier, Amri, and Thompson 2005; Li, Musilek, and Wyard-Scott 2004; Hsieh et al. 2004; Shajari and Ghorbani 2004). Fuzzy logic has been employed in the agent decision-making process and in the definition of agent behavior. However, these studies do not explore or demonstrate how the fuzzy rules, which define the agent behavior, are constructed. This study attempts to fill this gap, by proposing a method for capturing and organizing qualitative data when modeling agent behavior, mainly when considering the application of fuzzy-inference mechanisms. Before introducing the method, the next section consists of a review of the research techniques employed in collecting and structuring data for the development of agent-based models.

LITERATURE REVIEW

Modeling the agent behavior depends on available knowledge about the real-world dominion under analysis, amenable to being captured for the construction of the model. To the best of our knowledge there is no formal agent behavior modeling methodology for the development of multi-agent models described in the literature. Some researchers have employed knowledge elicitation techniques, developed in the context of artificial intelligence, specifically expert systems. According to Moody, Blanton, and Will (1998–1999), various elicitation techniques have been recommended for the development of expert systems, the most frequently employed being: (i) protocol analysis; (ii) repertory grids; (iii) multi-dimensional scaling; (iv) card-sorting; (v) interviews; and (vi) case study reviews. The process of knowledge acquisition, notwithstanding the existence of diverse subjective elicitation techniques, has so far proved to be a major challenge for knowledge engineers (Moody et al. 1998–1999). In the expert systems literature, knowledge elicitation has often been cited as the main bottleneck in developing such systems (Burton, Shaboldt, Rugg, and Hedgecock 1990; Holsapple, Rai, and Wagner 2008; Kim and Courtney 1998).

There are few studies using the above-cited knowledge elicitation techniques for the development of agent-based models (Lee 2004; Bossomaier et al. 2005; Dixon and Reynolds 2005; Bharwani 2006). It should be noted, though, that the paradigm of agent-based models is a world “under construction,” where the research techniques and methodologies are not, as yet, consolidated (Terna 1998). Despite that, within the research methods employed in the studies, there is a strong trend towards carrying out interviews with experts, chiefly for the modeling of cognitive agents.

Elicitation techniques are employed, specifically, for knowledge acquisition from experts. However, if owing to research limitations, or to the nature of the domain of the application, the knowledge required for the modeling is not amenable to being captured directly from human sources, these traditional techniques are not adequate for defining the agent's behavior. Most of the knowledge elicitation methods previously cited were proposed to deal with the acquisition of well-structured domain knowledge from single experts. However, as multiple experts may have different experiences and knowledge on the same application domain, it is necessary to elicit and integrate knowledge from multiple experts in building an effective knowledge base (Chu and Hwang 2008).

Given its complexity and comprehensiveness, the financial sector is a clear example of this situation. The knowledge about this system is not circumscribed by a group of experts, that is, the expertise required for its comprehension resides not only in different locations, rules, legislations, and actors, but is also dynamically created from interactions between these elements. In this case, elicitation techniques, viewed in isolation, are unable to offer an adequate definition of the agent's behavior in the model. It becomes necessary to develop new techniques for the definition of multi-agent systems and the capture of their peculiarities. The integration of these techniques with other techniques used in the context of qualitative research, such as content analysis, case studies, observation, etc., may open new perspectives in terms of achieving the best description of agent behavior.

The method introduced in this study integrates content analysis research and in-depth interviews for the modeling of agent behavior. The construction and implementation of this method, along with its application to the Brazilian financial system, are the main contributions of this research article to the literature on agent-based systems.

DESCRIPTION OF THE METHOD

The process of modeling agent-based models begins with the definition of the individual components for the understanding of the emergent behavior of the system. Considering that the conceptual model is employed for the specification of the agent-based model, the identification of actors and the characterization of their behaviors are admittedly the major challenges for a fair representation of the system under analysis. The term behavior in this article is related to attitudes, interests, intentions, perceptions, motivations, and other forms of action and reaction of the actors within situations of interest. This section describes the developed method to the behavior of agent' elicitations.

The method is a three-stage process (see Figure 3) that incorporates qualitative elicitation techniques. There are various approaches to qualitative research that may be of help to the researcher in collecting and analyzing data for the construction of a conceptual model. Denzin and Lincoln (2000) is a good reference, bringing together and describing diverse methods and techniques of qualitative research. Because the information concerning FSs is usually found in textual form and in the minds of experts from the area, it was decided to apply two qualitative data research techniques, namely, content analysis and in-depth interviews. The three stages of the process are to a great extent arbitrary and whether these are applied depends naturally on the specific domain. The method can be further described as follows.

FIGURE 3 Method of eliciting agent' behavior.

Stage 1: Domain Metamodel

The main objective of stage 1 is to define a domain metamodel of a specific problem domain. This stage comprises the use, construction, or development of the frames, rules, constraints, models, and theories applicable and useful for modeling a particular domain. “As a problem domain becomes better understood, formal theories or normative models can be constructed. In the absence of this formalization, problem-solving and understanding are more likely to depend on informal, intuitive, possibly unarticulated models” (Kim and Courtney 1998). The metamodel offers a (preferentiality) theoretical framework that can guide the subsequent stages.

Stage 2: Content Analysis

The main objective of the content analysis is the characterization of the behaviors and attitudes of the actors who influence and are influenced by the FS regulatory governance. Content analysis is “a research method that employs a set of procedures for making valid inferences from a given text” (Weber 1990). It is used to describe and construct the content of a whole class of documents and texts (Moraes 1999). The central ideas of content analysis are the classification of text extracts into categories in order to attain a better comprehension of their semantic contents, and to establish a model of the concepts called on by a person when he/she is facing a given situation. Words, sentences, or other registry units, which are classified into the same category, have similar or close meanings. According to Weber (1990), the decision regarding the choice and definition of the text unit to be classified is one of the major and most fundamental steps in content analysis and depends directly on the objectives of the analysis.

In the present study, two registry units are used: theme and actor. Theme is the meaning unit that is naturally released from an analyzed text, according to criteria relative to the research (Bardin 1977). The clippings from the text are obtained in accordance with the semantic level, each theme corresponding to one or more of the content segments deemed relevant. The actor is the other registry unit employed in the research. During execution of content analysis the actor will be linked to the themes, which are the contexts (scenarios) where their actions and behaviors occur.

Registry units are grouped and classified into categories. According to Bardin (1977), categories are classes that bring together a group of elements (registry units) under a generic title. They represent the result of an effort at data synthesis (Moraes 1999). The criterion used for the categorizations is semantic, that is, the grouping of registry units is carried out according to common themes, giving rise to thematic categories.

The categorization process may involve various cycles, that is, the data may be grouped into various categorization levels. According to Moraes (1999, p. 19), “the analysis of the material is processed in a cyclic and circular form, rather than in a sequential and linear form. The data do not speak for themselves. It is necessary to glean meaning from them.” It is important that the categories respect a set of criteria. At the end of each categorization process, categories should be (Moraes 1999): (i) valid: validity or pertinence refers to the analysis objectives, that is, all the categories created should be useful and significant in terms of the work proposed; (ii) exhaustive: no significant datum capable of classification should be left out; (iii) homogeneous: their organization should be based on a single classification principle or criterion (unique analysis dimension); (iv) mutually exclusive: each element should be classified into only one category, and; (v) objective: classification rules should be defined with sufficient clarity, so that they can be consistently applied across the analysis (classification is not affected by the subjectivity of the coder).

Stage 3: In-Depth Interviews

With the conclusion of the content analysis, the main elements and precepts required for the design of the conceptual model are mapped. Nevertheless, the information collected should be checked in order to revise any assumptions and fill any gaps that might exist. Refining of the information derived from the content analysis is achieved by means of in-depth interviews with specialists acting within the environment of the system under analysis. An in-depth interview is a technique for data collection, which is closer to conversation than to a formal structured interview (Marshal and Rossman 1989). It is a nonstructured direct interview, in which only one respondent is tested, by an interviewer, relative to a topic (Malhotra 1999). In the present case, the interviews will be planned in such a manner that the precepts and specifications of the governance model are assessed through formulated questions, taking into consideration the areas of competence and experience of the interviewees. Based on the in-depth interview technique, questions may be changed, omitted, or additional questions may be included, in order to obtain relevant information for the refinement of the model.

Vennix (1996) provides some useful hints for preparing and conducting interviews aimed at the construction of a conceptual model. Among the author's suggestions are: (i) elaboration of a list of topics serving as a guide; (ii) the formulation of questions about opinions; (iii) the use of questions of the type why; (iv) do not begin interviews with controversial issues; (v) do not express any value, either positive or negative, about what is declared by the interviewee; (vi) show a genuine interest towards the other person's ideas and opinions; (vii) provide adequate feedback or support to the respondent; (viii) use a recording device and have a transcript of the interview contents so as not to disturb the communication by going through lengthy notes.

METHOD IMPLEMENTATION

As mentioned in the previous section, the research is conducted by means of the analysis of qualitative data. For the purposes of this study, a datum is considered to be every and any type of text contributing to the mapping of the actor' behavior in the model. The data sources should be reliable, and they can be, for example, newspapers, magazines, report services, and information from experts acting in the environment of the system under analysis.

For the transcription of the content analysis, a database and forms were created in the MS ACCESS database manager. The MS ACCESS database was used to organize the data during the content analysis research, structuring and capturing the most important aspects of the qualitative information. It records the research results and the links between information grouped in different categorization cycles. At the conclusion of the research, this database stores the actors as well as the manifestation of their behaviors to be made available in particular contexts. Figure 4 represents how the data was structured using the developed database. On the left side is represented the basic entity-relationship (ER) model used to design the database and, on the right side, the schematic diagram shows how the data was organized after the conclusion of the content analysis research. It should be pointed out that no text processing technique, such as text-clustering algorithms (Aiello and Pegoretti 2006), was employed during the analysis of the data. The outcomes of the content research were expressed by means of production rules.

FIGURE 4 Representation of the research data structure.

In the case of the in-depth interviews, as previously commented, their purpose is to complement and refine the information examined during the content analysis research. The interview scripts are elaborated based on the thematic categories which are created during the content analysis research. Additionally, for the preparation and conducting of interviews, Vennix's (1996) suggestions were taken into consideration.

APPLICATION OF THE METHOD

This section describes the application of the method developed for the definition of agents in terms of their goals, intentions, and behaviors for the modeling of the Brazilian financial system. For the identification of the actors in the model and the definition of their behaviors, the authors of this article were granted access to the following data: (i) digital files of the news material on the financial system published by the mainstream print news media in Brazil from January 2000; (ii) minutes of monthly meetings of the Brazilian Central Bank's Monetary Policy Committee (COPOM)—the official body responsible for setting the monetary policy guidelines and defining the interest rate; and (iii) financial market specialists. The subsequent subsections will describe the application of the method to the BFS.

Domain Metamodel

In order to carry out the knowledge elicitation process in the BFS, Kooiman's (2003) governance framework was defined as our domain metamodel. This framework is one of the most complete in analyzing governance domains. Figure 5 presents the main elements of the framework. In Kooiman (2003), the governance emerges from the interactions among social, economical, and political entities (individuals, organizations, institutions). Interaction is the core concept in the framework. An interaction can be considered as a mutually influencing relation between two or more actors or entities, and consists of processes and structures (Kooiman 2003). The processes are the outcome of the capacity of actors or entities to act, while the structure of interactions is related to the contexts in which the interactions come about (material, social, and cultural contexts). At the intentional level of interactions the goals, interests, and purposes of actors are considered. The structural dimension of interactions consists of those circumstances that limit, broaden, and condition the interactions of the intentional level, such as institutions, patterns of communication, technological possibilities, and societal power distributions.

FIGURE 5 Framework to governance analysis (Kooiman 2003).

Kooiman (2003) discriminates the structural and intentional aspects of interactions and the relation between these two levels to analyze and interpret the socio-political processes. The author also distinguishes three forms of interactions: (i) interferences; (ii) interplays (reciprocal action and reaction); and (iii) interventions. Interferences are basically the least organized kind of interaction, e.g., they are very informal. Interplay can be considered as an interaction where there is no formal authority, domination, or subordination. The actors aim at reaching goals collectively and on a generally equal basis. Finally, interventions are the most formalized kind of interaction, characterized by a clear hierarchy. They aimed at directing exertion of formalized influence, such as the regulation behavior of central banks.

The intentional level of interactions is presented in three different elements: (i) images; (ii) instruments; and (iii) action. The public and private actors, in their interactions, make systematic use of these elements. They form images about what they are managing. During image formation, challenges will be defined and formulated, information will be gathered, opinions will be collected and selected. Image is also the point of departure for the selection of instruments and to take actions. Instruments are used by actors to move from one state of affairs to another. The use of a particular instrument or a set of them characterizes different forms of interactions. For interferences, information is the most characteristic instrument; for interplays, organization is important, while interventions require the application of rules. Finally, a certain amount of action is needed to bring these instruments to place. The actions are embedded in broader contexts. They occur at the intentional level of interactions, while simultaneously societal circumstances limit or enhance such actions.

It is also important to distinguish the three types of interactions (interferences, interplays, interventions) under particular contexts of governance or structures. These contexts are designated modes of governance and they represent a scheme for the structural level. There are three modes of governance: (i) self-governance; (ii) co-governance; and (iii) hierarchical governance. Self-governance refers to the capacity of social entities to govern autonomously. Self-governance is embedded in the sphere of the most spontaneous form of societal interactions, e.g., the interferences. Co-governance is associated with interacting parties that have something in common to pursue together. The actors communicate, collaborate, or co-operate without a central or dominating governing actor. Therefore, co-governance is directly related to interplay interactions. The hierarchical mode of societal governance displays types of governing interactions that have a top-down characteristic, e.g., those governing are in some way superimposed above those governed. For this reason, the hierarchical governance is associated with an intervention form of interaction.

Overall, the Kooiman's framework has acted as a template to support the development of the content analysis. The concepts and ideas presented in Figure 5 have guided the identification of how the real-world actors in the BFS interact, in which contexts they interact, and how they interact given different social, political, and economical environments.

Content Analysis

The definition of the content analysi' themes followed Kooiman's framework (Figure 5), which establishes that interactions among socio-political actors (at the intentional or action level) occur in particular contexts of governance (structural level). Elements involved in any interaction were considered as actors.

In this study, because of the extent of available data (especially the daily news regarding the BFS), a sampling criterion was established. Accordingly, the research is restricted to the analysis of the news released on the 5 working days of the week in which the Copom meets, on the day of the release of the minutes of the meeting, as well as on the following day, thus totaling seven dates for the analysis of the news and information concerning the BFS on a monthly basis. Since the year 2000, the ordinary Copom meetings have been held on a monthly basis, the meeting schedule for each year being announced by October of the preceding year. The minutes of the Copom meetings are released in the week subsequent to each meeting, within the regulatory period of 6 working days. Additionally, at the meetings of the National Monetary Board, for the definition of the inflation goals (monetary policy guideline) for the next periods, or owing to changes in the current goals, news analyses were carried out on the day of the meeting, as well as on the subsequent day. On a monthly basis, it was observed that the reports on the facts and events regarding interest rate policy (the specific subject analysis) are concentrated during these periods.

Due to a lack of time for investigating the myriad of newspapers, and other available sources of information, the analysis was performed using the reports regarding the BFS from one newspaper only. It should be pointed out that the purpose of this work is not to carry out an in-depth analysis of the agent' behavior, but rather to demonstrate the feasibility of the method. The chosen newspaper was the Folha de São Paulo (http://www.folha.com.br). This newspaper has the widest circulation among the mainstream news media, according to the Brazilian National Newspaper Association (http://www.anj.org.br). As the premise of its editorial guidelines, Folha de São Paulo pursues a critical, nonpartisan, and pluralist journalism. It favors diverse types of news, with an emphasis on that of a political and financial nature, which are fundamental for the construction of the model. In any case, the information from Folha de São Paulo, collected and analyzed in the periods chosen for the sample, are not conclusive. In-depth interviews with financial market experts add fresh information and new meanings to the data, for the specification and design of the conceptual model. With regard to the defined sample, 86 issues of Folha de São Paulo were analyzed (newspaper dates). This quantity represents the analysis of 730 news items and the identification of 1439 text segments (clippings) within the scope of the study.

Three categorization cycles were carried out for the content analysis research. The execution of the research procedure is represented in a simplified form in Figure 6. Difficulties were faced mainly during the classification of the rules for defining the agent's behaviors into independent categories, particularly during the second categorization process.

FIGURE 6 Content analysis research.

The first categorization cycle corresponds to the selection of 1439 text segments of news items from 86 issues of Folha de São Paulo. At this stage, it was also possible to classify 88 distinct denominations for actors acting in the BFS and expressing some type of significant behavior for the scope of the research, such as Central Bank, government, opposition to the government, analysts, economists, and unions.

The second categorization cycle involved the generation of production rules for the text segments selected from the first categorization cycle. The idea was to express the actor' behavior by means of rules of the type “IF-THEN.” The thematic and actor' categories were also grouped in order to reduce their quantity. At this stage, topics from 12 categories were identified from the prior classification, which were not used for the analysis (corresponding to 450 text segments). During the analysis, it was decided to restrict a little further the scope of the study in this first version of the model. Table 1 presents the nine categories resulting from the execution of the second categorization cycle conducted in the 23 categories of the first stage of the content analysis research. These categories group together the actor' behaviors under the form of production rules. The categorization process has also grouped the 88 actors from the preceding stage of the research into 10 distinct denominations. The actors are associated with the behavior rules in the categories.

TABLE 1 List of the Thematic Categories of the Second Categorization Cycle of the Content Analysis

Thematic category	# Text segments	Description
Aspects relative to the dynamics of the economy	273	Cause-and-effect relationships, commonly based on the agent' views acting in the FS; aspects relative to the behavior and communication among the agents; and the agent' perception of agents in relation to the economy, based on their expectations about certain future facts.
Aspects of Central Bank practices	231	Agent' perception in relation to the Central Bank's practices and behavior. In this category texts criticizing the Central Bank activities were also classified. Situations characterizing the Central Bank credibility, notwithstanding their association with the banks activities, were classified into a category of their own: “Situations impacting on Central Bank credibility.”
Aspects of the inflation target	24	Describes actor' behaviors and situations associated with definitions, adjustments, and failure to meet inflation target.
Aspects of monetary policy	55	In this category the clippings concerning reports on the actor' opinions and declarations related to the monetary policy are classified.
Effects of the decision on the Selic rate (Brazilian prime rate)	46	Impacts on the economy and the stances taken by actors in the FS as a result of the Copom decisions on the Selic rate.
Expectation in relation to decision on Selic rate	113	Actor' manifestations on the decision on interest rate by the Copom.
Pressures on the decision-making process of the monetary policy	62	Pressures exerted by actors in the FS in relation to decisions on interest rate.
Situations impacting on Central Bank credibility	102	Criticisms and praise in relation to the Central Bank practices, reinforcing or weakening its credibility before the other FS actors.
Noncategorized	83	Refers to text segments that were not classified into any category, although they are of interest because of the topic they deal with. The reason for the noncategorization is due to the difficulty/impossibility in creating production rules (IF-THEN) from the information contained in these texts.

The third stage of the categorization of the content analysis had as its central focus on the simplification and reduction of information. The items of information grouped in categories ‘non-categorized’ and ‘aspects relative to the dynamics of the economy’ were not considered at this stage of the research. The category ‘non-categorized’ was not considered, as per description shown in Table 1. In the case of the category ‘aspects relative to the dynamics of the economy’, the information within this grouping is insufficient to describe the cause-and-effect relationships between the macroeconomic variables and, consequently, the dynamics ruling the macroeconomic environment of the financial system. Since the study focus is not directly related to the characterization of the macroeconomic environment, but rather with the interactions and behaviors of the actors in this environment, it was decided not to analyze the data classified within this category. Also, production rules associated with some actors for whom the quantity of collected and analyzed information was insignificant were not considered, as, for example, with the rules mapped for actor ‘opposition to the government’, with only related 10 production rules.

In Table 2, the thematic categories resulting from the third categorization cycle are listed. These categories are related with the categories from the second categorization cycle. Column ‘# 3rd’ shows the quantity of production rules that are classified within each category from the third categorization. Table 3 describes the actors resulting from the execution of the third categorization cycle. Figure 7 illustrates the categorization process of the content analysis research, presenting an example of the construction of a rule that defines the behavior of one of the actors in the system.

FIGURE 7 Example of the categorization process of the content analyses research.

TABLE 2 List of Thematic Categories from the Third Categorization Cycle of the Content Analysis with the Categories from the 2nd Categorization Cycle

3rd categorization categories	# 3rd	2nd categorization categories
Stances in relation to Central Bank practices	380	Aspects of the Central Bank practices
		Effects of decision on Selic rate (Brazilian prime rate)
		Expectation in relation to the decision on Selic rate
Stances in relation to Central Bank credibility	102	Situations impacting Central Bank credibility
Stances in relation to inflation target	24	Aspects of the inflation target
Stances in relation to monetary policy	51	Aspects of the monetary policy
Pressures on monetary policy decision process	62	Pressures on monetary policy decision process
Total production rules		619 rules

TABLE 3 List of the Actor Categories from the Third Categorization Cycle of the Content Analysis

Category actors	# 3rd	Actors
Government	81	Government, working politicians aligned with the government, ministers, official bodies linked to the government, such as the National Treasury, etc., expressing opinions and issuing statements about the monetary policy and issues linked to the financial market.
Central Bank	35	Central Bank of Brazil
Financial market	58	Representatives of financial institutions or entities representing financial market interests
Analysts	228	Market analysts, economists, consultants, service providers, journalists, etc., having a connection with the market, divulge their opinions, and consequently, help build expectations in the ambit of the financial sector (journalists act as spokesmen on behalf of the market matters).
Industry and commerce	134	Unions, trade association representatives, industry, entrepreneurs/businessmen, etc.
Civil society	83	Newspaper editorial line columnists, politicians, sociologists, etc. not linked with the financial sector, expressing their opinions regarding the economy, financial sector, and related issues.
Total production rules		619 rules

Rule-based systems may sometimes grow very large, making rule consistency and contradiction quite problematic. According to Guillaume and Magdalena (2006), the model can be interpretable if it expresses the modeled system in an understandable way. However, this is a subjective property that relies on issues such as the model structure, the number of variables/rules, or the dimension and structure of the fuzzy sets. Some authors use the term transparency. It is a property that enables us to understand the influence of each parameter on the system's output. Transparency is related to the readability of the rules. In this work, the notion of transparency is based on the use of a moderate number of rules, variables, and fuzzy sets. Since the number of agents is low, a small number of rules helps the avoidance of inconsistencies. In addition, the BDI architecture helps to ensure that few fuzzy rules can guarantee the behavior of the agent based after it has committed to an intention. Also, agents in the real world can have contradicting behaviors. They depend on the granularity of their own model specification. Because of the BDI architecture, the intentions (plans) are chosen based on desires and beliefs. The intentions are specific to some conditions. Then, the fuzzy rules are triggered only after the desires and beliefs are tested. This condition helps to avoid contradiction.

Besides the use of the BDI framework, in order to avoid these two classic issues in rule-based systems, we decided to keep the number of rules and agents as small as possible. The following procedures were executed to reduce the model complexity: (i) the identification, selection, and classification of text segments into thematic and actor categories; (ii) grouping together of the categories, for data reduction and projection of text segments into production rules; and (iii) reduction in the number of categories by means of new groupings and through the exclusion of groups of rules that will not be used in the model design.

Although the quantity of thematic (5) and actor categories (6) are small in Tables 2 and 3, it can be seen that the number of production rules is still large (619 rules) to avoid some inconsistency and contradiction given the complexity of the problem domain. Instead of carrying out a new categorization process, we decided to proceed with in-depth interviews, benefiting from expert' knowledge to perform an additional (and proper) reduction in the number of agents and rules.

In-Depth Interviews

The purpose of the in-depth interviews was to verify and refine the information obtained through the content analysis research. The interviews were carried out over a period of 1 month, with six specialists of the financial system area as follows: (i) medium- and high-ranking managers and administrators of the Central Bank of Brazil (consultant, head of department, and heads of subsector areas linked to banking supervision and monetary policy of the financial system); and (ii) financial director of a Brazilian commercial bank attending quarterly meetings concerning monetary policy held by the Central Bank with the financial market.

The interview scripts were elaborated based on the thematic categories created during the content analysis research and on the precepts of our metamodel (see Figure 5). Additionally, Vennix's suggestions (1996) were taken into consideration for the preparation and conducting of the interviews. The topics defined in the interview scripts aimed at: (i) attesting to the production rules created through the content analysis research; (ii) detailing the dynamics and relationships among BFS actors mapped through the content analysis research; (iii) obtaining suggestions for the model specification; (iv) restricting the scope of the model through the identification of the most important behaviors of the actors within regulatory governance of the BFS.

Table 4 summarizes the topics that are part of the script and were addressed with the interviewees. These topics were differentiated according to the level at which they refer—structural or intentional (see Figure 5).

TABLE 4 Topics of the In-Depth Interview Scripts

Structural Level (macro)

Time allowed for the practices to yield effect in NFS

Governance measures

Inflation

Exchange

Interest rate

Economic policies

Intentional Level (micro)

Characterization of FS actors

Formal and informal interactions among FS actors

Standard behaviors (typical) and practices of FS sociopolitical actors

Regulatory body practices

Level of confidence between FS and the regulatory body (attitudes and expectation, reputation mechanisms in communications, etc.)

The interviewee' responses were recorded on the research instrument and transcribed into the tables by means of a text editor. During the interview analyses, the data were grouped together in order to facilitate the interpretation of the findings. Thus, interview responses are organized within a topic-based structure. Because of the informal nature of the in-depth interviews, not all questions covered by the proposed script of topics were addressed with the interviewees.

Generally, the interviews complemented the findings found with the content analysis research, giving important hints concerning a process of selecting the most important actors and defining the base interest rate as the most important regulatory instrument of a financial system. All rules defined in the content analysis not related to this regulatory instrument were neglected. The following important aspects resulted from the interviews are worth highlighting: (i) identification of the following agents: “monetary authority,” “financial market,” and “real sector economy,” as the main representatives of the actors within the BFS; (ii) the complex relationship of the currency exchange rate and the public debt with the interest rate, which reinforces the decision not to consider them within the scope of the first model version; and (iii) the relationship existing between the inflation target, inflation, and the other macro-economic indicators in the context of the monetary policy.

Based on the in-depth interviews, the rules were organized into categories according to the variables, agents, and behaviors expressed on them, as shown in Table 5. Each category associated with an agent includes several different rules. Table 5 gives a generic description of the rules associated with any agent and category. This final categorization facilitated the selection of the final set of production rules and the comprehension of the outcomes of the content analysis research. The six actors of the third categorization (Table 3) were also grouped into three actors for a more concise presentation of the rules in Table 5. It was decided that only rules relevant to the scope of interest in this study should be analyzed. The selection of the rules to model the behavior of the agents was made based on the following two criteria: (i) incorporating variables used to integrate the behavior of each agent to the environment of the financial market (data from environment used as an input to the agent-based model, data generated by the agent-based model used as an input to the environment); and (ii) variables that can generate governance measurements, e.g., variables used to measure the regulatory governance of the model. These two criteria revealed a lower set of production rules to model the fuzzy inference mechanisms of the agents: 280 rules. Then, from the 619 rules resulting from the content analysis research, 280 rules were selected. The selected rules were the ones considered relevant by the experts during the interviews. The discarded production rules were considered irrelevant, since the data and variables involved in them were considered by the experts very diverse and ambiguous. The BDI framework was also reliable to avoid inconsistencies and contradictions among the rules and agent' behavior.

TABLE 5 Rules Categories and Generic Descriptions

Agent: Financial Market (Financial Market+analysts)
Category: stances in relation to Bacen practices	# Rules
Perception of decisions related to Bacen's interest rates (effects and practices)	20
Beliefs on the Bacen stance in relation to its practices	16
Expectations regarding the Bacen practices in relation to interest rate decisions in the period (short term)	42
Expectations on the Bacen practices in relation to interest rate decisions in the period (medium term)	4
Expectations on the Bacen practices in relation to interest rate decisions in the period (long term)	2
Perceptions of the effects of the expectations on the expectations regarding the Bacen practices in relation to the interest rate.	1
Category: stances in relation to Bacen credibility
Perceptions of situations reinforcing Bacen's credibility	25
Perceptions of situations weakening Bacen's credibility	7
Category: stances in relation to the monetary policy
Expectations and desires in relation to the monetary policy	1
Category: pressures on the monetary policy decision
Perceptions of the effects of the pressures on the monetary policy decision process.	4
Pressures aiming to influence Bacen's practices	4
Agent: Economy Real Sector (industry and commerce+civil society)
Category: stances in relation to Bancen's practices
Perceptions of the decisions relative to Bacen's interest rates (effects and practices)	36
Beliefs on the Bacen's posture in relation to its practices	18
Expectations on Bacen's practices related to the decisions on interest rates in the period (short term)	5
Expectations on Bacen's practices related to the decisions on interest rates in the period (medium term)	2
Expectations on Bacen's practices related to the decisions on interest rates in the period (long term)	7
Category: stances in relation to Bacen's credibility
Perceptions of situations reinforcing Bacen's credibility	11
Perceptions of situations weakening Bacen's credibility	21
Category: stances in relation to the monetary policy
Negative evaluation of the monetary policy (criticisms)	2
Perceptions of the monetary policy (effects and practices)	1
Category: pressures the monetary policy decision process
Perceptions of the effects of the pressures on the monetary policy decision process.	3
Pressures aiming to influence Bacen's practices	10
Agent: Monetary Authority (govern+Central Bank)
Category: stances in relation to Bacen's practices
Perceptions of the decisions relative to Bacen's interest rates (effects and practices)	13
Beliefs on Bacen's posture in relation its practices	2
Expectations on Bacen's practices related to the decisions on interest rates in the period (short term)	4
Expectations on Bacen's practices related to the decisions on interest rates in the period (medium term)	1
Expectations on Bacen's practices related to the decisions on interest rates in the period (long term)	2
Category: stances in relation to the monetary policy
Negative evaluation of the monetary policy (criticisms)	1
Perceptions of the monetary policy (effects and practices)	2
Expectations and desires in relation to the monetary policy	2
Government statements about monetary policy	6
Category: pressures on the monetary policy decision process
Government statements aiming to influence the expectations in relation to the monetary policy	3
Pressures aimed at influencing the Bacen's practices	2

Then, these final sets of rules were analyzed for the extraction of the main cause-and-effect relationships describing the actor' behaviors and the subsequent definition of the fuzzy rules. Appendix A shows a synthesis of the fuzzy rules effectively used in modeling the agent behavior in the agent-based model. The fuzzy rules were derived from the 280 rules after carrying out the procedures equivalent to those employed in the content analysis research.

Model Validation

This subsection succinctly describes the procedure employed in demonstrating the adequacy of the agent' rules, that is, the validation of the behavior of each agent in the model. Accordingly, for the purposes of model validation, we have opted to analyze each agent separately and verify to what extent the results of the experiments correspond to real-world data. The experiments compare the decisions taken by real-world agents with computational ones, given the same set of input data. Real input data are used in the validation process for each new cycle for an independent outcome analysis.

Figure 8 exemplifies the validation of the decision on interest rates by the agent “monetary authority,” confronting the outcomes of the computational agent with the real values on interest rates defined by the Copom. For each cycle, real macroeconomics input data is used, validating the behavior of an agent. In this specific case, the data refers to: (i) expectations regarding inflation; (ii) variation in inflation in relation to the previous period; (iii) country-risk; (iv) situation regarding economic activity; and (v) interest rate decision in the previous period. The analysis period for validation is from August 2000 til September 2005, corresponding to 62 cycles. The definition of the period is due to the input data availability for the financial market inflation expectations.

FIGURE 8 Validation results for agent “monetary authority” on decisions on interest rate.

It is observed that the curve adherence is not constant. However, there are few periods in which the tendencies of the curves differ as, for instance, in the period 24 (July 2002). In this case, the decision on interest rates made by the agent “monetary authority” corresponds to an increase of 0.83 percentage points (p.p), while Copom's decision reduces the rate to 0.50 p.p. The differences between the agent's decision on the interest rate and Copom's effective decisions are smaller or equal to 0.25 p.p. in 56.5% of the 62 cycles studied. This percentage increases to 72.6% for differences smaller than or equal to 0.50 p.p. Generally, it is observed that the decision of the agent “central bank” follows the trends and variations of Copom's decisions, and even captures the most accentuated changes of these decisions.

CONCLUSIONS

The purpose of this article is to describe a method for the collection and structuring of qualitative data in order to develop agent-based models. The core objective of the method is to develop a conceptual model that captures the behaviors, intentions, attitudes, beliefs, and interactions of agents in the system under analysis. The method is based on the integration of content analysis research and in-depth interview techniques, and provides systematic data collection and interpretation. The complementarity of the two techniques makes it possible to interpret information and data produced in textual form as well as knowledge from experts.

Overall, the method has proven to be suitable for the formal representation of agents and their behavior within the BDI architecture by means of fuzzy logic. With the application of the method, it was possible to establish a set of relevant validated rules related to agent' behavior, which can be converted into fuzzy rules. These rules were encapsulated as intentions to execute the agent' actions (fuzzy-inference system module), after the beliefs and desires were processed by the agent (intention selection module), according to the agent' internal model in Figure 2. Hence, it is understood that the method can contribute to the development of further studies and methodologies for the modeling and representation of agent' behavior in social applied simulation models.

However, the whole process is far from simple. The content analysis research took longer to complete than originally planned. The number of rules generated by the content analysis was too large. It is possible to define a set of guidance principles towards the selection of “relevant” rules, restraining the number of them kept for further computer implementation. The main structural properties to be considered when analyzing a set of rules are completeness, consistency, and nonredundancy (Guillaume and Magdalena 2006). These questions are crisp properties, in the sense that a rule base is or is not complete, and two rules are or are not consistent or redundant. We can also quantify the level of satisfaction of those properties by defining levels of incompleteness, inconsistency, or redundancy (Casillas, Cordón, Herrera, and Magdalena 2003). In addition, one important question is the simplicity of the rules base, also called compactness, a property that is mostly qualitative. Simplicity refers to the number of rules in the rule base, and the number of variables that are considered in each rule. There are different methods for rule extraction which pay attention to some of these properties or qualities (Setnes, Babuska, and Verbruggen 1998). In any case, what is important to consider is what kind of property from a partial rule base will influence or not the integrated rule base, in order to define the key properties of the rule extraction process. As a result, the most important question to be considered during rule generation is the compactness of the rules, while consistency and nonredundancy should be managed during integration and conflict reduction.

The experience gained from this analysis leads to the conclusion that the method can be further enhanced. Several steps could be carried out in a different way. The selection of texts could have been optimized, based on a preanalysis of the “news” items, identifying those of greater relevance and discarding those of little relevance. Domain experts could have been used in this task with great efficacy. Also, the relevance of actors could be used in order to reduce the number of texts selected, eliminating those that are related to actors of little significance in terms of the objectives of the analysis. These observations allow us to define the problem formulation, in terms of the definition of objectives and scope of analysis, as the critical success factor for the application of the developed method. It seems that the method can work very well with ill-structured domains given that the system analysts have set well-defined objectives. All rules not directly involved with these objectives can be considered as serious candidates to be eliminated from the analysis. Another aspect that could be rationalized is the ordering of the application of the content analysis and the in-depth interviews. We believe that altering this order, firstly carrying out interviews, might speed up the content analysis process, offering aspects and elements that would guide us in the selection of rules and in the structure of the content analysi' themes and actors.

Although the developed method accomplished its objective of establishing a verified conceptual model, the authors realize that the method is still a first attempt in the right direction. The generality of the method is an issue that requires a deeper analysis. The application field of agent-based modeling is vast and increasing as its effectiveness is proven. As a consequence, it is possible to argue whether a single structuring method is able to cover the whole range of applications and situations. The answer is obviously no. There are intrinsic factors to each area of application that make it impossible to define a single method. Nevertheless, the developed method presents generic ideas that can be used as guidance to any such process such as the integration of different techniques into one integrated, formal approach. Multiple techniques in an appropriate sequence is an effective way of building upon the information elicited from each stage. It also facilitates the collection of different types of data and knowledge allowing a measure of triangulation, which can be used to confirm the validity and consistency of acquired behaviors. This issue is currently being addressed by the authors and it will constitute subject of future research. Moreover, the research proceeds for implementing the conceptual model using JAVA and JASON framework (Bordini and Hübner 2005). JASON is an interpreter for an extended version of AGENTSPEAK, a BDI agent-oriented logic programming language. The agent-based model is integrated to an econometric model, which is responsible in characterizing the macroeconomic environment. The dynamic behavior of the model is obtained through simulation experiments to characterize the regulatory governance in different monetary policy situations. It will generate measures of the financial sector credibility, pressures on the monetary authority agent, interest rate expectations, interest rate decisions, among others.

Acknowledgments

The authors would like to thank Central Bank of Brazil and CNPq, for the financial support to the research. Also, the authors would like to express their sincere thanks to the anonymous referee for their valuable suggestions.

APPENDIX A: EXAMPLE OF FUZZY RULES FOR THE MODELING OF AGENT BEHAVIOR

Table A.1 displays the fuzzy rules derived from the 280 production rules that were selected from the content analysis research. The fuzzy rules are organized according to the behaviors of the agents: monetary authority, financial market, and real sector economy, which represent the grouping of six actors (government, central bank, financial market, analysts, industry and commerce, civil society). Equally, the rules are described into a suitable format by means of fuzzy variables and standardized linguistic terms.

TABLE A.1 Example of the Fuzzy Rules Describing the Actor's Behavior in the Model

Pressure of the agent financial market on the agent monetary authority

IF (interest rate quite high) AND (inflation not quite high or not high) THEN (high pressure for interest rate reduction)

Pressure of the agent real sector economy on the agent monetary authority

IF (interest rate quite high) THEN (high pressure for interest rate reduction)

IF (economic activity quite low) THEN (high pressure for interest rate reduction)

IF (more than 2 periods of low economic activity) THEN (high pressure for interest rate reduction)

IF (average interest rate) AND (quite low inflation) THEN (high pressure for interest rate reduction)

Expectations of the agent financial market in relation to decisions on interest rates

IF (inflation increases very much) THEN (expectation: aggressive increase interest rate)

IF (inflation decreases very little) THEN (expectation: reasonable increase interest rate OR expectation maintenance of interest rate)

IF (inflation decreases) THEN (expectation: reduction interest rate)

IF (interest rate quite high) AND (inflation decreases) AND (low economic activity) THEN (expectation: aggressive reduction interest rate)

IF (interest rate quite high) AND (inflation decreases reasonable) AND (average economic activity) THEN (expectation: reasonable reduction interest rate)

IF (interest rate quite high) AND (reasonable inflation) AND (economic activity decreases) AND (country-risk high) THEN (expectation: increase high interest rate)

IF (country-risk high) AND (conditions for reduction interest rate) THEN (expectation: small interest rate reduction)

IF (economic activity decreases) THEN (expectation: small reduction interest rate)

IF (reasonable inflation) AND (low economic activity) THEN (expectation: high OR reasonable reduction interest rate)

IF (prior interest rate decision aggressive reduction) THEN (expectation: decision non-aggressive interest rate reduction)

IF (high interest rate) AND (inflation decreases) AND (prior decision reduces interest rate) THEN (expectation: small interest rate reduction)

IF (interest rate close to equilibrium interest rate) THEN (expectation: decision small interest rate reduction)

IF (statement of agent monetary authority optimistic) AND (monetary authority credibility high) THEN (expectation: decision interest rate improves)

Statements of agent monetary authority to influence expectations on interest rate decisions

IF (inflation quite high OR high) AND (pressure reduction interest rates) THEN (statement difficulties)

Decision on interest rate by the agent monetary authority

IF (high inflation) THEN (decision increase interest rate)

IF (inflation decreases little) THEN (decision small interest rate increase)

IF (high inflation) THEN (decision aggressive interest rate increase)

IF (monetary authority concerned growth economic activity) THEN (reduction interest rate higher than expectation)

IF (country-risk high) THEN (decision maintain interest rate)

IF (prior decision interest rate large reduction OR aggressive) THEN (decision interest rate reduction lower than prior period)

Perception of agent financial market in relation to the credibility of agent monetary authority

IF (inflation decreases) THEN (monetary authority credibility increases)

IF (inflation decreases) AND (decision small interest rate increase) THEN (monetary authority credibility increases)

IF (reasonable inflation) AND (maintain interest rate) THEN (monetary authority credibility is maintained)

IF (inflation does not decrease) AND (decision interest rate reduction) THEN (monetary authority credibility decreases)

IF (decision non-aggressive reduction) AND (expectation aggressive interest rate reduction) THEN (monetary authority credibility increases)

IF (high interest rate) AND (decision interest rate unchanged OR small increase) THEN (monetary authority credibility is maintained)

IF (high interest rate) AND (decision small interest rate reduction) AND (inflation decreases) THEN (monetary authority credibility increases)

IF (decision interest rate not surprising) THEN (monetary authority credibility increases)

IF (decision aggressive interest rate reduction) AND (expectation reasonable interest rate reduction) THEN (interest rate decision surprising)

IF (decision aggressive interest rate reduction) AND (expectation non-aggressive reduction) AND (economic activity quite low) THEN (decision interest rate not surprising)

IF (interest rate quite high) AND (decision maintain interest rate) THEN (interest rate decision not surprising)

IF (decision reasonable interest rate reduction) AND (interest rate quite high) AND (inflation decreases) THEN (interest rate decision not surprising)