Paulo André Lima de Castro

Complex network study of Brazilian soccer players

Although being a very popular sport in many countries, soccer has not received much attention from the scientific community. In this paper, we study soccer from a complex network point of view. First, we consider a bipartite network with two kinds of vertices or nodes: the soccer players and the clubs. Real data were gathered from the 32 editions of the Brazilian soccer championship, in a total of 13 411 soccer players and 127 clubs. We find a lot of interesting and perhaps unsuspected results. The probability that a Brazilian soccer player has worked at N clubs or played M games shows an exponential decay while the probability that he has scored G goals is power law. Now, if two soccer players who have worked at the same club at the same time are connected by an edge, then a new type of network arises (composed exclusively by soccer player nodes). Our analysis shows that for this network the degree distribution decays exponentially. We determine the exact values of the clustering coefficient, the assortativity coefficient and the average shortest path length and compare them with those of the Erdös-Rényi and configuration model. The time evolution of these quantities are calculated and the corresponding results discussed.

Nonlinear Barabási–Albert network

In recent years there has been considerable interest in the structure and dynamics of complex networks. One of the most studied networks is the linear Barabasi–Albert model. Here we investigate the nonlinear Barabasi–Albert growing network. In this model, a new node connects to a vertex of degree k with a probability proportional to kα (α real). Each vertex adds m new edges to the network. We derive an analytic expression for the degree distribution P(k) which is valid for all values of m and α⩽1. In the limit α→−∞ the network is homogeneous. If α>1 there is a gel phase with m super-connected nodes. It is proposed a formula for the clustering coefficient which is in good agreement with numerical simulations. The assortativity coefficient r is determined and it is shown that the nonlinear Barabasi–Albert network is assortative (disassortative) if α 1) and no assortative only when α=1. In the limit α→−∞ the assortativity coefficient can be exactly calculated. We find r=713 when m=2. Finally, the minimum average shortest path length lmin is numerically evaluated. Increasing the network size, lmin diverges for α⩽1 and it is equal to 1 when α>1.

Optimization and self-organized criticality in a magnetic system

We propose a kind of Bak–Sneppen dynamics as a general optimization technique to treat magnetic systems. The resulting dynamics shows self-organized criticality with power-law scaling of the spatial and temporal correlations. An alternative method of the extremal optimization (EO) is also analyzed here. We provided a numerical confirmation that, for any possible value of its free parameter τ, the EO dynamics exhibits a non-critical behavior with an infinite spatial range and exponential decay of the avalanches. Using the chiral clock model as our test system, we compare the efficiency of the two dynamics with regard to their abilities to find the systems ground state.

Towards automated trading based on fundamentalist and technical data

Autonomous trading is often seen as artificial intelligence applied to finance by AI researchers, but it may also be a way to motivate the development of autonomous agents, just like robot soccer competitions are used to motivate the research in mobile robots. In fact, some initiatives could be observed in recent years, for instance [1] and [2]. In this paper, we present a multiagent system composed by several autonomous analysts that use fundamentalist information in their reasoning process. These fundamentalist information are composed by company profit, dividends, data related to the company economic sector among others. This kind of information is rarely used on autonomous trading, because most of the agents deal only with technical information, which is composed by price and volume time series. Furthermore, we do not find a open source stock market simulator with support to fundamentalist trader agents. We then created a significantly extended version of the open source financial market simulation tool, called AgEx. This designed version provides also fundamentalist information about the traders assets. As well as, makes more efficient the exchange of messages within AgEx. This efficiency allows traders that may submit orders in very short intervals of just some seconds or even some fraction of second, to use AgEx as a test platform. Using this new version of AgEx, we implemented and tested the multiagent system based on fundamentalist agents, that we call FAS. The achieved results are presented and analyzed.

AgEx: A Financial Market Simulation Tool for Software Agents

Many researchers in the software agent field use the financial domain as a test bed to develop adaptation, cooperation and learning skills of software agents. However, there are no open source financial market simulation tools available, that are able to provide a suitable environment for agents with real information about assets and order execution service. In order to address such demand, this paper proposes an open source financial market simulation tool, called AgEx. This tool allows traders launched from distinct computers to act in the same market. The communication among agents is performed through FIPA ACL and uses a market ontology created specifically to be used for trader agents. We implemented several traders using AgEx and performed many simulations using data from real markets. The achieved results allowed to test and assess comparatively trader’s performance against each other in terms of risk and return. We verified that the effort to implement and test trader agents was significantly diminished by the use of AgEx. Furthermore, such results indicated new directions in trader strategy design.

Automated asset management based on partially cooperative agents for a world of risks

Despite the fact any investor prefers lower risk and higher return, investors may have different preferences about what would be an acceptable risk or a minimal return. For instance, some investors prefer to have a lower bound risk rather than gaining a higher return. In portfolio theory, it is commonly assumed the existence of one risk free asset that offers a positive return. This theoretical risk free asset combined with a risky portfolio creates a new portfolio that presents a linear relation between risk and return as the risk free asset weight (wf) changes. Hence, any level of risk or of return is easy to achieve separately, just by changing wf. However, in a world without any risk free assets, the combination between assets creates nonlinear portfolios. Achieving a specific level of risk or return is not a trivial task. In this paper, we assume a risky world rather than the existence of a risk free asset, in order to model an automated asset management system. Furthermore, some automated asset managers give very different results when evolving in different contexts: hence, a very profitable manager can have very bad results in other market situations. This paper presents a multiagent architecture, aiming to tackle these problems. The architecture, named COAST (COmpetitive Agent SocieTy), is based on competitive agents that act autonomously on behalf of an investor in financial asset management. It allows the simultaneous and competitive use of several asset analysis techniques currently applied in the finance field. Some dedicated agents, called advisors, apply a particular technique to a single asset. The results provided by these advisors are then submitted to and analyzed by a special agent called coach, who evaluates its advisors’ performance and defines an expectation about the future price of one specific asset. Within COAST, several coaches negotiate to define the best money allocation among different assets, by using a negotiation protocol defined in this paper. We also propose an investor description model that is able to represent different investors’ preferences with defined acceptable limits of risk and/or return. The COAST architecture was designed to operate adequately with any possible investor’s preference. It was implemented using a financial market simulator called AgEx and tested using real data from the Nasdaq stock exchange. The test results show that the architecture performed well when compared to an adjusted market index.

SELF-ORGANIZED CRITICALITY, OPTIMIZATION AND BIODIVERSITY

By driving to extinction species that are less or poorly adapted, the Darwinian evolutionary theory is intrinsically an optimization theory. We investigate two optimization algorithms with such evolutionary characteristics: the Bak–Sneppen and the Extremal Optimization. By comparing their mean fitness in the steady state regime, we conclude that the Bak–Sneppen dynamics is more efficient than the Extremal Optimization if the parameter τ is in the interval [0, 0.86]. The determination of the spatial correlation and the probability distribution of the avalanches show that the Extremal Optimization dynamics does not lead the system into a critical self-organized state. Through a discrete form of the Bak–Sneppen model we argue that biodiversity is an essential prerequisite to preserve the self-organized criticality.We propose a kind of Bak–Sneppen dynamics as a general optimization technique to treat 9 magnetic systems. The resulting dynamics shows self-organized criticality with power-law scaling of the spatial and temporal correlations. An alternative method of the extremal optimization (EO) 11 is also analyzed here. We provided a numerical con3rmation that, for any possible value of its free parameter , the EO dynamics exhibits a non-critical behavior with an in3nite spatial range 13 and exponential decay of the avalanches. Using the chiral clock model as our test system, we compare the e5ciency of the two dynamics with regard to their abilities to 3nd the system’s 15 ground state. c

Expected utility or prospect theory: Which better fits agent-based modeling of markets?

Abstract Agent-based simulations may be a way to model human society behavior in decisions under risk. However, it is well known in economics that Expected Utility Theory (EUT) is flawed as a descriptive model. In fact, there are some models based on prospect theory (PT), that try to provide a better description. If people behave according to PT in finance environments, it is arguable that PT based agents may be a better choice for such environments. We investigate this idea in a specific risky environment, a financial market. We propose an architecture for PT-based agents. Due to some limitations of the original PT, we use an extension of PT called Smooth Prospect Theory (SPT). We simulate artificial markets with PT and traditional (TRA) agents using historical data of many different assets over a period of 20 years. The results showed that SPT-based agents provided behavior that is closer to real market data than TRA agents, and that the improvement when using SPT rather than TRA agents is statistically significant. It supports the idea that PT based agents may be a better pick to model the behaviour of agents in risky environments.

Towards Modeling Securities Markets as a Society of Heterogeneous Trading Agents

In recent article, Farmer and Foley [1] claimed that the agent-based modeling may be a better way to help guide financial policies than traditional mathematical models. The authors argue that such models can accurately predict short periods ahead as long as the scenario remains almost the same, but fail in times of high volatility. Another real world problem that is rarely addressed in agent-based modeling is the fact that humans do not make decisions under risk strictly based on expected utility. This context inspired the goal of this work: modeling trading agents to populate an artificial market and use it to predict market price evolution in high and low volatility periods. We developed a set of simple trading agents and executed a set of simulated experiments to evaluate their performance. The simulated experiments showed that the artificial market prediction performance is better for low volatility periods than for higher volatility periods. Furthermore, this observation suggests that in high volatility period trading agent strategies are influenced by some other factor that is not present or is smaller in other period. These facts lead us to believe that in high volatility period human agents can be influenced by psychological biases. We also propose in this paper one simple trading agent model that includes prospect theory concepts in his decision making process. We intend to use such model in future work.

Towards autonomous investment analysts — Helping people to make good investment decisions

Since early days of computer science, researchers ask themselves where is the line that separates tasks machine can do from those only human beings can really accomplish. Several tasks were pointed as impossible to machines and later conquered by new advances in Artificial Intelligence. Nowadays, it seems we are not far from the day when driving cars will be included among the tasks machines can do in an efficient way. Certainly, even more complex activities will be dominated by machines in the future. In this paper, we argue that investment analysis, the process of assessment and selection of investments in terms of risk and return, should and can be among the tasks performed efficiently by machines in the (maybe not so far) future. Investment decisions have to be faced not only by financial professionals but by all people. Naturally, these professionals have more complex and often decisions to make, but everybody needs to invest to warrant good standard of living in the old age. In fact, there is significant research effort to create algorithms and/or quantitative methods to analyze investments. We present a brief review of them. Through this review, we may realize that there are many interconnected challenges in the quest for autonomous investment analysis. In this paper, we propose an adaptive multiagent architecture that deals with these three dimensions of complexity (nature of assets, multiple analysis algorithms per asset and horizon of investment) and keeps an explicit model of investors preferences. This architecture breaks down the complexity faced by AIA in problems that can be addressed by a group of agents that work together to provide intelligent and customized investment advices for individuals. We believe that such architecture may contribute to development of AIA that deals with the complexity of the problem in a tractable way. Furthermore, this architecture allows the incorporation of known algorithms and techniques that may help to solve part of the issue.