Artículos de Nature sobre modelos económicos basados en agentes

8/3/2019 Artculos de Nature sobre modelos econmicos basados en agentes

1/8

Compared with physics, it seems fair to saythat the quantitative success of the economicsciences has been disappointing. Rockets flyto the Moon; energy is extracted from minutechanges of atomic mass. What is the flagshipachievement of economics? Only its recurrentinability to predict and avert crises, includingthe current worldwide credit crunch.

Why is this so? Of course, to paraphrase Isaac

Newton, modelling the madness of people ismore difficult than modelling the motionof planets. But statistical regularities shouldemerge in the behaviour of large populations,just as the law of ideal gases emerges from thechaotic motion of individual molecules. Tome, the crucial difference between modellingin physics and in economics lies rather in howthe fields treat the relative role of concepts,equations and empirical data.

Classical economics is built on very strongassumptions that quickly become axioms: therationality of economic agents (the premisethat every economic agent, be that a person or

a company, acts to maximize his profits), theinvisible hand (that agents, in the pursuit oftheir own profit, are led to do what is best forsociety as a whole) and market efficiency (thatmarket prices faithfully reflect all known infor-mation about assets), for example. An econo-mist once told me, to my bewilderment: Theseconcepts are so strong thatthey supersede any empiri-cal observation. As econo-mist Robert Nelson arguedin his book, Economics asReligion (Pennsylvania State Univ. Press, 2002),the marketplace has been deified.

Physicists, on the other hand, have learnedto be suspicious of axioms. If empirical obser-vation is incompatible with a model, themodel must be trashed or amended, even ifit is conceptually beautiful or mathematicallyconvenient. So many accepted ideas have beenproven wrong in the history of physics thatphysicists have grown to be critical and queasyabout their own models.

Unfortunately, such healthy scientificrevolutions have not yet taken hold in econom-ics, where ideas have solidified into dogmas.These are perpetuated through the educationsystem: students dont question formulas they

can use without thinking. Although numer-ous physicists have been recruited by financial

institutions over the past fewdecades, they seem to haveforgotten the methodologyof the natural sciences as theyabsorbed and regurgitated theexisting economic lore.

The supposed omnis-cience and perfectefficacy of a free mar-

ket stems from eco-nomic work done inthe 1950s and 1960s,which with hind-sight looks more likepropaganda againstcommunism thanplausible science. Inreality, markets arenot efficient, humanstend to be over-focusedin the short-term and blindin the long-term, and errors get amplified, ulti-mately leading to collective irrationality, panic

and crashes. Free markets are wild markets.

Picture imperfectReliance on models based on incorrect axiomshas clear and large effects. The BlackScholesmodel, for example, which was invented in1973 to price options, is still used extensively.

But it assumes that theprobability of extremeprice changes is negligi-ble, when in reality, stockprices are much jerkier

than this. Twenty years ago, unwarranted use ofthe model spiralled into the worldwide October

1987 crash; the Dow Jones index dropped 23%in a single day, dwarfing recent market hiccups.Ironically, it was the very use of a crash-freemodel that helped to trigger a crash.

This time, the problem lies, in part, in thedevelopment of structured financial productsthat packaged subprime risk into seeminglyrespectable high-yield investments. The mod-els used to price them were fundamentallyflawed: they underestimated the probabilitythat multiple borrowers would default ontheir loans simultaneously. These models againneglected the very possibility of a global crisis,even as they contributed to triggering one.

Surprisingly, classical economics has noframework through which to understand

wild markets, even thoughtheir existence is so obvious to

the layman. Physics, on the otherhand, has developed several

models that explain howsmall perturbations canlead to wild effects. Thetheory of complexityshows that although

a system may have anoptimum state, it issometimes so hardto identify that the

system never settlesthere. This opti-mum state is notonly elusive, it isalso hyper-fragile

to small changesin the environment,

and therefore oftenirrelevant to understanding what is going on.There are good reasons to believe that this

paradigm should apply to economic systems ingeneral and financial markets in particular. Weneed to break away from classical economicsand develop completely different tools. Somebehavioural economists and econo-physicistsare attempting to do this now, in a patchy way,but their fringe endeavour is not taken seri-ously by mainstream economics.

While work is done to enhance models,regulation also needs to improve. Innovationsin financial products should be scrutinized,crash-tested against extreme scenarios outsidethe realm of current models and approved byindependent agencies, just as we have done with

other potentially lethal industries (chemical,pharmaceutical, aerospace, nuclear energy).Crucially, the mindset of those working in

economics and financial engineering needs tochange. Economics curricula need to includemore natural science. The prerequisites formore stability in the long run are the develop-ment of a more pragmatic and realistic rep-resentation of what is going on in financialmarkets, and to focus on data, which shouldalways supersede perfect equations and aes-thetic axioms. Jean-Philippe Bouchaud is head of researchof Capital Fund Management and a physicsprofessor at cole Polytechnique in France.e-mail: [email protected]

Economics needs a scientific revolutionFinancial engineers have put too much faith in untested axioms and faulty models, says Jean-Philippe

Bouchaud. To prevent economic havoc, that needs to change.

Classical economics has no

framework through which to

understand wild markets.

D

P A R K I N S

1181

NATURE|Vol 455|30 October 2008 OPINION

ESSAY


2/8

A model approachMore development work is needed to help computer simulations inform economic policy.

Models are everywhere in economics. They range from thepencil-and-paper equations used for academic analyses ofmarket behaviour, to the computer forecasts used by central

banks, such as the Bank of England and the US Federal Reserve Sys-tem, to determine the likely effects of interest-rate adjustments.

But the reputation of economic models has been tarnished of late.Virtually none anticipated the global financial meltdown that begantwo years ago this summer (see pages 680 and 685). The finger-pointing seems likely to go on indefinitely: were the models flawed?

Or were policy-makers at fault for ignoring the warnings?What is clear is that economic models need to improve. The abil-

ity to run policy options through a believable set of what-if sce-narios could be useful to forestall future economic crises, and toinform debate, such as that over the labyrinthine efforts to reformthe US health-care system.

The field could benefit from lessons learned in the large-scalemodelling of other complex phenomena, such as climate change andepidemics (see page 687). Those lessons, taken together with lessonsfrom the downturn, suggest an ambitious research agenda not justfor economists, but for psychologists, political and social scientists,computer researchers and more.

First, details matter. Government regulators rely on dynamic

stochastic general equilibrium (DSGE) simulations, which can makesophisticated extrapolations of past economic data. But these modelsdo little to incorporate information about the financial sector, whichis where the current crisis began. Which company was entering intowhat kind of arrangements with another, for example, and how werethey all interconnected? And most models dont even attempt toincorporate the psychological insights gained from behavioural

economics, and so ignore shifting attitudes towards risk, and thespread of fear both major contributors to the crisis. The com-paratively few modelling efforts that do try to include these factorsdeserve support and many more such efforts are needed.

Second, models should evolve through vigorous competition. Asthe articles in this issue show, advocates of agent-based modellingtechniques, which represent each individual or company with anagent, claim that their programs can often account for economicphenomena much better than can DSGE simulations. Such claims

need to be addressed empirically. The economics community shouldtry to agree on a standard set of test cases analogous to those usedby climate modellers, whose challenges can include being able toreproduce El Nio oscillations. Economic modellers should alsoconsider adopting the modular architecture used in many climatemodels. This approach makes it easy to aggregate smaller modelsinto more comprehensive simulations, while still allowing steadyimprovement in each piece. A sub-model for ocean circulation,say, can be switched for an alternative circulation module withoutchanging anything else.

Third, modellers seeking to make a real difference in the worldshould concentrate on the tangible, immediate questions thatdecision-makers actually worry about. A good example to follow

is that of pandemic planning, in which simulations are already inwidespread use to help officials decide when to close schools andother public gathering places, and how best to mount a vaccinationcampaign. The simulations alone cannot answer such questions,nor can they replace judgement. But by helping officials frame theproblem, organize the available information and identify whichfactors matter, they can make judgements better informed.

Science under attackCongress should stop playing politics with the

peer-review process.

In a depressingly familiar display of irresponsible politicking, theUS House of Representatives has taken aim at three studies fundedby the National Institutes of Health (NIH). Representative Darrell

Issa (Republican, California) introduced an amendment killing theprojects on 24 July, during a debate on the NIHs 2010 budget. TheHouse passed the amendment by a voice vote.

Issa was unhappy that the studies looked at substance abuse and HIVrisk behaviour, and that the subjects were outside the United States.One focused on Russian alcoholics, another on female sex workers inChina and a third on female and transgender prostitutes in Thailand.All three passed muster with NIH peer reviewers, and together would

cost about $5 million over five years. Issa wanted that money to be

spent at home, and complained that HIV had been heavily studiedalready. But his reasoning is specious: alcoholism, prostitution andHIV do not respect borders, and any behavioural information thatcould help slow the transmission of HIV is crucial. Some 33 million

people are infected worldwide, and a vaccine is nowhere in sight.Issas tactic is not new. Since 2003, conservative House Republicans

have tried at least five times to strip funding from peer-reviewedprojects that drew their ire. Such meddling threatens to underminethe peer-review process as well as potentially eroding the publicstrust that science is above politics.

Also worrying is the House Democrats acquiescence to Issasamendment. Democrats facing tough re-election bids hoped tododge Republican attacks in media adverts in their home districtsthat might have resulted from opposing Issa. Their assumption isthat the amendment can be quietly removed when House and Senatenegotiators meet to square their versions of the NIH bill before a final

vote on it. But Congress should renounce all tactics that undermine

peer review and cease indulging those who use them.

667

www.nature.com/nature Vol 460 | Issue no. 7256 | 6 August 2009

2009 Macmillan Publishers Limited. All rights reserved


3/8

The economy needs agent-based modellingThe leaders of the world are flying the economy by the seat of their pants, say J. Doyne Farmer and

Duncan Foley. There is, however, a better way to help guide financial policies.

In todays high-tech age, one naturallyassumes that US President BarackObamas economic team and its inter-

national counterparts are using sophis-ticated quantitative computer modelsto guide us out of the current economiccrisis. They are not.

The best models they have are of twotypes, both with fatal f laws. Type one iseconometric: empirical statistical models

that are fitted to past data. These suc-cessfully forecast a few quarters aheadas long as things stay more or less thesame, but fail in the face of great change.Type two goes by the name of dynamicstochastic general equilibrium. Thesemodels assume a perfect world, and bytheir very nature rule out crises of thetype we are experiencing now.

As a result, economic policy-makersare basing their decisions on commonsense, and on anecdotal analogies toprevious crises such as Japans lostdecade or the Great Depression (see

Nature457, 957; 2009). The leaders ofthe world are flying the economy by theseat of their pants.

This is hard for most non-economists tobelieve. Arent people on Wall Street usingfancy mathematical models? Yes, but for acompletely different purpose: modelling thepotential profit and risk of individual trades.There is no attempt to assemble the piecesand understand the behaviour of the wholeeconomic system.

There is a better way: agent-based models.An agent-based model is a computerized simu-lation of a number of decision-makers (agents)

and institutions, which interactthrough prescribed rules. The agentscan be as diverse as needed fromconsumers to policy-makers and WallStreet professionals and the institu-tional structure can include everythingfrom banks to the government. Suchmodels do not rely on the assumptionthat the economy will move towardsa predetermined equilibrium state, as othermodels do. Instead, at any given time, eachagent acts according to its current situation, thestate of the world around it and the rules gov-erning its behaviour. An individual consumer,

for example, might decide whether to save orspend based on the rate of inflation, his or her

current optimism about the future, and behav-ioural rules deduced from psychology experi-ments. The computer keeps track of the manyagent interactions, to see what happens overtime. Agent-based simulations can handle a farwider range of nonlinear behaviour than con-ventional equilibrium models. Policy-makerscan thus simulate an artificial economy underdifferent policy scenarios and quantitativelyexplore their consequences.

Why is this type of modelling not well-developed in economics? Because of his-

torical choices made to address thecomplexity of the economy and theimportance of human reasoning andadaptability.

The notion that financial econo-mies are complex systems can betraced at least as far back as AdamSmith in the late 1700s. More recentlyJohn Maynard Keynes and his fol-

lowers attempted to describe and quantifythis complexity based on historical patterns.Keynesian economics enjoyed a heyday in thedecades after the Second World War, but wasforced out of the mainstream after failing a cru-

cial test during the mid-seventies. The Keyne-sian predictions suggested that inflation could

pull society out of a recession; that, asrising prices had historically stimulatedsupply, producers would respond tothe rising prices seen under inflationby increasing production and hiringmore workers. But when US policy-makers increased the money supply inan attempt to stimulate employment, itdidnt work they ended up with bothhigh inflation and high unemployment,

a miserable state called stagflation.Robert Lucas and others argued in1976 that Keynesian models had failedbecause they neglected the power ofhuman learning and adaptation. Firmsand workers learned that inflation isjust inflation, and is not the same as areal rise in prices relative to wages.

Realistic behaviourThe cure for macroeconomic theory,however, may have been worse than thedisease. During the last quarter of thetwentieth century, rational expectations

emerged as the dominant paradigmin economics. This approach assumes

that humans have perfect access to informa-tion and adapt instantly and rationally to newsituations, maximizing their long-run personaladvantage. Of course real people often act onthe basis of overconfidence, fear and peer pres-sure topics that behavioural economics isnow addressing.

But there is a still larger problem. Even ifrational expectations are a reasonable model ofhuman behaviour, the mathematical machineryis cumbersome and requires drastic simplifica-tions to get tractable results. The equilibrium

models that were developed, such as those usedby the US Federal Reserve, by necessity strippedaway most of the structure of a real economy.There are no banks or derivatives, much lesssub-prime mortgages or credit default swaps these introduce too much nonlinearity andcomplexity for equilibrium methods to handle.When it comes to setting policy, the predictionsof these models arent even wrong, they are sim-ply non-existent (seeNature455, 1181; 2008).

Agent-based models potentially presenta way to model the financial economy as acomplex system, as Keynes attempted to do,while taking human adaptation and learning

into account, as Lucas advocated. Such mod-els allow for the creation of a kind of virtual

Agent-based models could help to evaluate policies designed to

foster economic recovery.

685

Vol 460|6 August 2009

OPINION



4/8

universe, in which many players canact in complex and realistic ways. In some other areas of science,

such as epidemiology or traffic con-trol, agent-based models already helppolicy-making.

Promising effortsThere are some successful agent-based models of small portions of theeconomy. The models of the finan-cial market built by Blake LeBaronof Brandeis University in Waltham,Massachusetts, for example, provide aplausible explanation for bubbles and crashes,reproducing liquidity crises and crashes thatnever appear in equilibrium models. Rob

Axtell of George Mason University in Fairfax,Virginia, has devised firm dynamics modelsthat simulate how companies grow and declineas workers move between them. These repli-cate the power-law distribution of companysize that one sees in real life: a very few largefirms, and a vast number of very small oneswith only one or two employees.

Other promising efforts include the credit-sector model of Mauro Gallegatis group at theMarche Polytechnic University in Ancona,Italy, and the monetary model developedby Robert Clower of the University of SouthCarolina in Columbia and Peter Howitt of

Brown University in Providence, RhodeIsland. These models are very useful, buttheir creators would be the first to say that theyprovide only a tentative first step.

To see in more detail how an agent-basedmodel works, consider the model that oneof us (Farmer) has developed with StefanThurner of the University of Vienna and JohnGeanakoplos of Yale University to explore howleverage affects fluctuations in stock prices(published in a Santa Fe Institute workingpaper). Leverage, the invest-ment of borrowed funds, ismeasured as the ratio of

total assets owned to thewealth of the borrower; if ahouse is bought with a 20%down-payment the lever-age is five. There are four types of agents inthis model. Noise traders, who trade more orless at random, but are slightly biased towarddriving prices towards a fundamental value;hedge funds, which hold a stock when it isunder-priced and otherwise hold cash; inves-tors who decide whether to invest in a hedgefund; and a bank that can lend money to thehedge funds, allowing them to buy morestock. Normally, the presence of the hedge

funds damps volatility, pushing the stockprice towards its fundamental value. But, to

contain their risk, the banks cap leverage at apredetermined maximum value. If the price ofthe stock drops while a fund is fully leveraged,

the funds wealth plummets and its leverageincreases; thus the fund has to sell stock to payoff part of its loan and keep within its leveragelimit, selling into a falling market.

This agent-based model shows how thebehaviour of the hedge funds amplifies pricefluctuations, and in extreme cases causescrashes. The price statistics from this model look

very much like reality. It shows that the standardways banks attempt to reduce their own risk cancreate more risk for the whole system.

Previous models of leverage based onequilibrium theory showed qualitatively howleverage can lead to crashes, but they gave no

quantitative information about how this affectsthe statistical properties of prices. The agentapproach simulates complex and nonlinearbehaviour that is so far intractable in equilib-rium models. It could be made more realisticby adding more detailed information aboutthe behaviour of real banks and funds, and thiscould shed light on many important questions.For example, does spreading risk across manyfinancial institutions stabilize the f inancialsystem, or does it increase financial fragility?

Better data on lendingbetween banks and hedgefunds would make it possi-

ble to model this accurately.What if the banks themselvesborrow money and use lever-age too, a process that played

a key role in the current crisis? The model couldbe used to see how these banks might behave inan alternative regulatory environment.

Agent-based models are not a panacea. Themajor challenge lies in specifying how theagents behave and, in particular, in choosingthe rules they use to make decisions. In manycases this is still done by common sense andguesswork, which is only sometimes sufficientto mimic real behaviour. An attempt to model

all the details of a realistic problem can rapidlylead to a complicated simulation where it is dif-

ficult to determine what causes what.To make agent-based modelling use-ful we must proceed systematically,

avoiding arbitrary assumptions,carefully grounding and testing eachpiece of the model against reality andintroducing additional complexityonly when it is needed. Done right,the agent-based method can providean unprecedented understanding ofthe emergent properties of interact-ing parts in complex circumstanceswhere intuition fails.

A thorough attempt to understandthe whole economy through agent-based mod-elling will require integrating models of financialinteractions with those of industrial production,

real estate, government spending, taxes, businessinvestment, foreign trade and investment, andwith consumer behaviour. The resulting simula-tion could be used to evaluate the effectivenessof different approaches to economic stimulus,such as tax reductions versus public spending.

Holistic approachSuch economic models should be able toprovide an alternative tool to give insightinto how government policies could affectthe broad characteristics of economic per-formance, by quantitatively exploring howthe economy is likely to react under different

scenarios. In principle it might even be possi-ble to create an agent-based economic modelcapable of making useful forecasts of the realeconomy, although this is ambitious.

Creating a carefully crafted agent-basedmodel of the whole economy is, like climatemodelling, a huge undertaking. It requiresclose feedback between simulation, testing, datacollection and the development of theory. Thisdemands serious computing power and multi-disciplinary collaboration among economists,computer scientists, psychologists, biologistsand physical scientists with experience in large-scale modelling. A few million dollars much

less than 0.001% of the US financial stimuluspackage against the recession would allow aserious start on such an effort.

Given the enormity of the stakes, such anapproach is well worth trying. J. Doyne Farmer is at the Santa Fe Institute, 1399

Hyde Park Road, Santa Fe, New Mexico 87501,

USA, and at LUISS Guido Carli in Rome, Italy, and

founded the quantitative trading firm Prediction

Company. Duncan Foley is Leo Model Professor of

Economics at the New School for Social Research,

6 East 16th Street, New York 10003, USA, and an

external professor at the Santa Fe Institute.

e-mails: [email protected]; [email protected]

See Opinion, page 687, and Editorial, page 667.

Further reading accompanies this article online.

S.P

LATT/GETTYIMAGES

The policy predictions

of the models that are in

use arent wrong, they

are simply non-existent.

686

NATURE|Vol 460|6 August 2009OPINION



5/8

I

ts 2016, and experts at a US govern-ment facility have detected a threatto national security. A screen on thewall maps the worlds largest finan-

cial players banks, governmentsand hedge funds as well as the webof loans, ownership stakes and otherlegal claims that link them. High-powered computers have been using theseenormous volumes of data to run throughscenarios that flush out unexpected risks. Andthis morning they have triggered an alarm.

Flashing orange alerts on the screen showthat a cluster of US-based hedge funds hasunknowingly taken large ownership positionsin similar assets. If one of the funds should haveto sell assets to raise cash, the computers warn,its action could drive down the assets value and

force others to start selling their own holdingsin a self-amplifying downward spiral. Many ofthe funds could be bankrupt within 30 min-utes, creating a threat to the entire financialsystem. Armed with this information, financialauthorities step in to orchestrate a controlledelimination of the dangerous tangle.

Alas, this story is likely to remain fiction. Nogovernment was able to carry out any such warroom analyses as the current financial crisisemerged, nor does the capability exist today.Yet a growing number of scientists insist thatsomething like it is needed if society is to avoidsimilar crises in future.

Financial regulators do not have the toolsthey need to predict and prevent meltdowns,

says physicist-turned-sociologist DirkHelbing of the Swiss Federal Instituteof Technology Zurich, who has spentthe past two decades modelling large-

scale human systems such as urbantraffic or pedestrian flows. They cando a good job of tracking an economyusing the statistical measures of stand-

ard econometrics, as long as the influences onthe economy are independent of each other,and the past remains a reliable guide to thefuture. But the recent financial collapse wasa systemic meltdown, in which intertwinedbreakdowns in housing, bank-ing and many other sectorsconspired to destabilize thesystem as a whole. And the pasthas been anything but a reli-

able guide of late: witness howUS analysts were led astray bydecades of data suggesting thathousing values would neversimultaneously fall across the nation.

Likewise, economists can get reasonablygood insights by assuming that human behav-iour leads to stable, self-regulating markets,with the prices of stocks, houses and otherthings never departing too far from equilib-rium. But stability is a word few would useto describe the chaotic markets of the past fewyears, when complex, nonlinear feedbacksfuelled the boom and bust of the dot-com and

housing bubbles, and when banks took extremerisks in pursuit of ever higher profits.

In an effort to deal with such messy realities,a few economists often working with physi-cists and others outside the economic main-stream have spent the past decade or so

exploring agent-based models that make onlyminimal assumptions about human behaviouror inherent market stability (see page 685). Theidea is to build a virtual market in a computerand populate it with artificially intelligent bitsof software agents that interact with oneanother much as people do in a real market.The computer then lets the overall behaviourof the market emerge from the actions of the

individual agents, without pre-supposing the result.

Agent-based models haveroots dating back to the 1940sand the first cellular autom-

ata, which were essentiallyjust simulated grids of onoffswitches that interacted withtheir nearest neighbours. But

they didnt spark much interest beyond thephysical-science community until the 1990s,when advances in computer power began tomake realistic social simulations more feasible.Since then they have found increasing use inproblems such as traffic flow and the spreadof infectious diseases (see page 687). Indeed,points out Helbing, agent-based models are thesocial-science analogue of the computationalsimulations now routinely used elsewhere in

science to explore complex nonlinear processessuch as the global climate.

Meltdown modellingCould agent-based computer models prevent another financial crisis? Mark Buchanan reports.

We have had a

massive failure

of the dominant

economic model. Eric Weinstein

ILLUSTRATIONSBY

JESSELEFKOWITZ

680


680

NATURE|Vol 460|6 August 2009NEWS FEATURE



6/8

That is why he is eager to bring social andphysical scientists together to develop computa-tional wind tunnels that would allow regulatorsto test policies before putting them into practice.The idea is to invest a lot in science, he says,and thereby save hundreds of times as much byavoiding or mitigating future crises.

Just more theory?That notion is a tough sell among mainstreameconomists, many of whom are less than

thrilled by offers of outside help. After anycrisis, says Paul Romer of Stanford University,California, a leading researcher in the econom-ics of innovation, you hear recommendationsto recruit scientists from other fields who canpurge economics and finance of ideology andfailed assumptions. But we should ask if thereis any evidence that more theory, developed bypeople who dont have domain experience, isthe key to scientific progress in this area.

Others think some fresh thinking is longoverdue. We have had a massive failure of thedominant economic model, says Eric Wein-stein, a physicist working in mathematical

finance for the Natron Group, a hedge fund inNew York, and were trying to find the rightpeople to deal with this failure. At least some ofthose people are likely to be unfamiliar voicesand come from other parts of science.

At least some economists agree. Themeltdown has shown that regulatory poli-cies have to cope with far-from-equilibriumsituations, says economist Blake LeBaron ofBrandeis University in Waltham, Massachu-setts. Even fairly simple agent-based modelscan be used as thought experiments to see ifthere is something that hasnt been consideredby the policy-makers.

LeBaron has spent the past decade and a halfworking with colleagues, including a number

of physicists, to develop an agent-based modelof the stock market. In this model, several hun-dred agents attempt to profit by buying and

selling stock, basing their decisions on pat-terns they perceive in past stock movements.Because the agents can learn from and respondto emerging market behaviour, they often shifttheir strategies, leading other agents to changetheir behaviour in turn. As a result, prices dontsettle down into a stable equilibrium, as stand-ard economic theory predicts. Much as in thereal stock market, the prices keep bouncing upand down erratically, driven by an ever-shiftingecology of strategies and behaviours.

Nor is the resemblance just qualitative, saysLeBaron. Detailed analyses of the agent-basedmodel show that it reproduces the statistical

features of real markets, especially their sus-ceptibility to sudden, large price movements.Traditional models do not go very far inexplaining these features, LeBaron says.

Another often-cited agent-based model gotits start in the late 1990s, as the NASDAQ stockexchange in New York was planning to stop list-ing its stock prices as fractions such as 12 andinstead list them as decimals.The goal was to improve theaccuracy of stock prices, but thechange would also allow pricesto move by smaller increments,which could affect the strate-

gies followed by brokers withunknown consequences forthe market as a whole. So beforemaking this risky change, NASDAQ chief MikeBrown hired BiosGroup, a company based inSanta Fe, New Mexico, to develop an agent-based model of the market to test the idea.

Over ten years on the NASDAQ Board,says Brown, I grew increasingly disappointedin our approach to studying the consequencesof proposed market regulations, and wanted totry something different.

Once the model could reproduce price fluc-tuations in a mathematically accurate way,

NASDAQ used it as a market wind tunnel.The tests revealed that if the stock exchangereduced its price increment too much, traderswould be able to exploit strategies that wouldmake them quick profits at the expense of over-all market efficiency. Thus, when the exchangewent ahead with the changeover in 2001, it wasable to take steps to counter this vulnerability.

Agent-based models are also being usedelsewhere in the private sector. For example,the consumer-products giant Proctor & Gam-ble of Cincinnati, Ohio, has used agent-basedmodels to optimize the flow of goods throughits network of suppliers, warehouses and stores.

And Southwest Airlines of Dallas, Texas, hasused agent-based models for routing cargo.

Despite such successes, however, financialregulators such as the US Securities andExchange Commission (SEC) still dont use

agent-based models as practical tools. Whenthe SEC changes trading rules, it typically haseither flimsy or modest support from econo-metric evidence for the action, or else noempirical evidence and the change is driven byideology, claims computational social scientistRob Axtell of George Mason University in Fair-fax, Virginia. You have to wonder why MikeBrown is doing this, while the SEC isnt.

Risk of the newA big part of the answer is that agent-basedmodels remain at the fringe of mainstreameconomics, and most economists continue

to prefer conventional mathematical models.Many of them argue that agent-based modelshavent had the same level of testing.

Another problem is that an agent-basedmodel of a market with many diverse playersand a rich structure may contain many variableparameters. So even if its output matches reality,its not always clear if this is because of care-

ful tuning of those parameters,or because the model succeedsin capturing realistic systemdynamics. That leads manyeconomists and social scientiststo wonder whether any such

model can be trusted. But agent-based enthusiasts counter thatconventional economic models

also contain many tunable parameters and aretherefore subject to the same criticism.

Familiarity wins out, notes Chester Spatt,former chief economist at the SEC. Regulatorsfeel duty-bound to adhere to generally acceptedand well-vetted techniques, he says. It wouldbe problematic for the rule-making process touse methods whose foundation or applicabilitywere not established.

Still, agent-based techniques are beginningto enter the regulatory process. For example,

decision-makers in Illinois and several otherUS states use computational models of com-plex electricity markets. They want to avoida repeat of the disaster in California in 2000,when Enron and other companies, followingmarket deregulation, were able to manipulateenergy supplies and prices for enormous profit.Rich computational models have made it pos-sible to test later market designs before puttingthem in place.

Weve had a lot of success in developingthese models, says economist Leigh Tesfat-sion of Iowa State University in Ames, who hasled the development of an open-source agent-

based model known as the AMES Whole-sale Power Market Test Bed. It has worked

We still implement

new economic

measures without

any prior testing.

Dirk Helbing

681


681

NATURE|Vol 460|6 August 2009 NEWS FEATURE



7/8

because weve focused on all the details of thereal situation and can address questions thatpolicy-makers really care about, she says.

Other models have successfully simu-lated financial markets. At Yale University,for example, economist John Geanakoplos,working with physicists Doyne Farmer of theSanta Fe Institute and Stefan Thurner of theMedical University of Vienna,has constructed an agent-basedmodel exploring the systemicconsequences of massive bor-rowing by hedge funds tofinance their investments. Intheir simulations, the fundsfrequently get locked into aself-amplifying spiral of losses

(see page 685) much as real-world hedge funds did after August 2007.

At the University of Genoa in Italy, mean-while, Silvano Cincotti and his colleagues arecreating an agent-based model of the entireEuropean Union economy. Their modelincludes markets for consumer goods andfinancial assets, firms that interact with banksto obtain loans, and banks that compete withone another by offering different interest rates.Based on real economic data, the model cur-rently represents some 10 million households,100,000 firms and about 100 banks, all of whichcan learn and change their strategies if they

find more profitable ways of doing business.We hope that these simulations will have

an outstanding impact on the economic-policycapabilities of the European Union, says Cin-cotti, and help design the best policies on anempirical basis.

This is the kind of ambition that has inspiredHelbing. He doesnt pretend to be an eco-nomic modeller himself: since the early 1990s

his own work has focused on simulations ofhuman behaviour in relatively small groups how traffic ebbs and flows through a roadnetwork, for example, or how crowds crushtowards a door in a panic situation as well

as on experiments to test hispredictions with real data. Butthat work has given Helbing akeen appreciation for the waycomplex collective phenomenacan emerge from even the sim-plest individual interactions. Ifpedestrians can organize them-selves into smoothly flowing

streams just by trying to walkthrough a crowded shopping centre as hehas shown they do just imagine how muchricher the emergent phenomena must be in agroup the size of a national economy.

Crisis logicThat observation acquired fresh force for Hel-bing after last years global financial meltdownmade it clear that a regulatory system based onconventional economic theory had failed.

Its remarkable, he says, that while anynew technical device or medical drug hasextensive testing for efficiency, reliability and

safety before it ever hits the market, we stillimplement new economic measures withoutany prior testing.

To get around this impasse, he says, research-ers need to reimagine the social and economicsciences on a larger scale. I imagine experts

from different fields meeting in one place forextended periods of time, he says, so thattheir complementary knowledge could collide,

creating new ideas, much as particle colliderscreate new kinds of particles. Ultimately, suchan effort would bring together social scientists,economists, physicists, ecologists, computerscientists and engineers in a network of largecentres for socioeconomic data mining andcrisis forecasting, as well as in supercomputercentres for social simulation and wind-tunnel-like testing of policy.

That is a large ambition, Helbing admits especially as he has only recently got tentativeapproval for a one-year grant from the Euro-pean Commission to develop the idea. Butnow, in the aftermath of the meltdown, may

be the time to start.Axtell endorses that view. Left to their own

devices, he says, academic macroeconomistswill take a generation to make this transition.But if policy-makers demand better models, itcan be accomplished much more quickly.

The revolution has to begin here, agreesWeinstein, who helped organize a meeting inMay at the Perimeter Institute for TheoreticalPhysics in Waterloo, Canada, that assembledthe kind of interdisciplinary mix of expertsthat Helbing envisions. And I think ideas fromphysics and other parts of science really have achance to catalyse something remarkable.

Mark Buchanan is a science writer based inCambridge, UK. After writing this story, he

was involved in reviewing grant proposals on

the topic of agent-based modelling.

See Editorial, page 667, and Opinion, pages 685

and 687.

Experts

complementary

knowledge could

collide, creating

new knowledge.

Dirk Helbing

682


682

NATURE|Vol 460|6 August 2009NEWS FEATURE



8/8

Modelling to contain pandemicsAgent-based computational models can capture irrational behaviour, complex social networks

and global scale all essential in confronting H1N1, says Joshua M. Epstein.

As the world braces for an autumn waveof swine flu (H1N1), the relatively newtechnique of agent-based computational

modelling is playing a central part in mappingthe diseases possible spread, and designingpolicies for its mitigation.

Classical epidemic modelling, which beganin the 1920s, was built on differential equa-tions. These models assume that the popula-tion is perfectly mixed, with people moving

from the susceptible pool,to the infected one, to therecovered (or dead) one.Within these pools, every-one is identical, and no oneadapts their behaviour. A tri-umph of parsimony, this approachrevealed the threshold nature of epidem-ics and explained herd immunity, wherethe immunity of a subpopulation can stifleoutbreaks, protecting the entire herd.

But such models are ill-suited to captur-ing complex social networks and the directcontacts between individuals, who adapt their

behaviours perhaps irrationally based ondisease prevalence.

Agent-based models (ABMs) embrace thiscomplexity. ABMs are artificial societies: everysingle person (or agent) is represented as a dis-tinct software individual. The computer modeltracks each agent, her contacts and her healthstatus as she moves about virtual space travel-ling to and from work, for instance. The modelscan be run thousands of times to build a robuststatistical portrait comparable to epidemic data.ABMs can record exact chains of transmissionfrom one individual to another. Perhaps mostimportantly, agents can be

made to behave something likereal people: prone to error, bias,fear and other foibles.

Such behaviours can have ahuge effect on disease progres-sion. What if significant num-bers of Americans refuse H1N1 vaccine out offear? Surveys and historical experience indicatethat this is entirely possible, as is substantialabsenteeism among health-care workers. Fearitself can be contagious. In 1994, hundreds ofthousands of people fled the Indian city of Suratto escape pneumonic plague, although by WorldHealth Organization criteria no cases were con-

firmed. The principal challenge for agent mod-elling is to represent such behavioural factors

appropriately; the capacity to do so is improv-ing through survey research, cognitive science,and quantitative historical study.

Robert Axtell and I published a full agent-based epidemic model1 in 1996. Agents withdiverse digital immune systems roamed a land-scape, spreading disease. The model trackeddynamic epidemic networks, simple mecha-nisms of immune learning, and behavioural

changes resulting from disease progression, allof which fed back to affect epidemic dynamics.However, the model was small (a few thousandagents) and behaviourally primitive.

Now, the cutting edge in performance is theGlobal-Scale Agent Model (GSAM)2, developedby Jon Parker at the Brookings InstitutionsCenter on Social and Economic Dynamics inWashington DC, which I direct. This includes6.5 billion distinct agents, with movementand day-to-day local interactions modelled asavailable data allow. The epidemic plays outon a planetary map, colour-coded for the dis-

ease state of people in different

regions black for suscepti-ble, red for infected, and bluefor dead or recovered. Themap pictured shows the state ofaffairs 4.5 months into a simu-lated pandemic beginning in

Tokyo, based on a plausible H1N1 variant.For the United States, the GSAM contains 300

million cyber-people and every hospital andstaffed bed in the country. The National Centerfor the Study of Preparedness and CatastrophicEvent Response at Johns Hopkins University inBaltimore is using the model to optimize emer-gency surge capacity in a pandemic, supported

by the Department of Homeland Security.Models, however, are not crystal balls

and the simulation shown here is not a pre-diction. It is a base case which by design ishighly unrealistic, ignoring pharmaceuticals,quarantines, school closures and behaviouraladaptations. It is nonetheless essential because,base case in hand, we can rerun the model toinvestigate the questions that health agenciesface. What is the best way to allocate limitedsupplies of vaccine or antiviral drugs? Howeffective are school or work closures?

Agent-based models helped toshape avian flu (H5N1) policy,through the efforts of the NationalInstitutes of Healths Models ofInfectious Disease Agent Study(MIDAS) a research network

to which the BrookingsInstitution belongs. TheGSAM was recentlypresented to officialsfrom the Centers for

Disease Control andPrevention in Atlanta,

Georgia, and other agencies, and will be inte-

gral to MIDAS consulting on H1N1 and otheremerging infectious diseases. In the wake ofthe 11 September terrorist attacks and anthraxattacks in 2001, ABMs played a similar part indesigning containment strategies for smallpox.

These policy exercises highlight anotherimportant feature of agent models. Becausethey are rule-based, user-friendly and highlyvisual, they are natural tools for participatorymodelling by teams clinicians, public-healthexperts and modellers. The GSAM executesan entire US run in around ten minutes, fastenough for epidemic war games, giving deci-sion-makers quick feedback on how interven-

tions may play out. This speed may even permitthe real-time streaming of surveillance data fordisease tracking, akin to hurricane tracking. AsH1N1 progresses, and new health challengesemerge, such agent-based modelling effortswill become increasingly important. Joshua M. Epstein is director of the Center on

Social and Economic Dynamics at the Brookings

Institution, 1775 Massachusetts Avenue,

Washington DC 20036, USA.

e-mail: [email protected]

1. Epstein, J. M. & Axtell, R. L. Growing Artificial Societies:Social Science from the Bottom Up Ch. V. (MIT Press, 1996).

2. Parker, J. A.ACM Trans Model. Comput. S. (in the press).

See Opinion, page 685, and Editorial, page 667.

Further reading accompanies this article online.

Agents can be madeto behave something

like real people: prone

to error, bias, fear.

Simulation of a pandemic beginning in Tokyo.

687

NATURE|Vol 460|6 August 2009 OPINION

Artículos de Nature sobre modelos económicos basados en agentes

Documents

Transcript of Artículos de Nature sobre modelos económicos basados en agentes