EPJ manuscript No.(will be inserted by the editor)

A complex systems approach to constructing better models formanaging financial markets and the economy

J. Doyne Farmer1, Mauro Gallegati2, Cars Hommes3, Alan Kirman4, Paul Ormerod5, Silvano Cincotti6, AnxoSanchez7, and Dirk Helbing8

1 Santa Fe Institute, 1399 Hyde Park Road, Santa Fe, NM 875012 DiSES, Universit Politecnica delle Marche, Ancona, Italy3 CeNDEF, University of Amsterdam, Netherlands4 GREQAM, Marseille, France5 Volterra Partners, London and University of Durham, UK6 DIBE-CINEF, University of Genoa7 GISC, Universidad Carlos III de Madrid, Spain8 ETH, Zurich

March 14, 2012

Abstract. We outline a vision for an ambitious program to understand the economy and financial marketsas a complex evolving system of coupled networks of interacting agents. This is a completely different visionfrom that currently used in most economic models. This view implies new challenges and opportunitiesfor policy and managing economic crises. The dynamics of such models inherently involve sudden andsometimes dramatic changes of state. Further, the tools and approaches we use emphasize the analysis ofcrises rather than of calm periods. In this they respond directly to the calls of Governors Bernanke andTrichet for new approaches to macroeconomic modelling.

PACS. PACS-key discribing text of that key – PACS-key discribing text of that key

Contents

1 Visionary Approach . . . . . . . . . . . . . . . . . . 12 State of the Art . . . . . . . . . . . . . . . . . . . . . 83 Innovative approach . . . . . . . . . . . . . . . . . . 114 Expected Paradigm Shifts . . . . . . . . . . . . . . . 155 Expected Impact . . . . . . . . . . . . . . . . . . . . 16

1 Visionary Approach

In November 2010, European Central Bank (ECB) thenGovernor Jean-Claude Trichet opened the ECBs flagshipannual Central Banking Conference with a challenge tothe scientific community to develop radically new ap-proaches to understanding the economy:

When the crisis came, the serious limitations of ex-isting economic and financial models immediately be-came apparent. Macro models failed to predict the cri-sis and seemed incapable of explaining what was hap-pening to the economy in a convincing manner. As apolicy-maker during the crisis, I found the availablemodels of limited help. In fact, I would go further: inthe face of the crisis, we felt abandoned by conventionaltools.

. . . we need to develop complementary tools to im-prove the robustness of our overall framework. In thiscontext, I would very much welcome inspiration fromother disciplines: physics, engineering, psychology, bi-ology. Bringing experts from these fields together witheconomists and central bankers is potentially very cre-ative and valuable. Scientists have developed sophisti-cated tools for analysing complex dynamic systems in arigorous way. These models have proved helpful in un-derstanding many important but complex phenomena:epidemics, weather patterns, crowd psychology, mag-netic fields.

An important component of the FuturICT proposalwill be to deliver the complementary tools that GovernorTrichet is calling for. The FuturICT project will exploitthe rapidly expanding capacity of ICT to develop a sys-tem to continuously monitor and evaluate the social andeconomic states of European countries and their variouscomponents, and, what is of particular importance for thisdiscussion, the real economy, the governmental sector, thebanking and finance sector, by developing and managinga massive repository of high-quality data in all economicand financial areas; This will provide a platform for thedevelopment and application of data mining, process min-

2 Farmer et al.: Complex systems for better management of the economy

ing, computational and artificial intelligence and everyother computer and complex science technique coupledwith economic theory and econometric methods that canbe devoted to identifying the emergence of social and eco-nomic risks, instabilities and crises; FuturICT will providethe framework and infrastructure to perform what-if anal-ysis, scenario evaluations and computational experimentsto inform decision makers and help develop innovative pol-icy, market and regulation designs. This will involve push-ing the envelope of what is technologically possible, usingnew information and computational capabilities to greatlyexpand the state of the art in the collection and modellingof data, at a hitherto unprecedented scale. Perhaps as im-portant we shall investigate the impact that new infor-mation technologies, coupled with a natural tendency ofmarket participants to “herd”, has had on the evolutionof financial market prices. As Governor Bernanke said:

“The brief market plunge was just an example of howcomplex and chaotic, in a formal sense, these systemshave become ... What happened in the stock market isjust a little example of how things can cascade, or howtechnology can interact with market panic” (Interviewwith Ben Bernanke, the IHT May 17th 2010).

The only way to make a major advance in economicmodelling is to explore entirely new approaches ratherthan make incremental modifications to existing mod-els. The economic models developed will be tightly in-tegrated with the broader sociological modelling effort ofFuturICT. Such an approach will take us away from thecomfortable belief that the macroeconomy is well under-stood and that only occasional shocks perturb it. To quoteLord Turner, head of the U.K. financial services authority:

“But there is also a strong belief, which I share, thatbad or rather over-simplistic and overconfident eco-nomics helped create the crisis. There was a dominantconventional wisdom that markets were always ratio-nal and self-equilibrating, that market completion by it-self could ensure economic efficiency and stability, andthat financial innovation and increased trading activitywere therefore axiomatically beneficial”.

The purpose of our project is precisely to take up LordTurner’s challenge and to build models and perform anal-ysis of the economy as a complex system prone to suddenand major changes of endogenous origin. Our analysis willalso show that the blind faith that has been put on thecapacity of markets to self-organise in a stable way is mis-placed. Better approaches, such as those we suggest, canhelp understand intrinsic fragilities of the economic sys-tem and help to provide appropriate policy measures tomitigate them.

1.1 Goals

We intend to use new tools in information and commu-nication technologies to implement an integrated complexsystems approach for understanding financial markets andthe economy. This includes the collection of new data, newmethods of empirical analysis, and the development of newmathematical and computational tools. This effort willbe guided by emerging new conceptual paradigms suchas network theory, market ecology, behavioural economicsand agent-based modelling, and empirically grounded bylaboratory experiments and micro data. By combiningthese ideas into a practical simulation and forecasting toolour goal is to build the foundations of a new paradigmwithin economics.

1.1.1 Modelling the economy as a complex evolving network

The view of the economy as a complex system is at leastas old as Adam Smith, who (though not in these words)described the economy and the social welfare that it cre-ates, as an emergent process based on the self-organizedbehaviour of independently acting, self-motivated indi-viduals. Specific examples of complex systems in socialand economic contexts include traffic flows, large supplychains, financial markets, group dynamics and crowd be-haviour. Examples of pioneering papers which view socio-economic issues as complex systems are in [5,10].

Complex systems, which predominate in modern socio-economic systems, raise a daunting set of problems requir-ing new and innovative approaches. These systems aremade up of a large number of interacting individual el-ements, such as people, companies, countries, cars. Theychallenge conventional thinking. For example, they are dy-namic rather than static. They are probabilistic and notdeterministic. They appear to very difficult to predict andcontrol, and are permeated by non-linear or network in-teractions amongst the component agents. The individualelements of a system are influenced directly by the be-haviour of the system as a whole, and at the same timetheir interactions lead to the emergent behaviour at theaggregate level of the system. The ’common sense’ connec-tion between the size of an event and its consequences nolonger holds. Small changes have the capacity to triggerlarge scale events [an early demonstration of this is [102].

Complex systems are characterised by critical pointsand regime shifts, an example being public opinion andpro- and anti-war views, attitudes towards smoking bans,and so on. From a policy perspective, these features canlead to serious malfunctioning of the system caused by cas-cades across the network of the system. Examples includeepidemic spreading, the spreading of congestion, powerblackouts and the collapse of the interbank market.

Extreme events in such systems are observed muchmore often than is suggested by the standard assump-tion in applied econometric work that data follow theGaussian, or normal, distribution. The frequency and sizesof floods, earthquakes, wars, financial asset prices and

Farmer et al.: Complex systems for better management of the economy 3

even economic recessions are characterised by highly non-Gaussian outcomes [81].

Techniques used in the analysis of complex systems arequite different from those used in conventional economictheory, with its emphasis on optimisation. They includedata mining, network analysis, systems dynamics, agentbased modelling, non-linear dynamics, catastrophe theoryand the theory of critical phenomena. Much of our em-phasis in extending economic theory is on agent basedmodelling, given that the economy is made up of interact-ing individual agents (people, firms, regulators, govern-ments), each with the capacity to act with purpose andintent, each of which is acting in the context of networksin which the fundamental behaviour of the agent is notfixed, but which evolves in response to the behaviour ofothers.

The network view. One of the biggest successes ofthe complex systems approach has been the developmentof techniques for studying systems in terms of networkmodels. Economics provides a rich set of networks to con-sider, in which the nodes can be many things. Exam-ples that have already received considerable attention in-clude interbank lending [62,23,22,63,61,45], internationaltrade [58,57], corporate ownership [101,15,17,100], andinput-output relationships in production [31]. There re-main many more relationships to be studied. Most impor-tantly, these networks all interact with each other, andthese interactions remain unexplored. The analysis of cou-pled networks is at the forefront of network theory [29,71].

The vision of the economy as a system of evolving cou-pled networks provides a completely different policy per-spective. A key feature is that the behaviour of the econ-omy at the aggregate level emerges from the interactions,both of the individuals within each network and of thenetworks themselves.

Traditional policy recommendations are based on thereactions of individual nodes (people, firms, institutions)to changes in policy. The network approach we proposeopens up the possibility of identifying and targeting keynodes within the system, thereby potentially increasingthe effectiveness of policy. It further makes it possible forpolicy makers to influence the way in which the structureof the network evolves. So, for example, the Basel agree-ments have focused upon controlling and improving theviability of individual institutions rather than on the waysin which they are connected and hence the viability andresilience of the system as a whole. A classic example ofthe failure of this approach was that of Dexia, a Belgianbank which had to be rescued by government interven-tion only 3 months after having passed the official stresstests without any problem. For example, a complex sys-tems network analysis can yield insights into the systemicbanks in the financial network, which must be preservedfrom falling. These latter considerations have only recentlycome to the fore. This point has been heavily emphasizedby the Bank of England (see [52] and [53]), but despitethe establishment of the European Systemic Risk boardthere has been little reaction in terms of macroeconomicmodelling to this issue.

The growth and evolution of the economy is reflectedin the growth and evolution of these networks. The fluc-tuations and changes in the networks during the businesscycle potentially give a much finer grained view of statechanges in the economy than a few aggregate numberssuch as GDP and unemployment. So far this remains al-most unexplored territory.

FuturICT will provide a platform with the raw dataneeded to construct these networks, and to develop thetools to visualize them and analyze their interactions,growth and dynamics. It will provide the capability todo this flexibly, focusing on a given set of issues and thenetworks that are relevant for those issues.

Understanding network structure A key aspect ofthe project will be the use of ICT tools to obtain infor-mation both on the structure of relevant networks at apoint in time, as well as to monitor how they evolve overtime. The models which are built must incorporate the keyfeatures of the network topologies, and have behaviouralrules of agents which allow the structure to evolve overtime. In particular, what was thought of as the key fea-ture of financial networks, their connectivity, has to besupplemented by other network characteristics, such asdegree distribution and clustering, which may provide in-dications of fragility.

In other words, by examining the evolution of severalkey characteristics, we aim to understand the dynamicsof the networks. How do the networks change in time andwhat causes such changes? In an economic system thisinevitably requires an understanding of the agents whomake the decisions that cause the networks to change,though again it has to be emphasized that these decisionsare not taken in isolation and are linked to those of otheractors in the system.

Agent-based models are the natural tool for simulat-ing complex systems in social science. In comparison toeconometric models or the DSGE models of mainstreameconomics, which are formulated in terms of aggregatequantities, agent-based modelling is done at a microscopiclevel. Agent-based models operate at the level of individ-uals, who can be householders, decision makers at firms,or government regulators. Agent-based models make useof computer power to represent as many different agentsas are needed. They do not rely on complex mathematicalderivations or closed form solutions. This makes it easy toimplement nonlinear behaviour without restrictions on thedegree of realism. Off course stylized agent-based models,where (partial) analytical results are available are a valu-able and complementary tool that can yield importantinsights.

In recent years behavioural economists have madegreat strides in understanding how real people behave ineconomic contexts. Agent-based modelling works hand-in-hand with behavioural economics, incorporating its in-sights to model the decision-making of agents, and usingthe power of the computer to simulate their behaviour inthe complex interacting coupled networks discussed above,to keep track of their interactions through their consump-

4 Farmer et al.: Complex systems for better management of the economy

tion, production, budgets, borrowing, lending, flows ofgoods and services, investment, trading, etc.

1.1.2 Data and ICT-tools

The economy generates vast amounts of data that are cur-rently not being gathered and recorded in an integratedmanner, and which can provide much deeper insight intothe workings of the economy than existing data sources.Most current data collection in economics is geared foreconometric and DSGE models, which only require ag-gregate data such as GDP, unemployment, etc. Networkmodelling and agent-based modelling, in contrast, are bestdone with finer grained data, such as information aboutthe choices of individual householders. Information aboutthe heterogeneity of behaviour is essential. Agent-basedmodels can potentially make use of many different typesof information, as described below.

The data collection that we propose here is motivatedby the complex systems picture of the world economy pre-sented earlier. In order to construct the coupled networksthat make up the economy, and in order to validate andcalibrate agent-based models that quantitatively describethe interactions that take place within these networks, wemust collect detailed data. It is beyond the scope of thisposition paper to present a comprehensive plan of whatshould be collected and how it should be collected, butwe give some examples to give the reader a feeling forthe philosophy and the scope of what we have in mind.It is perhaps in this area more than any other where ICTmakes a crucial contribution.

The data that one would like to obtain in order to get arealistic picture of the economy includes trades in financialmarkets with identity information, international trade,firm transactions (invoices and receipts), credit networks,transactions by individual consumers, and electronic textfrom the internet and other sources. While some of thesedata are already collected in piecemeal form, much of itis never collected or recorded, and even when it is, it isoften not available or easily usable by researchers.

A simple example is given by trading on the foreign ex-change market. Whilst traders can potentially trade withthousands of other traders they, in fact, trade with veryfew others. Getting data for the structure of the cluster-ing of trading would be very useful for understanding theevolution of prices on this market.

In another direction, one of the breakthroughs in eco-nomics in the last few decades is the advent of economicexperiments. ICT technology potentially enables this to bedone more efficiently, more comprehensively, and on muchlarger scales, either through use of the web or through useof technology that allows volunteers to be monitored intheir daily life, or data on activity such as cell phone us-age to be anonymously monitored.

Ultimately the economy is about the transformationof physical materials into manufactured goods, and theorganizing of human activity into services. On a longertimescale, data concerning products, technologies andfirms should be recorded so that we can understand more

directly how the economy transforms human activity intomaterial goods, information and services. (Here we define“technology” in a very broad way, to include everythingfrom electronics to new financial instruments to changesin the legal system).

Financial markets. There are hundreds of millionsof trades a day in financial markets. In foreign exchangemarkets alone almost three trillion euro are traded ev-ery day, about 25 times the daily world GDP. From thepoint of view of economic theory it remains a mysterywhy there should be so much trading. Why are foreignexchange markets so active? How much of this is due toreal economic activity, and how much is speculation? Towhat extent do speculators play a useful role in settingexchange rates, and to what extent do they act to desta-bilize them? As already mentioned mapping the networksof traders would contribute to our understanding of thefunctioning of these markets.

We believe that to get proper answers to these ques-tions one has to collaborate with selected governmentsand exchanges to obtain such data for some of the ma-jor markets in Europe, to make it completely anonymousand make it available. This could lead to a break-throughin understanding how markets really function and whatdrives instabilities.

Inter-firm transactions and international trade.There exist substantial data sets documenting interna-tional trade. However, these data are typically highly ag-gregated and contain inadequate detail about the firmsdoing the trading and the products being traded. Theyare also not well integrated with data on domestic trades.One would like a much more fine-grained and texturedview, integrated together with other data that can putsuch data in perspective and illuminate the underlyinginteractions.

For example, to document real economic activity onewould like a record of the invoices and receipts of compa-nies. Each invoice and receipt is for a particular productor service. If one had a record of invoices and receipts,one would have a detailed record of economic activity, arecord of what is made, where it is made, and who makesit, making it possible to track the inputs and outputs ofeconomic production at a detailed level. Most importantly,such a record would chronicle the interactions between dif-ferent goods, and between goods and services. For exam-ple a large part of the trade between the countries of theEU is made up of goods which at an aggregate level arethe same. Germany, France and Italy export cars to eachother. Only a much finer categorisation of these goods willlead to a satisfactory explanation of this trade. Such datamay be collected directly or indirectly. For example whengoods are imported or exported they are insured and theinsurance policy records the nature of the goods and theirvalue.

While it is beyond the scope of even this project tocollect such data comprehensively, by working togetherwith a selected group of the largest international firms,and with cooperation (and some pressure) from the gov-ernments of a few EU countries, one could begin collecting

Farmer et al.: Complex systems for better management of the economy 5

such data and, by so doing, illustrate its power for gettinga more fundamental understanding of how the economygrows and why such growth fluctuates in time. This couldalso be put together with financial and banking transac-tions to better understand the linkages within the econ-omy.

Credit networks. The recent crisis made it clear howessential credit is to the functioning of the entire economy.It also made it clear how poor our information is aboutcredit – the Federal Reserve Bank of the United States,for example, did not have the ability to track U.S. creditmarkets during the crisis, even though they have regula-tory authority over U.S. banks. This has triggered a bigpush by central banks around the world to have bettercapabilities for tracking credit markets, and is leading tothe collection of new data sets. The U.S. Federal ReserveBank now has, for example, a detailed record of the trans-actions of all the commercial banks in the US on a minuteby minute basis. Similar data collection projects are un-derway by the Bank of England and the European CentralBank.

There are serious confidentiality problems with pub-lic access of these data sets. We believe, however, that byworking in partnership with these banks to create datawhich has been made completely anonymous, it should bepossible to make useful data about credit networks avail-able to researchers.

Transactions by individual consumers.There are billions of consumer transactions everyday.

These are increasingly being recorded in electronic form,which provides an enormous opportunity to understandconsumer behaviour. Many retail firms analyze these indetail to better understand how to market their goods;this unfortunately is also an impediment because it meansthat these data sets have considerable proprietary value.While some analysis has been published in the marketingliterature, a great deal remains to be done. We intendto develop collaborations with retail firms to develop thepotential of these data.

Electronic text. One of the central lessons of eco-nomics is that expectations are important. Unfortunately,however, they are very difficult to measure. The currentmethod for measuring expectations is through surveys,but this method is slow, expensive, time consuming, andunless done very carefully, statistically unreliable. ModernICT technology presents us with a much more efficient al-ternative: The modern world abounds in electronic textflows. These include an enormous number of internet sites,news, social media such as twitter, and mobile phone tex-ting. By gathering this data and developing better meth-ods for text analysis we have the potential to measureexpectations in real time. Such measurements could beextremely useful for economic modelling, for example tomonitor sentiment and make use of this in agent-basedmodels.

Individual expectations are influenced by beliefs andopinions of other individuals through social networks. Thisraises the issue of contagion and the pace of diffusion of

expectations and beliefs through social networks, whichneeds to be taken into account in agent-based models.

1.2 Opportunities

At the outset we quoted Governor Trichet, who has issueda challenge to create new tools for understanding the econ-omy. Such tools will be designed to answer the questionsof most interest to policy makers. The agent-based com-plexity models that we are proposing will provide an al-ternative method for answering these questions, that willgive a different perspective than current models. Thereare two standard approaches. Firstly there are economet-ric, or time series models which seek to find structure inthe macroeconomic data for economies. Secondly, dynamicstochastic general equilibrium (DSGE) models, which arethe benchmark of modern macroeconomics. These arebased on the rational economic individual who is capa-ble of understanding the evolution of the economy andacting accordingly.

We intend to base our models on a very different vi-sion. Under our approach individuals may have varyinginformation, depending on what is behaviorally plausible.Aggregate phenomena are generated by the interaction ofsuch individuals, who influence each other and interactwith each other. Such an approach requires intensive useof ICT both for simulating our models and for gatheringthe data against which to test them. To this end, an im-portant goal of FuturICT is to build a “flight simulator”for policy makers that can be used to obtain experiencein managing financial-economic crises. The simulator thatwill be built will use advanced ICT-tools collecting realdata and having state of the art graphics, to make it easyto visualize what the simulator is doing and what the out-comes of the simulation and different policy measures are.

An additional and important by-product of our projectwill be the increased accessibility of the data that wegather. As it becomes available this will enable a muchwider audience to develop a more comprehensive under-standing of the nature and evolution of our socio-economicsystems.

1.3 Challenges

The grand challenge for us here is to greatly expand thestate of the art in economic analysis and forecasting byusing agent-based complexity models, along with integra-tion of real-time data and real-time simulation throughadvanced ICT tools, to enable policy makers to monitorthe global economy and detect potential crises in an earlystage and to develop tools to prevent them. Particularchallenges that stand out are the following:

1.3.1 Financial Stability

An important challenge for FuturICT is to explore the(in)stability and resilience of global financial markets fo-cussing on issues of agent heterogeneity, network effects,

6 Farmer et al.: Complex systems for better management of the economy

spreading of information, market psychology, social learn-ing, and expectations. At the same time, many instabili-ties in financial markets are driven by nonlinear instabil-ities that are inherently mechanical or that derive frommarket structure. (Good examples are nonlinear instabili-ties driven by derivatives ([25], [78], or leverage ([96], [44],[25]). We will use the complex systems network approachto understand the nonlinear feedbacks that exist in mar-kets more comprehensively, explore their interactions witheach other, and in particular to model their interactionswith human decision making.

There are many questions to be addressed with this ap-proach: What are the causes of extreme events and crises?What preventive measures should be taken? If a crisis doesoccur, how should it be managed from a complexity per-spective? What is the role of financial innovation? Howcan institutional design and market regulation contributeto the stability and resilience of global financial markets?It is now widely believed that modern markets have be-come much more vulnerable to sudden changes as a resultof the development of automatic trading algorithms. Thelatter, for example, often incorporate stop loss instructionsto sell when a price descends to a certain level. If manyalgorithms have the same thresholds this can lead to acascade of sales as in May 2010 on the NYSE. This inter-action between modern technology and market dynamicswill play an essential role in the way FuturICT models fi-nancial markets. A question that is now being treated in arather pragmatic way is how to limit the impact of auto-matic trading algorithms on price dynamics. This projectwill make concrete proposals as to how to do this withoutdisrupting basic market functions.

1.3.2 Macrostability

Macroeconomics and business cycles have been studiedfor a long time but what we are proposing is a new per-spective. A macroeconomy is a complex system, and cur-rent economics lacks a satisfactory theory of the sort offluctuations in economic activity that are empirically ob-served. The EU macroeconomy is a multi-scale complexsystem of interacting national economies, where problemsof small countries (Greece, Ireland, Portugal) may conta-giously spread and threaten macrostability of big countries(Germany, France, UK) and even the global EU and worldeconomy.

Once we accept that the volatility of financial marketsis liable to increase rapidly from time to time withoutany exogenous shock, then we need to examine propos-als that are deliberately designed to mitigate volatility. Acase in point is the “Tobin Tax” on financial transactions.The basic argument is extremely simple: Adding frictionwill hinder large amounts of short term trades for verysmall profits. The assumption here is that such trades aredestabilising, but this has to be shown, and the experiencewith markets where this tax exists already is inconclusive.Some have argued that it will almost certainly decreaseliquidity by reducing market making, and therefore in-crease volatility. The Taiwan Stock Market, for example,

has a substantial Tobin tax (25 basis points) and is amongthe most volatile. Thus, although the tax is being widelyproposed as a way to reduce volatility in markets, theresults of implementing it are far from obvious. This isjust the kind of question that we intend to address usingagent-based complexity modelling.

The current crisis has shown once more the importanceof the feedback between the macro economy and world-wide financial markets. The global economy is a multi-scale complex system and a multi-disciplinary approachis necessary to study its functioning. In particular, inter-actions and feedbacks between financial markets and themacro economy need to be studied to understand crisesand improve their early detection and to develop new com-plexity based economic policy.

1.3.3 Expectations and learning in a complex economy

Expectations feedback and adaptive behaviour throughlearning are key ingredients distinguishing socio-economicsystems from complex systems in the natural sciences. Ineconomics the “particles think”: They learn from expe-rience and adapt their behaviour accordingly. A socio-economic complex system is an expectations feedback sys-tem between individual learning and emerging aggregatebehaviour. A fundamental question is: What is the rela-tionship between heterogeneous individual learning at themicro level and the emerging aggregate macro behaviourwhich it co-creates? This fundamental question may be ad-dressed with stylized agent-based models with few agenttypes, to get insights in the interactions of heterogeneousrules and their aggregate behaviour. But a daunting chal-lenge is to address interaction through simulations incor-porating thousands or millions of highly heterogeneousagents, which do not reduce to simulation of interactingpopulations of a few types of agents, using advanced ICTtools and programming levels.

An empirically grounded theory of heterogeneous in-dividual expectations and social learning is needed as afoundation for a complexity research programme in eco-nomics. Empirical testing of such a theory, both at themicro and at the macro level – through laboratory, fieldand web experiments and in empirical financial-economicdata – should yield key insights into which emerging pat-terns are most likely to occur in complex economic envi-ronments and how policy makers can manage expectationsdriven socio-economic complex systems.

1.3.4 Inequality

Where does inequality come from? The core model of mod-ern economics, general equilibrium theory, does not enableus to say anything about the distribution of income. Theissue of inequality cannot even be addressed in the stan-dard representative agent model in macro. Is inequality ainherent feature of complex systems with heterogeneousagents? But inequality is not just a matter of the distri-butions of income and wealth. A major concern of policy

Farmer et al.: Complex systems for better management of the economy 7

makers about, say, outcomes in health care or educationacross hospitals and schools is that such outcomes are “un-equal” or “inequitable” in the key sense that they differat any point in time. This seems inevitable in any com-plex system of interacting agents, but it is often a majorconcern to voters and hence to politicians.

Inequality also has an important geographical (e.g.North-South) component, which poses an additional chal-lenge. Simulations should include agents that live in het-erogeneous environments.

Distributional properties have an important impacton aggregate economic outcomes; for instance, it is well-known that the effectiveness of tax-cut policies aimed atsustaining consumption demand depends on the incomedistribution of tax payers. Other distributional issues areprobably still undervalued in economics, as for instancethe distribution of debt among and within the differentsectors of the economy: public, corporate and households.

Policy makers face the problem of predicting and con-trolling distribution in an economy. For example, what arethe best strategies to achieve a more egalitarian society,given recent findings by British sociologists [103]? Agent-based models naturally take into account the distributionof economic variables at individual levels and thereforecan be valuable tools for policy design. A particularly im-portant question is that of intergenerational inequality. Towhat extent should we make changes now to protect futuregenerations?

1.3.5 Sustainability

The problem of protecting both current and future gen-erations is manifest when considering the supply of food,water and energy, and understanding how we can maintaina high standard of living for the whole world without de-pleting natural resources and destroying biodiversity. Howdo we avoid or at least mitigate these key social, economicand security-related problems? These extremely difficultproblems require trans-disciplinary teams and play to thestrengths of FuturICT .

At present the state of the art for understandingthe dynamic aspects of sustainability is system dynam-ics modelling, as used in the original Club of Rome study,Limits to Growth. Such models have certainly been use-ful in providing an understanding of the relationships be-tween key components of earth systems. FuturICT willexpand considerably the analytical techniques which areapplied to such systems, beyond that of ordinary differ-ential equations. For example, network theory, complexsystems theory, multi-agent simulations, multi-level mod-els, experiments and participatory platforms. We also takeinto account explicitly spatial and network effects, as wellas heterogeneity of agents and randomness. FuturICT willalso use new methods of investigation such as the Liv-ing Earth Simulator, the Planetary Nervous System, theGlobal Participatory Platform and Interactive Explorato-ries.

Another important area is climate mitigation. Whatare the optimal strategies to deal with climate change it-

self and its consequences, and how much will they cost?Current climate mitigation models assume general equi-librium. Production decisions maximize the utility of arepresentative agent, a typical person who exemplifies theaverage worker and consumer. Such models assume fullemployment and assume that firms have no unused inven-tories – everything that is produced is consumed. Sinceby assumption industries operate at full capacity thereis no need to stimulate demand. Such models count thecosts of converting to new technologies without giving anyweight to the economic stimulus that such conversionsmight generate, i.e. they do not allow for the possibilitythat developing new technologies might put people whoare otherwise unemployed to work, and thus stimulate de-mand and make the economy operate at a higher capacity.Perhaps even more important, they typically make highlyquestionable assumptions about technological progress (asdiscussed in more detail in the next section). Such modelshave never been back-tested against historical data andtheir predictive accuracy is highly questionable.

Building on the other work outlined in the FuturICTproject in developing an agent-based model of the worldeconomy, we intend to develop an alternative type of cli-mate mitigation model. By constructing the model at thelevel of individual agents we have far more historical datathat can be used to calibrate the model. By making useof results from behavioral economics we do not need toassume that agents are rational – we can instead use deci-sion rules that have been calibrated against the behavior ofreal people. By collecting and calibrating against an exten-sive database on technological change, as described in thenext section, we will employ more realistic models of tech-nological progress. Most importantly, we do not have toassume that the economy is in equilibrium. We will studythe effects of stimulus for new technologies, modelling theconsequences of putting more people to work and makingprogress in new technologies, e.g. possible revolutionarytransformations to a green energy economy.

1.3.6 Technological progess and economic growth

At the same time that modern economic growth theoryembraces technological progress as the agent of changeunderlying growth, it has traditionally dealt with it ina very simple way. In a typical economic growth theorya technology is a black box embodied by a very simpleproduction function. In contrast, Arthur and others haveargued that to understand the patterns of technologicalprogress one needs to look inside the black box and care-fully model its constituents and their interactions witheach other [8,11]. Technologies are recursively built outof other technologies, and technological change happensin an evolutionary manner: Existing technologies are re-combined and only occasionally are genuinely new tech-nologies created. Indeed, it has been argued that mostinnovations result from the specific demands of producersfor improvements in their production technology. Recentwork ([12], [79]) models the way in which the components

8 Farmer et al.: Complex systems for better management of the economy

of a technology depend on each other, and shows that us-ing simple models for technological improvement, the rateof technological change and its diffusion depend on theinterconnectivity and separability of the technology.

In FuturICT we intend to gather detailed informationabout technologies so that we can empirically and theo-retically study the interaction graphs of technologies. Thiscould potentially allow us to construct a theory of tech-nological change that would be completely different thanpresent economic theories, and that could have substantialpredictive power about the circumstances required for eco-nomic growth. Indeed, empirical analysis done along theselines already suggests the value of this approach ([58,57]).

2 State of the Art

2.1 Traditional economics

Most of current economic theory is based on the postulatesof the agent who is rational in the following sense of theterm. The agent gathers all available information whichis relevant to a decision. The behavioural decision rulewhich is then used is that the agent makes the optimaldecision, given the information and given his fixed tastesand preferences. The decisions of others only affect theagent indirectly via the effects of such decisions on the setof prices facing agents. Agents operate autonomously andare not influenced directly by the decisions of others withtheir tastes and preferences held fixed.

With this as the model of agent behaviour, the majorchallenge of economic theory from the time of Jevons andWalras in the 1870s was to discover the most general setof conditions under which an existence proof of generalequilibrium could be established. In general equilibrium,supply and demand balance in all markets, and there areno unused resources, so it is efficient in that sense. Impor-tant contributions were made by, for example, von Neu-mann [80] and Arrow and Debreu [7]. The main challengein the decades around 1950 was to prove existence in aworld in which time existed, and the final results on thiswere obtained in by Radner [3]. Shortly afterwards, it wasdemonstrated that in general equilibrium no a priori con-straints could be placed on the shape of market demandand supply functions ([93], [37]). Further, that there wasno theoretical presumption that factors of production werepaid their marginal products [21]. In short, general equilib-rium contains no clear testable propositions which wouldenable it to be refuted empirically. A detailed critique isgiven, for example, by Kirman [66]; see also [42] for a dis-cussion of the limits and strengths of equilibrium.

The essential problem which is paramount in theframework that we will develop is how is aggregate be-haviour related to individual behaviour. In other wordsthe problem is that of aggregation.

In macroeconomics this problem was circumvented bymaking the assumption that the aggregate behaved likea rational individual. The representative rational agentmodel which is based on this assumption has become the

standard tool of analysis. Moreover, modern macro dy-namic models assume rational expectations, that is theyassume that all the agents correctly specify the stochasticprocess that governs the evolution of the economy. Theythen find equilibria which satisfy this hypothesis ([74],[75], for example). As an empirical tool, the model of theeconomically rational agent has essentially been the intel-lectual basis of a great deal of both social and economicpolicy. Much of the latter is based on the principle thatagents react individually and “rationally” to changes in in-centives such as tax and benefit rates. Under the guise ofdynamic stochastic general equilibrium (DSGE), in recentdecades the rational agent has become the foundation ofmainstream macroeconomic models as well ([20], [104]).A key development over the past forty years or so hasbeen to incorporate imperfect information into this modelof rationality. In particular, situations in which differentagents have access to different amounts of information,so-called ’asymmetric information’. The pioneers in thisfield were Akerlof [2] and Stiglitz [95]. This relaxation ofthe assumption that agents have complete information un-doubtedly extends the empirical relevance of the model ofthe economically rational agents. The model of economi-cally rational agents operating with imperfect informationis the intellectual basis of the ’market failure’ approach toeconomic policy, and specifically to regulatory policy, inthe past few decades. However, the model is still basedon the assumptions a) that agents have fixed tastes andpreferences and behave optimally, b) that agents operateautonomously and are not influenced directly in their be-haviour by other agents, and c) that agents have rationalexpectations.

2.2 Recent Developments

Bounded rationality In the last decades the aware-ness of the limitations of the rational paradigm amongeconomists has steadily increased. In fact, these limita-tions have been stressed already in the past by well knowneconomists. For example, Pareto said that people spentsome of their time making non-rational decisions andthe rest of their time rationalising them. Keynes empha-sized that “expectations matter” and stressed the impor-tance of market psychology (“animal spirits”). Simon [92]forcefully argued that economic man is boundedly ratio-nal, unable to compute optimal decisions, but instead us-ing satisficing rules of thumb. Tversky and Kahnemann[99] showed in laboratory experiments that individual de-cisions under uncertainty are much better described byheuristics, and that these may lead to systematic biases.Schumpeter and Hayek already had a complexity view ofeconomics with competing strategies and evolutionary se-lection. But these ideas did not withstand the rationalexpectations revolution in macroeconomics in the 1970sand 1980s.

Adaptive learning Modern macro economics hasalso been influenced by ideas from bounded rationality(see e.g. Sargent [90]) and studied macro fluctuations un-der adaptive learning, where agents use a statistical model

Farmer et al.: Complex systems for better management of the economy 9

and update the parameters based on observable quanti-ties to form expectations about the future (an extensiveoverview is Evans and Honkapohja [39]). Sometimes adap-tive learning converges to rational behaviour, but more of-ten than not it does not do so or, if it does, only extremelyslowly. The latter observation is important because it sug-gests that most of the time even if it were theoreticallypossible, people will not have come to have ”rational ex-pectations” and that therefore it is the situation wherethis assumption does not hold that must be analysed.

Behavioural economics is developing alternativemodels for non-rational behaviour (e.g. altruistic be-haviour, reciprocity, tit-for-tat, etc.) and behaviouralgame theory focuses on strategic behaviour which is notfully rational in the standard sense (see e.g. [30]). Labo-ratory experiments with human subjects have become astandard tool to test behavioural models, since one of itsfounders, Vernon Smith received the Nobel Prize in Eco-nomics 2002 (together with Daniel Kahnemann for study-ing the role of psychology in economics); see [33] on theimportant role of laboratory, field and web experimentswithin FuturICT. The evidence that people do not obeythe standard axioms of rationality does not mean that weshould develop other axioms but rather base our mod-els on behavioural rules which seem to be consistent withobserved behaviour. Akerlof and Shiller [3] have recentlystressed the importance of ”animal spirits” and the urgentneed for behavioral economic modelling; see also Colanderet al. [35] for a general critique on the failure of economicsin the light of the financial crisis.

Behavioural finance argues that some financial phe-nomena can be understood using models in which at leastsome agents are not fully rational (see e.g. Barberis andThaler [14]). In particular, Kahnemann and Tversky [64]developed an empirically supported alternative model,prospect theory, for decisions under risk as an alternativeto traditional expected utility. Prospect theory incorpo-rates psychological effects into economic decision making,such as the fact that there is an asymmetry between per-ceived profits and losses: individuals are less willing togamble with profits than with losses. The advantage ofthe sort of model we propose in FuturICT is that we donot have to assume that all individuals behave accordingto prospect theory, for example, but we can allow for moreheterogeneous types of behaviour.

Agent-based models. In the standard efficient mar-ket model, the hypotheses of rationality and homogeneityare normally invoked, so that the agents’ individual be-haviour can be neglected and strategies have all to beequal. As an alternative, taking individual interactionsinto account, agent-based models of financial markets havebeen studied extensively, e.g. the artificial Santa Fe stockmarket in Arthur et al. [9] and LeBaron et al. [70] and theGenoa Artificial Stock Market (GASM) [84,84,83]; see thesurvey [69]. Simpler, stylized models with essentially thesame results have been introduced e.g. in Kirman [65],Brock and Hommes [28] and Lux and Marchesi [77]; seethe survey [59]. Many more ABMs in finance have been

developed and these capture the above mentioned stylizedfacts nicely (see [76] for a recent survey).

ABMs have also been developed to describe macroeconomic fluctuations, e.g. in Delli Gatti et. al. [38]. Inan ABM scenario, empirical macroeconomic regularitiesshould be expressed in terms of statistical distributions,such as the distribution of firms according to their size orgrowth rate (Steindl [94]). Also, suitable modelling strate-gies should be adopted and these should be capable ofcombining a proper analysis of the behavioural character-istics of individual agents and the aggregate properties ofsocial and economic structures (Sunder, 2005).

A recent and prominent example of a large-scale ar-tificial economy has been developed within the EU-FP6project Eurace. The model captures a rich scenario of in-teractions between real and financial variables. This pointsout the validity of Agent-based Computational Economics(ACE) as innovative methodology for the study of eco-nomics. Eurace is a large-scale agent-based model andsimulator representing a fully integrated macroeconomyconsisting of three economic spheres: the real sphere (con-sumption goods, investment goods, and labour markets),the financial sphere (credit and financial markets), andthe public sector (Government and Central Bank). Fol-lowing the agent-based approach, Eurace economic agentsare characterized by bounded rationality and adaptive be-havior as well as pairwise interactions in decentralizedmarkets. The balance-sheet approach and the stock flowconsistency checks has been followed as a key modellingparadigm in Eurace. The computational results show thereal effects on the artificial economy of the dynamics ofmonetary aggregates, i.e., endogenous credit money sup-plied by commercial banks as loans to firms and fiat moneycreated by the central bank by means of quantitative eas-ing [32,85]. In particular, Eurace shows the emergence ofendogenous business cycles which are mainly due to theinterplay between the real economic activity and its fi-nancing through the credit market, thus shedding light onthe relation between debt, leverage and main economic in-dicators [85]. Furthermore, Eurace shows that a quantityeasing monetary policy coupled with a loose fiscal policygenerally provide better macroeconomic performance interms of real variables, despite higher price and wage in-flation rates [32]. See also [36] for a recent discussion ofagent-based macro model related to Eurace.

One of the standard misconceptions about large agentbased models (ABM) is that they are difficult to calibrateto empirical data. However, the main purpose of ABM, israther to be able to reproduce qualitatively some of thesalient features of the real economy. It is not to fit the datain the way that is the ambition of econometricians. For thelatter, one can assume that one has a good understandingof the stochastic process followed by the economy, in whichthere are a number of relatively standard tools such as”bootstrapping” available.

One of the standard problems with large agent basedmodels (ABM) is that they are difficult to calibrate to em-pirical data. For a number of economists the main purposeof ABM is to be able to qualitatively reproduce some of the

10 Farmer et al.: Complex systems for better management of the economy

salient features of the real economy. When the stochasticprocess is imperfectly known or exhibits non-stationaritythe econometric approach has to be modified and one canno longer think of a system converging to equilibrium, (see[34]).

One of the major challenges that FuturICT proposesto meet is to provide approaches that allow one to fit thedata with ABM models even when the stochastic processis imperfectly known or exhibits non-stationarity and thesystem does not converge necessarily to an equilibrium;see e.g. the discussion of an ABM for the housing marketin the Washington DC area in Section 3.6.

2.3 The fundamental limits of mainstream thinking

We recognise that economics has made advances in recentdecades, and that it is by no means a completely emptybox. However, to quote again from Jean-Claude Trichet:“As a policy maker during the crisis, I found the availablemodels of limited help. In fact, I would go further: in theface of the crisis, we felt abandoned by conventional tools”.

A similar attitude towards the crisis in the autumn of2008 was adopted by policy makers in the United States.Olivier Blanchard [20], chief economist at the IMF, hadwritten in August of that year, only a few weeks beforeLehman Brothers failed, that “For a long while after theexplosion of macroeconomics in the 1970s, the field lookedlike a battlefield. Over time however, largely because factsdo not go away, a largely shared vision both of fluctua-tions and of methodology has emerged The state of macrois good”. The state of macro is good! In August 2008! Hewent on to say “DSGE models have become ubiquitous.Dozens of teams of researchers are involved in their con-struction. Nearly every central bank has one, or wants tohave one”.

Despite this convergence of academic opinion aroundDSGE models, in the crisis American policymakers paidno attention to them whatsoever. Instead, they looked atwhat happened in the 1930s and tried to avoid those mis-takes. So, for example, they nationalised the main mort-gage companies, effectively nationalised AIG, eliminatedinvestment banks, forced mergers of giant retail banks andguaranteed money market funds.

It is therefore necessary to move beyond, and in manyways to break decisively, with the fundamentals of main-stream thinking (e.g. [55]). The rational agent hypothe-sis is still pervasive in economics, however it might bequalified by mainstream practitioners. But the conditionsunder which the existence of general equilibrium can beproved ([86]) are so restrictive as to be completely divorcedfrom any empirical picture of the real world. Human cog-nitive capacities are known to be bounded ([91]). In theface of NP-hard optimisation problems, even supercom-puters are facing limits so that optimisation jobs cannotbe performed in real time any more (so, for example, thereare 32 pieces at the start of a chess game, but computershave only been able to solve all possible positions whenthere are just 6 pieces on the board, and the computa-

tional task scales super-exponentially with the addition ofeach extra piece).

Above all, it is the equilibrium paradigm and the rep-resentative agent approach which we are intending to re-place. From a policy perspective, the efficient market hy-pothesis has been very important in practice. This hypoth-esis implies the equilibrium paradigm. Yet, for example,financial markets are systems of extremely many dynam-ically coupled variables, and it is not at all obvious theo-retically that such a system will have a stationary solution([72]). If a stationary solution exists, it is not clear thatit is unique ([1]. Further, as is well known since the for-mulation of the Lotka-Volterra equations in the 1920s, insystems of non-linearly interacting variables, the existenceof a stationary solution does not necessarily imply that itis stable.

In short, the equilibrium, efficient markets paradigmcannot explains phase transitions; it does not allow us tounderstand innovations which occur as a result of endoge-nous system dynamics; it does not permit us to studythe effects of different time scales of response, when, forexample, fast self-reinforcing effects and slow inhibitoryeffects may lead to pattern formation in space and time([98]), and neglects friction, without which it is difficultto understand entropy and other path-dependent effects.To remedy some of the deficiencies of the standard equi-librium approach to economics and the social sciences ingeneral, a number of the partners in FuturICT are alreadyparticipating in a project (NESS) on non equilibrium so-cial science (for details see http://www.nessnet.eu/).

A common simplification in economic modelling is thatof the representative agent. This is certainly a curious as-sumption to make in the light of the financial crisis, whenmuch of the focus is on the differing behaviors of creditorsand debtors. In the Euro area crisis at the time of writing(November 2011), for example, it makes no sense at all tospeak of the ’representative agent’, the German govern-ment has quite different behavioral rules and constraintsfrom that of, say, the Greek and Italian administrations.

A fundamental problem with the simplifying assump-tion of the representative agent is that it cannot ex-plain the basic existence of economic exchange. As Arrowpointed out ([6]): “if we did not have [agent] heterogene-ity, we would have no trade”. A similar point was made byKeynes in his 1936 magnum opus, The General Theory ofEmployment, Interest and Money, when he noted that ifall agents had identical expectations, market prices wouldfluctuate between zero and infinity.

In modern social and economic systems, percolationphenomena are ubiquitous. In other words, ideas, marketsentiments, technologies, viruses spread across networks.The actual spread of any individual example is not deter-mined so much by the average node degree, but by veryparticular features of network structure (e.g. [82]). More-over, both specialization and innovation imply the exis-tence of heterogeneous agents, and the latter in particular

Farmer et al.: Complex systems for better management of the economy 11

cannot be understood using the concept of the represen-tative agent1.

3 Innovative approach

3.1 Networks

Networks provide a good starting point to visualize anycomplex system. The economy can be thought of as a setof overlapping and strongly interacting networks. Some ofthe key networks needed to understand the economy inthese terms include:

– Production A good or service is a node, the goods andservices required to produce it are its incoming links,and the goods and services it helps produce are itsoutgoing links.

– Firms. Individual firms are nodes, their incoming linksare firms they invoice, and their outgoing links are thefirms they get receipts from. This is strongly overlap-ping with the production network described above, butfocuses on institutions rather than goods. One can or-ganize the production and firm networks based eitheron monetary flows or on the flow of physical goods.

– Ownership. This is an inherently bipartite graph, withtwo types of nodes. One type of nodes are owners, whocan be individuals or firms, and the other type of nodesare assets, which can be firms, commodities, bonds,derivatives, etc.

– Trading. This is closely related to the ownership net-work and describes with which other holders ownerstrade assets.

– Credit. Banks are specialized firms that lend money toindividuals and other firms. The credit network is abipartite graph in which one type of nodes are banks,and the other type of nodes are the individuals andfirms, including other banks, that they lend to. Sincebanks also own assets and hold them as collateral theyplay a prominent role in the ownership and tradingnetworks as well.

– Individuals. Ultimately the economy is about individu-als, who live in households, use their labour to produceor supply their labour to firms, and consume. Theseindividuals are connected to the networks above in avariety of different ways – they produce and consumegoods, work for firms, invest in firms, trade in stockmarkets (possibly through their pension plans), andreceive credit from banks. While it may be intractableto visually represent all 500 million individuals in theEuropean Union, this network is the vital conceptualcomponent of any economic model – the very fact thatit exists explains what is meant when we talk of aneconomy.

The networks delineated above are not comprehensive.One can refine these, e.g. by breaking assets down into

1 For a review of what equilibrium theories have accom-plished, what they cannot accomplish, and what is needed togo beyond them, see also Farmer and Geanakoplos [41].

different types, or looking at more specialized networks,such as the movement of workers between firms or thenetwork of patent citations and their relationship to firmsor physical goods.

There are flows of many different quantities along thelinks in the networks above. For the production network,for example, one can monitor the flow of money, physicalgoods, energy, pollutants such as CO2, or critical mate-rials. This makes it easy to relate economic activity tophysical activity and resource constraints – a perspectivethat is often missing from conventional economic analyses– and makes this approach particularly useful for studyingsustainability.

Breaking down a system into a set of interacting net-works is only the first step in a complex systems analy-sis, but it can nonetheless be very useful by providing aschematic view of what the key components of a systemare and how they interact. For example, see the discus-sion of leverage in Section 3.4 or the discussion of tech-nological growth in Section 1.3.6. The FuturICT programwill gather the data required to construct all the networkssketched above and more, and will use this representationas a basic tool for understanding the economy and its in-stabilities. This will go far beyond the state of the art –few of these individual networks have been constructed inany detail, and their interactions have largely never beenstudied. This will be by far the most comprehensive viewof the world economy ever constructed.

3.2 Market ecology and its role in financial instability

Real agents have cognitive and resource limitations thatforce them to specialize. They build up specialized knowl-edge based on past experience and make most decisionsusing rules of thumb, without going through the effort ofdeductively thinking each situation through from scratch.The diversity of behavior can be enormous. The ecologi-cal approach classifies different types of behavior, dividingagents into groups in order to simplify the study of theirinteractions and evolution [9,40].

Under this view small deviations from market effi-ciency support an ecology of interacting financial traders.This ecology is supported by the interactions of the finan-cial sector and the real economy: Basic economic func-tions, such as trading to get liquidity or offset risks, gen-erate profit-making opportunities. These are exploited byhighly specialized financial speculators, who play the roleof predators in a financial ecology. The diversity of strate-gies employed is enormous, including fundamental valueinvestors, technical traders, statistical arbitrage, high fre-quency trading, and many others.

As in biology, predators play a valuable role in regu-lating and maintaining balance. The introduction of newspecies can disrupt this balance, destabilizing an ecologyand causing irreversible changes (such as the introductionof rabbits into Australia). Similarly, financial strategiesare constantly evolving and new strategies may desta-bilize the market. For example, the introduction of newmortgage derivative products and the misunderstanding

12 Farmer et al.: Complex systems for better management of the economy

of these products is widely believed to have played a keyrole in triggering the current economic crisis.

Agent-based modelling and the study of market ecol-ogy provide a natural fit. The agent-based models ofLeBaron, for example, suggest that market instabilitiesare associated with decreases in the diversity of financialtraders [68]. This is not surprising – if there is no diver-sity in markets, there is no one to take the opposite sideof a position, which can give rise to extreme price move-ments. FuturICT will collect data to better study marketecologies, with the goal of developing early warning sig-nals based on shifts in the profit relationships betweenstrategies and shifts in market ecology.

3.3 A complex financial market

As an illustration, consider an example of a market, butthis time for a financial asset, ([4]). This starts with anempirical phenomenon, the collapse of the price of assetbacked securities early in the current crisis. The paperfirst presents a simple theoretical model to capture theessence of the phenomenon and then ran simulations of amore general dynamic model in which the agents act in thesame way as in the theoretical model to see if the modelevolves to the states predicted by the theory. The goalwas to model the general mechanism whereby investors,as a rule, trade securities without giving due diligence tofundamental information that is, they do not check on thetoxicity of the asset. The rationale motivating investors,is simply that it is profitable to adopt this rule, becauseother investors have already adopted it.

The market consists of agents, who, in the case of thesub-prime crisis, we can think of as the banks who wereboth the issuers and the investors in these Asset BackedSecurities,(ABS). Each agent decides whether or not tofollow a rule, which is to purchase an ABS, relying onsignals from the rating agencies, without independentlyevaluating the fundamental value of underlying assets. Ifenough other participants do so, the agent becomes con-vinced, not irrationally, that the ABS is highly liquid andhence easy to trade.

The ABS is toxic, with a certain probability the under-lying asset was incorrectly graded, (the original borrowerhas already defaulted or has a high probability of doingso). Agents are linked together with trading partners ina financial network. When an agent receives an offer tobuy a new ABS, she considers whether or not to check onthe underlying asset. Each agent now calculates the ex-pected gain to him of checking given the rules chosen bythe neighbours in his network and checks if the expectedpay-off is higher than that obtained by not checking.

It is not difficult to find the equilibria of this simplemarket, in terms of whether agents are checking or not,and there are two, one of which is always an equilibrium,and the other which only appears above a certain criticalvalue for the probability of default on the underlying asset.In the first equilibrium no banks check on the underlyingassets whilst in the second all banks do so. To test thestochastic stability of the two equilibria simulations were

CHAPTER 1 ASSESSING RISKS TO GLOBAL FINANCIAL STABILITY

12

While the U.S. housing sector may finally trough at some point in 2009, continued declines in house prices and sluggish growth are likely to deepen and broaden the default cycle.The combination of tighter lending standards, falling home prices, and lower recovery values would lead to a rise in charge-off rates on resi-dential mortgages from the current 1.1 percent rate to a peak of 1.9 percent by mid-2009, and they could remain elevated throughout 2010 (Figure 1.10).

Pressures on household balance sheets presage deterioration in consumer loans.

Charge-off rates on U.S. consumer loans have risen. In addition, there are rising signs of stress as consumers tap credit lines to support consumption amid higher mortgage and other costs. Additionally, the ability to pay down higher-interest credit card debt with cheaper home equity loans has diminished, suggesting that some consumers are being forced to shift from secured mortgage debt to higher-cost, unsecured credit card debt.6 However, with tighter lending standards, the availability of this type of credit may fall. Our analysis shows that tighter bank lending standards and slow-ing growth are likely to lead to consumer loan charge-off rates of about 3.9 percent by early 2009, slightly above the peak levels of 2002, before falling to more normal levels by 2010 (Figure 1.11). Under a stress scenario, charge-off rates climb to over 4 percent.

Stresses on U.S. consumers are also leading to credit weakening in commercial real estate loans.

Charge-off rates on U.S. CRE loans have already reached decade-high levels, as weaker consumer fundamentals weigh on the retail and condominium sectors. As with other loan categories, credit deterioration has been more pronounced on recently originated (2006–07) loans, which had weaker underwriting standards

6Banks have accommodated this increase so far, partly because credit card securitization has remained relatively robust over the last year.

0

2

4

6

8

10

12

14

16

18

0

5

10

15

20

25

30

35

40

0

0.5

1.0

1.5

2.0

2.5

3.0

Sources: Merrill Lynch; and LoanPerformance.

Figure 1.8. U.S. Mortgage Delinquencies by Vintage Year(60+ day delinquencies, in percent of original balance)

2005

2005

2004

2004

2003

2003

2001

2001

2002

2002

2000

2000

2007

2007

2006

2006

2005

20002001

2002

2004

2003

Months after origination

Alt-A

Subprime

2007

2006

Prime

0 10 3020 40 50 60

0 10 3020 40 50 60

0 10 3020 40 50 60

13

(e.g., higher loan-to-value and debt service coverage ratios). Econometric analysis indicates that private consumption strongly affects the level of CRE charge-off rates. Charge-offs may rise to a 17-year high of about 1.7 percent by the end of 2009, or to 1.9 percent under our stress scenario, remaining elevated for some time, though still below the levels reached in the early 1990s.

Tighter access to credit is pressuring leveraged companies and small and mid-sized enterprises, while nonfinancial investment-grade firms’ access remains relatively robust.

A weakening economic environment is already leading to corporate credit deteriora-tion, especially for firms closely tied to the consumer. Credit quality has deteriorated on leveraged buyout deals in the last few years, as shown by the rising ratio of rating downgrades to upgrades in this sector.7 Secondary market liquidity for leveraged loans remains low and banks and managers of collateralized loan obligations are selling loans at significant losses. Some of these sales have been to private equity firms, partly encouraged by lower prices and seller-provided financing for the purchases. Consequently, the leveraged loan pipeline has declined to $70 billion from a peak of $304 billion in mid-2007, relieving one source of potential stress on asset prices.

High-yield corporate bond issuance has slowed considerably, and firms are facing reduced access, higher rates, and shorter dura-tions on their commercial paper obligations. As the cycle has begun to turn, default rates have started to increase, rising to 2.5 percent. Through mid-September of this year, globally, 57 corporate issuers have defaulted, compared with just 22 issuers in all of 2007.8 The cur-

7A more pronounced deterioration in recent leveraged loans may ultimately materialize where “covenant-lite” agreements may have hindered early intervention by lenders.

8In the United States, the ratio of rating agency upgrades to downgrades on high-yield bonds is at its low-est level in four years.

Figure 1.9. Prices of U.S. Mortgage-Related Securities(In U.S. dollars)

Jumbo MBSAgency MBS

ABX BBBABX AAAAlt-A

Sources: JPMorgan Chase & Co.; and Lehman Brothers.Note: ABX = an index of credit default swaps on mortgage-related

asset-backed security; MBS = mortgage-backed security.

0

20

40

60

80

100

120

2006 07 08

Sources: Federal Reserve; S&P Case-Shiller; and IMF staff estimates.1As a percent of loans outstanding; annualized rate. 2Series standardized over the period from 1991:Q1 to 2010:Q4.

Figure 1.10. U.S. Residential Real Estate Loan Charge-Off Rates(In percent)

Charge-off rate1Estimate

Tighter lending standards;lower home prices

Home price appreciation (inverted, right scale)2

Residential real estate lending standards(right scale)2

–2

–1

0

1

2

3

4

–0.2

0.2

0.6

1.0

1.4

1.8

2.2

1991 93 95 97 99 2001 03 05 07 09

THE DEFAULT CYCLE

0

0.2

0.4

0.6

0.8

1

0.001 0.01 0.1

!

p

Sim. Best responseSim. logit B=1000

B=800B=500B=200Theory

Figure 1: Prevalence ⇡ = E[z] of rule followers, as a function of p, for c = 0.9, k = 11 and an exponentialdistribution for �i with mean � = 0.01. The results of numerical simulations (best response) are comparedwith the theory (full line). Results for noisy best response P{zi = 1} = e

Bui(1)/(eBui(1) + e

Bui(0)), withB = 200, 500, 800 and 1000 are also shown.

Numerical simulations based on the dynamics with both Eq. (4) and Eq. (5), are pre-sented in Fig. 1. The state where everybody follows the rule (zi = 1, 8i) is a fixed pointof the B ! 1 dynamics for all values of p. Indeed, when everybody follows the rule, notrade is ever refused, independently of the quality of the assets, precisely because no-onechecks. For large values of p > pc, however, this fixed point coexists with one where mostof the agents perform their independent credit analysis (zi = 0).

Fig. 1 also shows, however, that the rule dominated equilibrium loses stability for highp, if the intensity of choice parameter B is not large. For intermediate values of B, as pincreases, one has a sharp transition from the state where the vast majority follows therule, to a situation where most of the agents do not follow it.

In order to see how this model leads to herding, imagine starting from a market inwhich there are only fundamentalists, i.e., no one follows the rule (zi = 0 8i). Those agentswith cost �i > pc will find it profitable to adopt the rule. A large cost may be understoodas the securities being very complex, for example after multiple repackaging. This, in turn,this will make it appear to be profitable for some of their neighbors to also follow the ruleand similarly for the neighbors of these neighbors. If the rule spreads to the extent thatz > c all agents, even those for whom information is available for free, will follow the rule,then no one will look for information.

8

Fig. 1. Critical transition in an agent-based model of assetbacked securities (ABS). Top panel: rising percentage of loanswith a given date of origin which default; Middle panel: pricesof ABS are stable, before suddenly collapsing; Bottom panel:critical transition from not checking to checking equilibrium.

run in which agents noisily learn (they use reinforcementlearning), whether to check or not. What transpires fromthe simulations is that the system always converges to theno-checking equilibrium if the probability of default is lowenough, but a small increase in that probability can leadthe market to collapse into the equilibrium in which every-

Farmer et al.: Complex systems for better management of the economy 13

one checks. Thus a small change in one of the parametersof the model can lead to catastrophic consequences at theaggregate level.

Consider Figures 1a-c. In the first we see the probabil-ity of default on an ABS issued at a given time. For eachdate we see the percentage of loans with a given date oforigin which default after each period. This is steadily ris-ing in Figure 1a. In Figure 1b we see the evolution of theprices of the ABS over time, which remains stable and thensuddenly collapses after a certain time. Finally in Figure1c where on the vertical axis is the proportion of agentsnot checking, the noisy best response process is shown col-lapsing to the checking equilibrium as the probability ofdefault increases. If all agents were perfectly rational theprocess would not leave the no-checking equilibrium, buteven a small amount of noise will cause the no checkingequilibrium to suddenly shift to the one where everybodychecks and the toxic assets are revealed to everyone andthe prices collapse. Thus what was examined was the co-evolution of the default rates on mortgages and the pricesof securities backed by those mortgages. The default ratessteadily increased but this was not reflected in the priceof assets until they suddenly collapsed and the interbankmarket froze. A continuous change at one level led to adiscontinuous change at the aggregate level. Whilst onecould establish the existence of the equilibria of the modelanalytically, it was necessary to resort to simulations tosee to which equilibrium the learning process converged.

This underlines an important message. As soon aswe are interested in real economic phenomena we cannotavoid examining how the economy behaves out of equilib-rium and the characteristics of the states through whichit passes, or to which it settles.

Another and very general example of cascading net-works which are a source of systemic risk is given by [73].

3.4 Expectations and learning

In a complex system, individual agents lack the knowl-edge necessary to fully understand their environment andcompute optimal decision rules. How do individual agentslearn in such an environment and adapt their behaviouras the economy evolves and how does individual learningco-create aggregate behaviour? An important question isthe extent to which agents are capable of learning betterheuristics in complex non-equilibrium situations.

A theory of heterogeneous expectations for complexsocio-economic systems has been proposed in Kirman[67] and Brock and Hommes [27]. Agents act purpose-ful, choose between simple forecasting heuristics and tendto attach more weight to rules that have been more suc-cessful in the recent past. This theory of heterogeneousexpectations has been successfully fitted to laboratory ex-periments with interacting human subjects [60].

A key finding is that in systems with negative expec-tations feedback, such as the classical supply driven hogcycle model, subjects are able to coordinate and learn therational expectations equilibrium price, even when theyhave very limited information about their environment

0

20

40

60

80

100

0 10 20 30 40 50

Price

simulation experiment

0

20

40

60

80

100

0 10 20 30 40 50

Price

simulation experiment

0

0.2

0.4

0.6

0.8

1

0 10 20 30 40 50

Impacts of Heuristics

ADA WTR STR LAA

0

0.2

0.4

0.6

0.8

1

0 10 20 30 40 50

Impacts of Heuristics

ADA WTR STR LAA

Fig. 2. Laboratory experiments with negative expectationsfeedback (left panels) and positive expectations feedback (rightpanels) markets. Realized prices in experiments (top panels,red dots) and simulated prices (top panels, green squares) andcorresponding evolution of fractions of 4 strategies in heuristicsswitching model (bottom panels). Negative feedback marketsquickly converge to fundamental equilibrium, with adaptive ex-pectations dominating the market (bottom left panel, purplecurve). Positive feedback markets persistently oscillate, due togood performance and amplification by trend following strate-gies (bottom right panel, blue and green curves).

(see [54] and Figure 2). In contrast, in systems with pos-itive expectations feedback, such as asset markets drivenby speculative demand, agents are not able to learn the ra-tional expectations price, but instead they coordinate onsuccessful and self-fulfilling simple trend following strate-gies causing price fluctuations, excess volatility and persis-tent deviations from the rational expectations fundamen-tal price [60]. These results generalize the initial findingsof [43] that aggregates do not become “well behaved” ifthere is interdependence in agents’ preferences or choices.A simple complexity model thus fits the experimental databoth at the micro level of individual forecasting heuristicsand the macro level of aggregate price data. Many finan-cial market and macro systems exhibit positive feedback,and a challenge for FuturICT is to study heterogeneousexpectations in more realistic complex markets and largeweb-based experiments.

More and more attention is being paid to situations inwhich cooperative behaviour evolves in situations wheregame theory suggests that the solution would be for peo-ple to behave in an individualistic way. See for example[24]. One line of reasoning which will be incorporated intoFuturICT models is that of “team reasoning” as intro-duced in Bacharach [13]. This suggests that people in-volved in a collective venture may redefine their utilityto take account of the welfare of the group. Laboratory

14 Farmer et al.: Complex systems for better management of the economy

experiments point out the importance of networks in theemergence of cooperation [50,97]. Cooperation and coordi-nation are, highly important features of modern societiesand economies and the way in which such cooperationemerges and can be encouraged should be a central themefor FuturICT.

The importance of social learning, of agents alteringtheir behaviour directly as a result of the behaviour ofothers, is shown by the tournament reported in [88]. Avery simple strategy based upon copying proved the mostsuccessful, a result not anticipated by the experts whodesigned the tournament: ’This outcome was not antici-pated by the tournament organisers, nor by the committeeof experts established to oversee the tournament, nor... bymost of the tournament entrants.’

FuturICT will investigate a wide range of behaviouralrules of agents based on different types of social learning,individual learning, reinforcement learning, heterogeneousexpectations models and examine which are more capableof explaining the key emergent macro-economic propertiesof the system than others. In addition to learning modelsbased on principles more familiar to economists, we willdraw on models of social learning behaviour from other so-cial sciences such as anthropology and cultural evolution,to build an empirically grounded theory of learning andadaptive behaviour for complex socio-economic systems(for example [26], [24], [89], [51], [56], [18], [19]).

3.5 Innovative policy and market regulation

A key factor driving the current economic crisis was theexcessive use of leverage, i.e. too much borrowing with-out proper collateral. John Geanakoplos [46] has docu-mented and analyzed the leverage cycle, in which leverageincreases due to competition during times of financial sta-bility. The excessive growth of leverage eventually leads tofinancial instability and triggers a crash; once the financialsystem stabilizes, the cycle repeats itself. The agent-basedsimulation of the leverage cycle by Thurner et al. [96]demonstrates this cycle explicitly, and shows how leveragecan drive clustered volatility and heavy tails in financialtime series. Margin calls, which are made to limit risk andmake good sense from the myopic perspective of individ-ual banks, can drive systemic risk when too many banksact in unison. This provides an example of how financialinnovations, which were in Lord Turner language mistak-enly considered to be “axiomatically beneficial”, may bedestabilizing in a complex system with boundedly rationalinteracting agents (Brock, Hommes and Wagener, [25]).

What level of leverage is optimal? Basel III sets lever-age levels based on rules of thumb and intuition. Thetoolkit of FuturICT can potentially make the regulation ofleverage much more scientific. In a world of complex inter-locking relationships and contracts, leverage is difficult toeven calculate. The proper level for leverage depends onnetwork relationships such as interbank lending (e.g. [49,48]), which can provide security in normal times but mayamplify the extent of a crash in bad times. Bank holdingcompanies are highly complex institutions that often have

thousands of independent entities – without understand-ing the network of ownership and control (see e.g. [101]),one cannot properly measure leverage; this is essential tounderstand what will happen when any given institutionis driven bankrupt. The data and network view of Fu-turICT will make such relationships clearer, and providea framework for making sound regulatory decisions.

3.6 Agent-based models as econometric and policytools

As we have already mentioned, up until now agent-basedmodels in economics have mainly been qualitative con-ceptual models. A project to build an ABM for housingmarkets in the Washington DC area provides a proof-of-principle that it is possible to go beyond this to builda model that is carefully calibrated to real data. Somepreliminary results are given in [47], and the results of atypical simulation are shown in Figure 3. This model canbe used to make predictions about the housing marketconditioned on factors such as interest rates, loan policy,immigration, or unemployment, or it can be used to inves-tigate changes in policy, such as interest rates or lendingpractices.

In this model agents are households who buy and sellhouses. Each house is randomly assigned a quality fac-tor based on the distribution of sale prices of real houses,and the assumption is made that quality remains constantand buyers prefer houses of higher quality. Every month afraction of households decide to move and put their housesup for sale, or attempt to buy a house, e.g. because theydecide it is a better alternative than renting. Sellers droptheir prices at regular intervals until they find a buyer orremove their house from the market.

The thing that makes this model unique is that mostof the characteristics of the households, such as income,employment, frequency of moving, etc. are calibrated toexogenous data. Income, for example, is matched to yearlyincome distributions from tax returns (which also pro-vide valuable information about the frequency with whichhouseholders move). Census data is used to understandmigration and wealth, data on individual mortgages isused to track loan policy, and data from real estate recordsis used to calibrate the strategies sellers use to adjust theirprices.

An example showing a typical run of the model isshown in Figure 3. The first six panels compare six dif-ferent outputs of the model to the data for the period ofthe simulation from 1997 - 2010. Many properties of themodel do a good job of matching the data, such as thehousing price index, units sold, or days on market. Theauthors then investigate counter-factual scenarios to as-sign causality by simulating how the market would havebehaved differently had exogenous factors changed. Forexample, in the lower left panel we see that if interestrates had been held constant there would have still beena bubble, but it would have been much smaller. The lowermiddle panel shows what would have happened if leverage

Farmer et al.: Complex systems for better management of the economy 15

Case Shiller

1998 2002 2006 2010

1.0

1.5

2.0

2.5

Index, first period = 1ModelData

Units Sold

1998 2002 2006 2010

5000

10000

15000

20000

Number*

*Data is smoothed with centered11−month moving average.

Days on Market

1998 2002 2006 2010

050

100150200250300

Days

Sold Price to OLP

1998 2002 2006 2010

0.90

0.95

1.00

1.05

1.10

FractionAverage Mortgage Rates

1998 2002 2006 20105

6

7

8

9

10

Percentage pointsFixed rateAdustable rate

Interest only

Average LTV at Origination

1998 2002 2006 2010

0.6

0.7

0.8

0.9

1.0

1.1Ratio

All originationsTop 50%Empirical data

Case Shiller: Constant Interest Rates

1998 2002 2006 2010

1.0

1.5

2.0

2.5

Index, first period = 1ModelData

Case Shiller: Constant LTV

1998 2002 2006 2010

1.0

1.5

2.0

2.5

Index, first period = 1ModelData

Case Shiller: Constant Interest Rates, LTV

1998 2002 2006 2010

1.0

1.5

2.0

2.5

Index, first period = 1ModelData

Figure 1: Housing Market Results

Fig. 3. A comparison of an agent-based simulation of the housing market in the Washington D.C. metropolitan area (solidlines) to real data (dashed lines).

for buying houses, as measured by the Loan-To-Value ra-tio (LTV), had remained constant. In this case the bubbledisappears entirely. Indeed the scenario where both inter-est rates and LTV are held constant (lower right) is verysimilar to the one in which only LTV is held constant.Thus, this model suggests that the dominant cause of thebubble was the extremely liberal lending policies that pre-vailed during the housing market. Such results would bedifficult to achieve with other more traditional methods.This project is still in its early stages and the results arestill evolving as the model is improved, but the results sofar suggest the potential power of this method.

4 Expected Paradigm Shifts

A success of FuturICT would constitute a major paradigmshift within the field of economics along at least threedimensions:

1. A new scientific paradigm of a complexity-network ap-proach to economics, emphasizing the idea that the

economy is a large but noisy system of interactingagents who are themselves noisy. What is interesting insuch systems is not their efficiency, but their capacityto coordinate;

2. The economic discipline needs to become an empiri-cally falsifiable science, both at the individual and atthe aggregate level; see e.g. the experiments discussedin 3.4;

3. A paradigm shift in economic policy, with the develop-ment of a “flight simulator” with advanced ICT toolsfor policy makers to gain experience in pro-active man-agement of crises in a realistic environment;

4. An analysis of the increased speed of communicationon economic activity and firms’ growth.

As already discussed, up until now the main methodsfor macro economic analysis and forecasting have beeneconometric and DSGE models. A success in this projectwould introduce a third category of model into the stan-dard toolkit. This would be more than just a technicalinnovation: It would be a revolution in the way that eco-

16 Farmer et al.: Complex systems for better management of the economy

nomics is done, and entirely different way of thinkingabout the economy.

In the FuturICT view, economics should become anempirically falsifiable science again, both at the micro andat the macro level. “As if” models at the aggregate levelare insufficient. Instead, the focus will be on interactingagent-models, calibrated to individual and firm specificdata and matching the emerging aggregate properties, andthus matching both micro and macro data. Simple labora-tory macro experiments, such as those described in subsec-tion 3.3, show that matching micro and macro behavioursimultaneously is possible. FuturICT will develop methodsand tools to match large individual data sets to individualbehavior and at the same time match aggregate behaviourof complex socio-economic systems.

In a complex economy, such as that which would bedepicted by the models developed in FuturICT, since theconsequences of individual choices depend on what all theothers are simultaneously doing, people take actions inan environment characterized by radical or endogenousuncertainty. The aggregate outcomes emerging from theircontinuous and asynchronous localized interactions are al-most incomprehensible at an individual level and risk anal-ysis has to be systemic, not individual. In spite of this,modern market economies display a reasonably coordi-nated state of affairs most of the time. For example theystay within a few percentage points of full-employment,and do not exhibit persistent pathological shortages orsurpluses of goods. In other terms, the macroeconomy ischaracterized both by a substantial resilience and a deepfragility.

This opens the way for some fundamental theoreticaland policy questions that FuturICT will address, regard-ing how built-in feedback mechanisms operate in a com-plex economy, and how government interventions shouldboth anticipate and react to them to prevent future crises.As Bernanke said “I just think it is not realistic to thinkthat human beings can fully anticipate all possible inter-actions and complex developments. The best approach fordealing with this uncertainty is to make sure that the sys-tem is fundamentally resilient and that we have as manyfail-safes and back-up arrangements as possible ”

(Interview with the IHT May 17th 2010)In the remaining part of this section we discuss three

of the important questions that arise.In the scenario proposed by FuturICT, economic pol-

icy prescriptions have fundamentally novel aspects. Theystabilize economic system by managing a complex systemwith heterogeneous agents who interact through networks.It evokes a passage similar to that one from chemotherapyto targeted medicine, with individualized treatment. Twocases: i) network of networks, and ii) immunization strat-egy or circuit breakers. Some statistical physicists have re-cently proposed the concept of networks of networks [29,71] where they develop a framework for understandingthe robustness of interacting networks subject to cascad-ing failures from one network to another and vice versa.The FOC project linked to the FuturICT proposal has asits aim to study the relationship between the structure of

these networks and their propensity to implode (see [16]and [17]). As they point out, it is not enough to look atany of the single nodes in the network. The case of BearStearns is an example of a default which occurred eventhough there was a capital cushion well above what is re-quired to meet supervisory standards calculated using theBasel II standard (Cox, 2008).

To come back to the basic message of FuturICT, reg-ulators have to face the fact that, in many situations, ac-tions ensuring the soundness of one institution (e.g., sol-vency, liquidity capacity, etc.) may not be consistent withensuring the soundness of another (Crockett, 2000). Onthe contrary they may even decrease the stability of thesystem as a whole. Indeed, local shocks can have systemicrepercussions and the requirement to have sounder indi-viduals can have the counter-intuitive effect to make theentire system more fragile. It is interesting to note theseissues are very common in the field of complex systems,where the network of agent interactions at the micro levelresults in the emergence of new behaviour that must betackled with new global strategies. To quote FOC, “Suchan integrated and network-oriented approach invoked byregulators does not exist at the moment. This projectaims at developing new indicators that are genuinely con-structed with a systemic risk approach starting from mi-croscopic data and, in particular, taking into account thenetwork of mutual exposures among institutions”.

Some progress has been made towards obtaining ana-lytical results for the fragility of a network. Indeed undersome simplifying assumptions, exact analytical solutionshave been obtained for the critical fraction of nodes that,when removed, will lead to a failure cascade and to a com-plete fragmentation into interdependent networks. Theynotice that a broader degree distribution increases the vul-nerability of interdependent networks to random failure,which is opposite to how a single network behaves. Weplan to use the concept of networks of networks in the in-vestigation of the interbank market and of the credit net-work. In fact there is a network interaction among banksand a network interaction among firms and the credit net-work among banks and firms links the two networks.

Recent papers have shown that connectivity might fa-vor disease spreading, with domino effects and bankruptcycascades (reference in sect. 3.3). FuturICT may envisagehow to set some circuit breakers which are activated when,e.g., the average leverage of the actors in a network reachessome critical threshold (two FP7 European projects , FOCand CRISIS, deal with that issue). In a sense, the prob-lem is that of finding the best strategy to immunize apopulation of agents with a minimal welfare loss. It hasbeen accepted that targeted strategies on most centralnodes are most efficient for networks. FuturICT will de-velop a graph-partitioning strategy, which requires a mini-mal wealth injection into the system, by isolating its mostfragile parts.

Farmer et al.: Complex systems for better management of the economy 17

5 Expected Impact

The current economic crisis has reminded us just how im-portant the economy is, and made us acutely aware of ourcurrent lack of understanding and our inability to manageit. The ability to avert such crises, or to at least lessentheir impact and manage them better when they occur,will be enormously valuable. On other levels this projectwill also have a major impact on science itself, as well ason technology and competitiveness.

5.1 Impact on Science

As already mentioned, the impact of this project on thefield of economics would be enormous. Central bankersare actively seeking tools that would give them a bet-ter understanding of the nature of the crises with whichthey have to deal and the effectiveness of what they aredoing in the short and long run. Many economics depart-ments throughout the world have focused considerable re-sources on providing better tools for central bankers. Reg-ulatory policy analysis is one of the most important prac-tical application of economics, and even small improve-ments can have big effects. Complex systems analysis andagent-based modelling are so different from the standardapproach that a practical success in this domain couldtrigger a sea-change in the way economics is done.

At present economics stands out as the branch of sci-ence where the study of complex systems has had the leastimpact. Yet, it is clear that periodic endogenous crises arean intrinsic feature of modern economies (see Reinhartand Rogoff’s “This time is different” [87]). It is at beston the periphery. A success here could bring economistsstrongly into the center of the complex systems commu-nity.

As should be evident just from the data sets thatwe envision collecting, the approach that we are takingis highly transdisciplinary. Demonstrating the practicalvalue of transdisciplinary approaches would also have amajor impact.

Lastly in this connection, we will make the data wecollect available (within legal constraints) to anyone whowishes to work on it. This should be a significant advan-tage for those who need more complete socio-economicdata and represents one of the major advantages of thisproject for the community at large.

5.2 Impact on Technology and Competitiveness

We envision using state of the art software and hardwareto achieve our goals. One of these goals is to develop stan-dard libraries of software tools for agent-based modelling.This could drive software capable of both rapid prototyp-ing and rapid execution.

As it currently stands the United States completelydominates the field of economics. The US has given al-most no support to applications of complex systems in

economics, in particular agent-based modelling. A successin this area could turn Europe into the leader of the field.

Science depends on funding, and science funding de-pends on the economy. As argued below, even a minorsuccess in improving our ability to manage the economyis worth an enormous amount, which ultimately impactsscience research budgets in every field.

5.3 Impact on Society

The current crisis has already cost the world an enormousamount, and may yet cost a lot more. Trillions of euroswere lost, unemployment soared, and the European Unioncame under severe stress. The stakes are very high – if thisproject could mitigate even one percent of the losses, thistranslates into tens of billions of euros, which would be anample return on the investment in a flagship project.

We sometimes forget that almost everything dependson the economy. Economic performance translates directlyinto human suffering. Poor economic performance meanshigh unemployment, and it also means less money for so-cial services, and lower budgets for the arts and sciences.It is no coincidence that Da Vinci and Michelangelo’s co-incided with the domination of Florence by the Medici(who directly supported their work).

Finally, a FuturICT Flagship would be a way to as-sure citizens that the stability and improvement of socio-economic system is one of the main worries of govern-ments, worthy of their attention and one of the main sci-entific challenges to be addressed in the first big scienceproject in the socio-economic sciences.

Acknowledgements

Doyne Farmer, Mauro Gallegati and Cars Hommes ac-knowledge financial support from the EU 7th frameworkcollaborative project ”Complexity Research Initiative forSystemic InstabilitieS (CRISIS)”, grant no. 288501. CarsHommes acknowledges financial support from the Nether-lands Organization for Scientific Research (NWO), project”Understanding Financial Instability through ComplexSystems”. None of the above are responsible for errorsin this paper.

References

1. H.N. Agiza, G.I. Bischi, and M. Kopel. Multistability ina dynamic cournot game with three oligopolists. Mathe-matics and Computers in Simulation, 51:63–90, 1999.

2. G.A. Akerlof. The market for ”lemons”: Quality uncer-tainty and the market mechanism. Quarterly Journal ofEconomics, 84:488–500, 1970.

3. G.A. Akerlof and R.J. Shiller. Animal Spirits: How Hu-man Psychology Drives the Economy, and Why It Mat-ters for Global Capitalism. Princeton University Press,Princeton, 2009.

18 Farmer et al.: Complex systems for better management of the economy

4. K. Anand, A. Kirman, and M. Marsili. Epidemics of rules,rational negligence and market crashes. European Journalof Finance, 0:1–10, 2011.

5. P. W. Anderson, K. J. Arrow, and Pines D., editors.The Economy as an Evolving Complex System. Addison-Wesley, Redwood City, 1988.

6. K. Arrow. Interview with ken arrow. In D. Colander,R.P.F. Holt, and Barkley J. Rosser., editors, The Chang-ing Face of Economics. Conversations with Cutting EdgeEconomists, pages 291–308. The University of MichiganPress, Ann Arbor, 2004.

7. K. J. Arrow and G. Debreu. Existence of an equilibriumfor a competitive economy. Econometrica, 22(3):265–290,1954.

8. W. B. Arthur. Inductive reasoning and bounded ratio-nality. The American Economic Review, 84(2):406–411,1994.

9. W. B. Arthur, J. H. Holland, B. LeBaron, R. Palmer,and P. Tayler. Asset pricing under endogenous expec-tations in an artificial stock market. In W. B. Arthur,S. N. Durlauf, and D. H. Lane, editors, The Economy asan Evolving Complex System II, pages 15–44. Addison-Wesley, Redwood City, 1997.

10. W. Brian Arthur, N. Durlauf, Steven, and A. Lane,David, editors. The Economy as an Evolving Complexsystem II. SFI Studies in the Complexity of Science.Perseus Books Group, 1997.

11. W.B. Arthur. Increasing Returns and Path Dependence inthe Economy. University of Michigan Press, Ann Arbor,1994.

12. P. Auerswald, S. Kauffman, J. Lobo, and K. Shell. Theproduction recipes approach to modeling technological in-novation: An application to learning by doing. Journal ofEconomic Dynamics and Control, 24:389–450, 2000.

13. M. Bacharach. Beyond Individual Choice: Teams andFrames in Game Theory, Edited by N.Gold and R. Sug-den. Princeton University Press, Princeton, 2006.

14. N. Barberis and R.H. Thaler. A survey of behavioral fi-nance. In G.M. Constantinides, M. Harris, and R. Stultz,editors, Handbook of the Economics of Finance. North-Holland (Handbooks in Economics Series), Amsterdam,2003.

15. S. Battiston, D. Delli Gatti, M. Gallegati, B. C. N. Green-wald, and J. E. Stiglitz. Credit chains and bankruptciesavalanches in production networks. Journal of EconomicDynamics and Control, 31:2061–2084, 2007.

16. S. Battiston, D. Delli-Gatti, M. Gallegati, B.C. Green-wald, and J.E. Stiglitz. Credit default cascades: Whendoes risk diversification increase stability? Journal of Fi-nancial Stability, forthcoming, 2012.

17. S. Battiston, D. Delli-Gatti, M. Gallegati, B.C. Green-wald, and J.E. Stiglitz. Liaisons dangereuses: Increasingconnectivity, risk sharing, and systemic risk. Journal ofEconomic Dynamics and Control, forthcoming, 2012.

18. R.A. Bentley, M. O’Brien, and P. Ormerod. Quality ver-sus mere popularity: a conceptual map for understandinghuman behaviour. Mind and Society, in press, 2011.

19. R.A. Bentley, P. Ormerod, and M. Batty. Evolving socialinfluence in large populations. Behavioural Ecology andSociobiology, 65:537–546, 2011.

20. O.J. Blanchard. The state of macro. Department of Eco-nomics Working Paper 08-17, MIT, 2008.

21. C.J. Bliss. Capital Theory and the Distribution of Income.North-Holland Publishing Company, Amsterdam, 1975,1975.

22. M. Boss, H. Elsinger, M. Summer, and S. Thurner. Net-work topology of the interbank market. Quantitative Fi-nance, 4:677–684, 2004.

23. M. Boss, M. Summer, and S. Thurner. Contagion flowthrough banking networks. Technical report, 2004.

24. R. Boyd, H. Gintis, and S. Bowles. Coordinated punish-ment of defectors sustains cooperation and can proliferatewhen rare. Science, 328:617–620, 2010.

25. W. Brock, C. Hommes, and F. Wagener. More hedging in-struments may destabilize markets. Journal of EconomicDynamics and Control, 33:1912–1928, 2009.

26. W. A. Brock and S. Durlauf. Discrete choice with socialinteractions. Review of Economic Studies, 68:235–260,2001.

27. W.A. Brock and C.H. Hommes. A rational route to ran-domness. Econometrica, 65:1059–1095, 1997.

28. William A. Brock and Cars H. Hommes. Heterogeneousbeliefs and routes to chaos in a simple asset pricing model.Journal of Economic Dynamics and Control, 22:1235–1274, 1998.

29. S.V. Buldyrev, R. Parshani, G. Paul, H.E. Stanley, andS. Havlin. Catastrophic cascade of failures in interdepen-dent networks. Nature, 464, 2010.

30. C.F. Camerer. Behavioral Game Theory: Experiments inStrategic Interaction. Princeton University Press, Prince-ton, 2003.

31. V.M. Carvalho. Aggregate fluctuations and the net-work structure of intersectoral trade. Technical report,CREI/UPF, 2010.

32. S. Cincotti, M. Raberto, and A. Teglio. Credit money andmacroeconomic instability in the agent-based model andsimulator eurace. Economics: The Open-Access, Open-Assessment E-Journal, 4(2010).

33. S. Cincotti, D. Sornette, P. Trelaeven, S. Battiston, Cal-darelli, and C.H.. Hommes. A european economic andfinancial exploratory. European Physical Journal B, thisissue, 2012.

34. M.P. Clements and D.F. Hendry. Economic forecastingin a changing world,.

35. D. Colander, H. Follmer, A. Haas, M.D. Goldberg,K. Juselius, A. Kirman, T. Lux, and B. Sloth. The finan-cial crisis and the systemic failure of academic economics.Economics Discussion Paper 09-03, Univ. of CopenhagenDept., 2009.

36. H. Dawid, S. Gemkow, P. Harting, S. van der Hoog,and M. Neugart. The euraceunibi model: an agent-based macroeconomic model for economic policy analysis.Working paper, University of Bielefeld, 2011.

37. G. Debreu. Excess demand functions. Journal of Mathe-matical Economics, 1:15–21, 1974.

38. D. Delli-Gatti, E. Gaffeo, M. Gallegati, G. Giulioni, andA. Pallestrini. Emergent macroeconomics. An agent-basedapproach to business fluctuations. Springer Verlag, Milan,2008.

39. George W. Evans and Seppo Honkapohja. Learning andExpectations in Macroeconomics. Princeton UniversityPress, 2001.

40. J. D. Farmer. Market force, ecology and evolution. In-dustrial and Corporate Change, 11(5):895–953, 2002.

Farmer et al.: Complex systems for better management of the economy 19

41. J. D. Farmer and D. Foley. The economy needs agent-based modelling. Nature, 460:685–686, 2009.

42. J.D. Farmer and J. Geanakoplos. The virtues and vices ofequilibrium and the future of financial economics. Com-plexity, 14:11–38, 2009.

43. H. Follmer. Random economies with many interactingagents. Journal of Mathematical Economics, 1:51–62,1974.

44. A. Fostel and J. Geanakoplos. Leverage cycles andthe anxious economy. American Economic Review,98(4):1211–1244, 2008.

45. P. Gai and S. Kapadia. Contagion in financial networks.Technical Report 383, Bank of England, 2010.

46. J. Geanakoplos. The leverage cycle. Technical Report1715, Cowles Foundation Discussion Paper, 2009.

47. J. Geanakoplos, R. Axtell, D. Farmer, P. Howitt, B. Con-lee, J. Goldstein, M. Hendrey, N. Palmer, and C-Y. Yang.Getting at systemic risk via an agent-based model of thehousing market. American Economic Review, forthcom-ing, 2012.

48. C.P. Georg. Basel iii and systemic risk regulation - whatway forward? Working Paper Series 2, 17-2011, GlobalFinancial Markets, 2011.

49. C.P. Georg. The effect of interbank network structure oncontagion and common shocks. Discussion Paper Series2, 12/2011, Deutsche Bundesbank, 2011.

50. J. Grujic, C. Fosco, L. Araujo, J.A. Cuesta, andA. Sanchez. Social experiments in the mesoscale: Hu-mans playing a spatial prisoner’s dilemma. PloS ONE,5(11):e13749, 2010.

51. M.W. Hahn and R.A. Bentley. Drift as a mechanism forcultural change: an example from baby names. Proc. Roy.Soc. B, 270:S1–S4, 2003.

52. A. Haldane. Rethinking the financial network, 2009.53. A.G. Haldane and R.M. May. Systemic risk in banking

ecosystems. Nature, 469(20 January):351–355, 2011.54. P. Heemeijer, C. H. Hommes, J. Sonnemans, and J. Tuin-

stra. Price stability and volatility in markets with positiveand negative expectations feedbacks: An experimental in-vestigation. Journal of Economic Dynamics and Control,33:1052–1072, 2009.

55. D. Helbing and S. Balietti. Fundamental and real-worldchallenges in economics. Science and Culture, 76:399–417,2010.

56. J. Henrich. The evolution of innovation-enhancing insti-tutions. In M.J. O’Brien and S.J. Shennan, editors, Inno-vation in cultural systems: contributions from evolution-ary anthropology, pages 99–120. MIT Press, Cambridge,2010.

57. C.A. Hidalgo and R. Hausmann. The building blocksof economic complexity. Proceedings of the NationalAcademy of Science, 106(26):10570–10575, 2009.

58. C.A. Hidalgo, B. Klinger, A.L. Barabasi, and R. Haus-mann. The product space conditions the development ofnations. Science, 27(2837):482–487, 2007.

59. C. H. Hommes. Heterogeneous agent models in eco-nomics and finance. In K.L. Judd and L. Tesfatsion,editors, Handbook of Computational Economics Vol. 2:Agent-Based Computational Economics. Elsevier/North-Holland (Handbooks in Economics Series), 2006.

60. C.H. Hommes. The heterogeneous expectations hypoth-esis: Some evidence from the lab. Journal of EconomicDynamics and Control, 35:1–24, 2011.

61. G. Iori, G. De Masi, O. Precup, G. Gabbi, and G. Cal-darelli. A network analysis of the italian overnight moneymarket. Journal of Economic Dynamics and Control,32:259–278, 2008.

62. G. Iori and S. Jafarey. Criticality in a model of bankingcrises. Physica A, 299:205–212, 2001.

63. G. Iori, S. Jafarey, and F. Padilla. Systemic risk on theinterbank market. Journal of Economic Behavior andOrganization, 61(4):525–542, 2006.

64. D. Kahneman and A. Tversky. Prospect theory: An anal-ysis of decision under risk. Econometrica, 47(2):263–292,1979.

65. A. Kirman. Epidemics of opinion and speculative bubblesin financial markets. In M. Taylor, editor, Money andFinancial Markets, chapter 17, pages 354–368. Blackwell,Cambridge, 1991.

66. A.P. Kirman. The intrinsic limits of modern economictheory: The emperor has no clothes. The Economic Jour-nal, 99:129–139, 1989.

67. A.P. Kirman. Ants, rationality and recruitment. Quar-terly Journal of Economics, 108:137–156, 1993.

68. B. LeBaron. Calibrating an agent-based financial market.Technical report, Brandeis Economics Dept., 2003.

69. B. LeBaron. Agent-based computational finance. In K.L.Judd and L. Tesfatsion, editors, Handbook of Compu-tational Economics Vol. 2: Agent-Based ComputationalEconomics. Elsevier/North-Holland (Handbooks in Eco-nomics Series), 2006.

70. B. LeBaron, W. B. Arthur, and R. Palmer. Time seriesproperties of an artificial stock market. Journal of Eco-nomic Dynamics and Control, 23(9-10):1487–1516, 1999.

71. E.A. Leicht and R.M. D’Souza. Percolation on interactingnetworks. Technical report, http://arxiv.org.

72. H.W. Lorenz. Nonlinear dynamical equations and chaoticeconomy. Springer, Berlin, 1993.

73. J. Lorenz, S. Battiston, and F. Schweitzer. Systemicrisk in a unifying framework for cascading processes onnetworks. European Physical Journal B, 71(4):441–446,2009.

74. R.E. Lucas. Econometric testing of the natural rate hy-pothesis. In O. Eckstein, editor, The econometrics of pricedetermination Conference, pages 50–59. Board of Gover-nors of the Federal Reserve System and Social ScienceResearch Council, Washington DC, 1972.

75. Robert E. Lucas and Edward C. Prescott. Investmentunder uncertainty. Econometrica, 39(5):659–681, 1971.

76. T. Lux. Stochastic behavioral asset pricing models andthe stylized facts. In T. Hens and K.R. Schenk-Hoppe,editors, Handbook of Financial Markets: Dynamics andEvolution. Elsevier, 2009.

77. Thomas Lux and Michele Marchesi. Scaling and critical-ity in a stochastic multi-agent model of financial market.Nature, 397:498–500, 1999.

78. Matteo Marsili. Spiraling toward market completenessand financial instability. Technical report, The AbdusSalam International Centre for Theoretical Physics, 2009.

79. J. McNerney, J.D. Farmer, S. Rdener, and J.E. Trancik.Role of design complexity in technology improvement.PNAS, 108(22):9008–9013, 2011.

80. J.V. Neumann, von. A model of general economic equi-librium. Review of Economic Studies, 13:1–9, 1945.

81. P. Ormerod. Risk, recessions and the resilience of thecapitalist economies. Risk Management, 12:83–89, 2010.

20 Farmer et al.: Complex systems for better management of the economy

82. P. Ormerod and E. Evans. Ex ante prediction of cas-cade sizes on networks of agents facing binary outcomes.Technical Report arXiv 1103.3457, 2011, 2011.

83. M. Raberto and S. Cincotti. Modeling and simulation ofa double auction artificial financial market. Physica A,355:34–45, 2005.

84. M. Raberto, S. Cincotti, S.M. Focardi, and M. Marchesi.Agent-based simulation of a financial market. Physica A,299:319–327, 2001.

85. M. Raberto, A. Teglio, and S. Cincotti. Debt deleveragingand business cycles. an agent-based perspective. Techni-cal Report 2011-31, Economics Discussion Paper, 2011.

86. R. Radner. Competitive equilibrium under uncertainty.Econometrica, 36:31–58, 1968.

87. C.M. Reinhart and K.S. Rogoff. This Time is Different: APanoramic View of Eight Centuries of Financial Crises.Princeton University Press, 2011.

88. L. Rendell, R. Boyd, D. Cownden, M. Enquist, K. Eriks-son, M.W. Feldman, L. Fogarty, S. Ghirlanda, T. Lil-licrap, and K.N. Laland. Why copy others? insightsfrom the social learning strategies tournament. Science,328:208–213, 2010.

89. P.J. Richerson, R. Boyd, and J. Henrich. Gene-culturecoevolution in the age of genomics. Proc. Nat. Acad. Sci.,107:8985–8992, 2010.

90. Thomas J. Sargent. Bounded Rationality in Macroeco-nomics. Oxford University Press, 1993.

91. H. Simon. A behavioral model of rational choice. TheQuarterly Journal of Economics, 69(1):99–118, 1955.

92. Herbert A. Simon. Models of Man: Social and Rational.John Wiley, New York, 1957.

93. H. Sonnenschein. Market excess demand functions.Econometrica, 40:549–63, 1972.

94. J. Steindl. Random Processes and the Growth of Firms., Griffin, London, 1965.

95. J. Stiglitz and S. Salop. Bargains and ripoffs: A model ofmonopolistically competitive price dispersion. Review ofEconomic Studies, 44:493–510, 1977.

96. S. Thurner, J. D. Farmer, and J. Geanakoplos. Leveragecauses fat tails and clustered volatility, 2009.

97. A. Traulsen, D. Semmann, R.D. Sommerfeld, H-J. Kram-beck, and M. Milinski. Human strategy updating in evo-lutionary games. PNAS, 107(7):2962–2966, 2010.

98. A.M. Turing. The chemical basis of morphogenesis. Philo-sophical transactions of the Royal Society, B, 237:37–72,1952.

99. A. Tversky and D. Kahneman. Judgement under un-certainty: Heuristics and biases. Science, 185:1124–1130,1974.

100. S. Vitali and S. Battiston. Geography versus topology inthe european ownership network. New Journal of Physics,13(6):063021, 2011.

101. S. Vitali, J.B. Glattfelder, and S. Battiston. The networkof global corporate control. PLos ONE, 6(10):e25995,2011.

102. D. Watts. A simple model of global cascades on randomnetworks. PNAS, 99:5766, 2002.

103. R. Wilkinson and K. Pickett. The Spirit Level: WhyEquality is Better for Everyone. , Penguin, 2009.

104. M. Woodford. Convergence in macroeconomics: ele-ments of the new synthesis. American Economic Journal:Macroeconomics, 1, 2009.