Agent-based Models of the Economy

From Theories to Applications

R. Boero, M. Morini, M. Sonnessa, P. Terna

• Editore: Palgrave Macmillan
• Anno: 2015
• ISBN:
  9781137339812
• Argomento: Economics
• Categoria: Econometrics

Agent-based models are tools that provide researchers in economic fields with unprecedented analytical capabilities. This book describes the power of agent-based models along their methodology, and it provides several examples of applications spanning from public policy evaluation to financial markets.

Part I

Concepts and Methods

1

A Few Good Reasons to Favor Agent-based Modeling in Economic Analyses

Riccardo Boero

An Agent-based Model (ABM) is a formal tool for scientific inquiry. When dealing with an unknown scientific method, several questions are worth asking: What is it? How to use it? And most important, is it useful?

Here, we address the last question, discussing how ABMs can improve economic analyses by solving some of the most important limitations of traditional modeling tools used in the discipline.

1.1 Integration of theory and data

Among the many complaints about and critiques of the common practice of research in economics, there is one about the separation of the approaches to it in theoretical and empirical works. Although in economics, even in recent times, there are many examples of fruitful integration of empirical data and theory development, it is also common to find evidence of a over-strict separation between theory and data. From one perspective, a over-strict separation leads to theoretical developments that rely more on theoretical consistency than on empirical salience. From the opposite perspective, evidence from empirical studies is often not studied in depth because it can not be harmonized with pre-existing theories.

Among the many causes of this phenomenon is the lack of flexibility in the tools traditionally used in economics. The main theoretical contributions take the form of equations when formalized, but, because of the complexity of the topics in question and the empirical data, such formalism is often inadequate.

To a certain degree, common formalisms in economics do not always support the linkage with empirical data, and in empirical analysis relying on equation-based models, like most econometric models, there is difficulty in identifying clear causal relationships. Furthermore, econometric validations of equation-based models present researchers with many other challenges beyond causality, such as those posed by a regression of several simultaneous equations.

The most common example of the short-circuit between theory and data in economics is the overwhelming evidence collected over centuries about the unrealistic nature of maximizing agents. Equation-based models are largely incapable of capturing the algorithmic nature of behavioral data, and economists thus have to rely on the “as-if” principle even in most recent theories. Empirical economists, behavioral ones in particular, continuously observe data that is inconsistent with theoretical models of behavior. However, the few economic theories of realistic economic behavior that are actually based on empirical evidence are not widely accepted today, and it is even harder to say that those have broken into the disciplinary mainstream. In other words, even if there is agreement on the unrealistic nature of some theoretical components, there is no substitute for them, or at least no a substitute that can be used within the traditional approach.

The technical limitations of traditional tools in economics oblige researchers who adopt them to give up a degree of realism, and to accept compromises whose consequences can not be evaluated, neither ex-ante nor ex-post. Standard economic tools thus impose a well-known trade-off between tractability and realism (Lawson, 2003). From a general perspective, ABMs avoid this trade-off by allowing a large degree of integration between theoretical and empirical knowledge, as shown in the discussion throughout this chapter.

ABMs can in fact be used for conducting abstract theoretical investigations (see, for instance, the thought experiments discussed in Bedau (1998; 1999)) that do not require empirical grounding, as well as highly specific applied analyses that use large amounts of data along a continuum of options (Boero and Squazzoni, 2005). Furthermore, with ABMs it is possible to use empirical knowledge in theoretical analysis, allowing researchers both to leave intact the loop of scientific discovery that goes back and forth between theory and data, and to properly address the different kinds of applied research questions posed by the contemporary world.

1.2 Causality and uncertainty

Since the beginning of the development of a methodology for ABMs there have been many contributions stressing the “emergentist” and “generativist” nature of this modeling tool (e.g. Epstein, 2006). The reason for this argument is that the flexibility in modeling provided by ABMs allows replication of the phenomenon of interest with a higher degree or realism than in other traditional models. In other words, with ABMs researchers are free to choose the degree of realism of their models, bounded by the availability of knowledge and data on the phenomenon of interest but not by the formalism they adopt, as is the case with equation-based models.

This feature of ABMs has a great impact on causality. In fact, the ability to realistically model the mechanism generating a phenomenon generates the ability to investigate which mechanisms are responsible for a phenomenon, their complex interplay and the conditions under which one mechanism, or cause, prevails over others. When testing equation-based models over empirical data, on the contrary, it is often difficult to discern causality from spurious correlation and to investigate endogeneity.

The possibility of modeling causal mechanisms in ABMs in conjunction with the adoption of opportune validation techniques (in particular structural validation, Troitzsch, 2004) and specific analytic procedures on model outputs (ABMs in fact generate complex outputs that could appear “opaque” – Di Paolo et al., 2000 – but that can be fully comprehended by means of sensitivity analysis) allows researchers to fully grasp causality in social and economic phenomena. This powerful feature of the agent-based approach has a positive impact not only on the explanation of phenomena and on theory development, but also on applied and predictive analyses.

In fact, the robust identification of causal mechanisms provides researchers and decision-makers with the possibility of fully developing effective policy measures. For instance, when a public policy is designed and valuated ex-ante with an ABM, it is possible not only to evaluate its systemic performance and to investigate the causes for the latter, but also to identify policy weak spots. By means of causal analysis it is possible, in other words, to fully support decision-makers by suggesting where to intervene with modifications.

This feature, furthermore, also allows researchers and decision-makers to have a higher degree of confidence in model results, since understanding the dynamics resulting from a proposed intervention is crucial, especially when the intervention has not been under experiment before.

A very similar argument can be presented on uncertainty. The capability to model causal mechanisms provided by ABMs extends the capability to fully investigate uncertainty propagation in these models, that is to say, to trace the sources of the uncertainty of systemic outcomes.

Returning to the example of the evaluation of a public policy, this feature of ABMs can be used to allow researchers and policy-makers to separate the uncertainty due to the incompleteness of the model (epistemic uncertainty) from the intrinsic uncertainty of human processes and decisions (aleatory uncertainty). A good understanding of uncertainty is thus crucial to a fruitful and productive interaction between economists and decision-makers (Brown et al., 2005).

1.3 Heterogeneity and interaction

Differences in economic actors codetermine most economic phenomena. However, most economic models rely on the concept of representative agents, that is to say economic actors with characteristics equal to the population average.

In equation-based models, heterogeneity is bound to be very low. For instance it is usual to separate the behavior and endowment of firms from those of consumers, but heterogeneity is not considered within each category of economic actors.

ABMs do not impose a priori any constraint on the degree of similarity between economic actors. From the perspective of an extreme case, economic agents in ABMs can be modeled one by one, with completely differen...

Indice dei contenuti

• Cover
• Title
• Part I Concepts and Methods
• Part II  Applications
• Index