The Challenge of Agent-based Modeling in Economics

!ESRC Conference on Diversity in Macroeconomics!

University of Essex

J.  Doyne  Farmer  Ins$tute  for  New  Economic  Thinking  at  the  Oxford  Mar$n  School  

Mathema0cal  Ins0tute,  Oxford  University  External  professor,  Santa  Fe  Ins$tute

Feb.  24,  2014

The  talk  I  almost  gave:  “Three  diverse  approaches  to  modeling  systemic  risk”  • Cascading  failure  network  models  – interbank  lending  – overlapping  porLolios  – combina$on  of  the  two  

• Dynamical  systems  models  – Liquida$on-­‐based  accoun$ng  – leverage  and  overlapping  porLolios  revisited  

• Agent-­‐based  models  – leveraged  value  investors  – housing  model  – CRISIS  (model  of  “main  components”  of  economy

�2

The Virtues and Vices of Equilibrium,and the Future of Financial Economics

J. Doyne Farmer� and John Geanakoplos†

January 11, 2008

Abstract

The use of equilibrium models in economics springs from the de-sire for parsimonious models of economic phenomena that take humanreasoning into account. This approach has been the cornerstone ofmodern economic theory. We explain why this is so, extolling thevirtues of equilibrium theory, then present a critique and describe whythis approach is inherently limited, and why economics needs to movein new directions if it is to continue to make progress. We stress thatthis shouldn’t be a question of dogma, but should be resolved empir-ically. There are situations where equilibrium models provide usefulpredictions and there are situations where they can never provide use-ful predictions. There are also many situations where the jury is stillout, i.e., where so far they fail to provide a good description of theworld, but where proper extensions might change this. Our goal is toconvince the skeptics that equilibrium models can be useful, but also tomake traditional economists more aware of the limitations of equilib-rium models. We sketch some alternative approaches and discuss whythey should play an important role in future research in economics.

Contents

1 Introduction 3

2 What is an equilibrium theory? 4

3 E�ciency 8�McKinsey Professor, Santa Fe Institute, 1399 Hyde Park Rd., Santa Fe NM 87501†Economics Department, Yale University, New Haven CT

1

(appeared in Complex Systems, 2009)

Paul  Krugman’s  view  of agent-­‐based  modeling    

!“Oh, and about RogerDoyne Farmer (sorry, Roger!) and Santa Fe and complexity and all that: I was one of the people who got all excited about the possibility of getting somewhere with very detailed agent-based models — but that was 20 years ago. And after all this time, it’s all still manifestos and promises of great things one of these days.” !Paul Krugman, Nov. 30, 2010, in response to an article about INET housing project in WSJ.

�4

Reminder:    All  economics  models  are  agent-­‐based  models

• ABMs  are  computa(onal  agent-­‐based  models  (ACE)

�5

Why  isn’t  ABM  the   mainstay  of  economics?

• Math  culture  is  deeply  rooted  – papers  scored  too  much  on  math  vs.  science  – disdain  and  distrust  of  simula$on  – fascina$on  with  ra$onality  and  op$mality  

• ABM  is  a  fringe  ac$vity,  hasn’t  delivered  home  runs  needed  to  enter  establishment  – chicken/egg  problem    

• Lucas  cri$que

�6

Lucas   Cri$que

• Recession  of  70’s.    “Keynesian”  econometric  models.  • Phillips  curve:    Rising  prices  ~  rising  employment  • Following  Keynesians,  Fed  inflated  money  supply  • Result:    Inflation,  high  unemployment  =  stagflation  • Problem:    People  can  think  • Conclusion:    Macro  economic  models  must  incorporate  human  reasoning  

• Solution:    Dynamic  Stochastic  General  Eq.  models

�7

Advantages  of  DSGE

• “Micro-­‐founded”  (unlike  econometric  models)  – can  be  used  for  policy  analysis.  

• Time  series  models  – ini$alizable  in  current  state  of  the  world,  can  make  condi$onal  forecasts  

• Describe  a  specific  economy  at  a  specific  $me.  • In  some  sense  parsimonious

�8

Why  agent-­‐based  modeling?

• Diversifies toolkit of economics: Complements DSGE and econometric models. Also microfounded

• Time is ripe: increased computer power, Big Data, behavioral knowledge. Never let a crisis go to waste.

• Hasn’t really been tried yet -- crude estimates:– econometric models: 30,000 person-years– DSGE models: 20,000 person-years– agent-based models: 500 person-years

• Successes elsewhere: Traffic, epidemiology, defense• Examples of successes in economics:

– Endogenous explanations of clustered volatility and heavy tails; firm size; neighborhood choice

�9

Advantages• Can faithfully represent real institutions• Easily captures instabilities, feedback, nonlinearities,

heterogeneity, network structure,...• Shocks can be modeled endogenously• Easy to do policy testing• Easy to incorporate behavioral knowledge• Can calibrate modules independently using micro

data -- much stronger test of models!– In some sense between theory and econometrics

• ABMs synthesize knowledge: – Possible to understand what is not understood

�10

Challenges• Little prior art • Developing appropriate abstractions –What to include, what to omit?– How to keep model simple yet realistic?• Micro-data to calibrate decision rules?• Data censoring problems• Realistic agent-based models are complicated.• No theoretical foundation

Cautionary tale of weather forecasting

�11

Formula$ng  decision  rules  • Make  something  up  • Take  from  behavioral  literature  • Perform  experiments  in  context  of  ABM  • Interview  domain  experts  • Calibrate  against  microdata  • Learning  and  selec$on,  Lucas  cri$que  !(ABM  can  respond  to  Lucas  cri$que)

�12

Exis$ng  ABMS  in  economics• Almost  all  are  qualita$ve  • Range  of  complexity,  e.g.  – zero/low  intelligence  con$nuous  double  auc$on  – latent  order  book  (Bouchaud  group)  – Lebaron,  Brock  Hommes  trend  follow/fundamentalist  – Axtell  firm  size  – Thurner  et  al.  leveraged  value  investors  – SFI  Stock  Market  – Dosi-­‐group  – EURACE

�13

Is  it  possible  to  make  a  quan$ta$ve  ABM  that  can  be  used  as  a  $me  series  model? (and  therefore  can  compete  with  DSGE)

�14

Housing  model  project

• Senior  collaborators:    Rob  Axtell,  John  Geanakoplos,  Peter  Howil  

• Junior  collaborators:  Ernesto  Carella,  Ben  Conlee,  Jon  Goldstein,  Malhew  Hendrey,  Philip  Kalikman  

• Funded  by  INET  three  years  ago  for  $375,000.

�15

Agent-­‐based  model  of  housing  market

• Goal:    condi$onal  forecasts  and  policy  analysis  • Simula$on  at  level  of  individual  households  • Exogenous  variables:  demographics,  interest  rates,  lending  policy,  housing  supply.  

• Predicted  variables:    prices,  inventory,  default  • 16  Data  sets:    Census,  mortgages  (Core  Logic),tax  returns  (IRS),  real  estate  records  (MLA),  ...    

• Current  goal:    Model  Washington  DC  metro  area  • Future  goal:    All  metro  areas  in  US  

�16

Module  examples• Desired  expenditure  model  – buyers’  desired  home  price  as  a  func$on  of  household  income  and  wealth  

• Seller’s  pricing  model  – seller’s  offering  price  as  a  func$on  of  home  quality,  $me  on  market,  and  total  inventory  

• Buyer-­‐seller  matching  algorithm  – links  buyers  and  sellers  to  make  transac$ons  

• Household  wealth  dynamics  – models  consump$on  and  savings  

• Loan  approval  – qualifies  buyers  for  loans  based  on  income,  wealth;  must  match  issued  mortgages

�17

Housing  model  algorithmAt  each  $me  step:  

• Input  changes  to  exogenous  variables  • Update  state  of  households  – income,  consump$on,  wealth,  foreclosures,  ...  

• Buyers:      –Who?    Price  range?    Loan  approval,  terms?  

• Sellers:      –Who?    Offering  price?    Price  updates?  

• Match  buyers  and  sellers  – Compute  transac$ons  and  prices

�18

Results  when  we  fit  parameters  to  match  the  target  data

�19

Tentative conclusion: Lending policy is dominant cause of housing bubble in Washington DC.

Results obtained by hand-fitting parameters

(this is an early slide)

Results  when  we  fit  each  module  separately  on  data  that  is  not  the  

target  data.

�21

Case Shiller

1998 2000 2002 2004 2006 2008 2010

1.0

1.5

2.0

2.5

Index, first period = 1ModelData

Average House Sale Price

1998 2000 2002 2004 2006 2008 2010

1e+05

2e+05

3e+05

4e+05

5e+05Dollars

Sold Price to OLP

1998 2000 2002 2004 2006 2008 20100.85

0.90

0.95

1.00

Fraction

Active Listings

1998 2000 2002 2004 2006 2008 2010

20000

40000

60000

80000

100000

120000 NumberUnits Sold

1998 2000 2002 2004 2006 2008 2010

5000

10000

15000

20000

25000

30000Number*

*Data is smoothed with centered11−month moving average.

Days on Market

1998 2000 2002 2004 2006 2008 2010

0

50

100

150

200

250Days

Months of Inventory

1998 2000 2002 2004 2006 2008 201005

101520253035

MonthsModelData

Homeownership Rate

1998 2000 2002 2004 2006 2008 2010

60

62

64

66

68

70Percent

Vacancy Rate

1998 2000 2002 2004 2006 2008 2010

1.0

1.5

2.0

2.5

3.0

3.5

4.0Percent

ModelData

Housing Market ResultsBaseline result

Case Shiller

1998 2000 2002 2004 2006 2008 2010

1.0

1.5

2.0

2.5

Index, first period = 1ModelData

Average House Sale Price

1998 2000 2002 2004 2006 2008 2010

2e+05

3e+05

4e+05

5e+05Dollars

Sold Price to OLP

1998 2000 2002 2004 2006 2008 2010

0.90

0.95

1.00

Fraction

Active Listings

1998 2000 2002 2004 2006 2008 2010

2e+04

4e+04

6e+04

8e+04

1e+05

NumberUnits Sold

1998 2000 2002 2004 2006 2008 2010

5000

10000

15000

20000

25000

30000

Number*

*Data is smoothed with centered11−month moving average.

Days on Market

1998 2000 2002 2004 2006 2008 2010

0

50

100

150

200

Days

Months of Inventory

1998 2000 2002 2004 2006 2008 20100

5

10

15

20

25Months

ModelData

Homeownership Rate

1998 2000 2002 2004 2006 2008 2010

60

62

64

66

68

70Percent

Vacancy Rate

1998 2000 2002 2004 2006 2008 2010

1.0

1.5

2.0

2.5

3.0

3.5

4.0Percent

ModelData

Housing Market Resultsfixed interest rate

Case Shiller

1998 2000 2002 2004 2006 2008 2010

1.0

1.5

2.0

2.5

Index, first period = 1ModelData

Average House Sale Price

1998 2000 2002 2004 2006 2008 2010

1e+05

2e+05

3e+05

4e+05

5e+05Dollars

Sold Price to OLP

1998 2000 2002 2004 2006 2008 2010

0.90

0.95

1.00

Fraction

Active Listings

1998 2000 2002 2004 2006 2008 2010

0

20000

40000

60000

80000

NumberUnits Sold

1998 2000 2002 2004 2006 2008 2010

5000

10000

15000

20000

25000 Number*

*Data is smoothed with centered11−month moving average.

Days on Market

1998 2000 2002 2004 2006 2008 2010

0

50

100

150

Days

Months of Inventory

1998 2000 2002 2004 2006 2008 2010

2

4

6

8

10

12

MonthsModelData

Homeownership Rate

1998 2000 2002 2004 2006 2008 2010

60

62

64

66

68

70Percent

Vacancy Rate

1998 2000 2002 2004 2006 2008 2010

1.01.52.02.53.03.54.0

PercentModelData

Housing Market Resultsfixed lending policy

• Complete  agent-­‐based  model  of  economy  • Agents:    Households,  firms,  banks,  mutual  funds,  central  banks.    Both  financial  and  macro.  

• Goals:  –  tool  for  policy  decision  making  – series  of  models  of  increasing  complexity  – create  standard  sopware  library    – Be  useful  for  central  banks

CRISISComplexity Research Initiative

for Systemic InstabilitieS

EUROPEAN

COMMISSION

�25

Mark  2

�26

CRISIS  schema$c

Produc$on  sector

• Input-­‐output  economy  – firms  are  myopic  profit  maximizers  that  use  heuris$cs  to  set  price  and  quan$ty  of  produc$on  

– variable  labor  supply  – finance  produc$on  via  mixture  of  credit  and  equity  – input-­‐output  structure  mimicking  real  economy  

• For  comparison  have  simpler  alterna$ves,  e.g.  fixed  labor  Cobb  Douglas,  exogenous  dividends.

�27

Financial  sector• Banks    – take  deposits  from  firms  and  households,  lend  to  firms,  buy  and  sell  shares,  par$cipate  in  interbank  market.  

– Investment  strategies:  trend  following,  fundamental  Central  bank  – conven$onal  and  unconven$onal  policy  opera$ons  – interest  rate  can  be  formed  endogenously  

• Firms  – borrow  from  banks  to  fund  produc$on

�28

Unconven$onal  policy  opera$ons:  purchase  &  assump$on,  bailout,  bail  in

�29

Conclusions• We  have  lots  of  work  to  do  to  make  models  that  can  seriously  compete  with  DSGE  

• Should  be  possible  to  make  model  with  rich  ins$tu$onal  structure,  calibrated  to  real  world  

• Capability  to  put  an  economy  in  current  state  of  a  real  economy,  make  condi$onal  forecasts  

• Economic  models  of  future  will  be  ABM  – but  when?  

• Chicken-­‐egg  problem  to  get  ABM  off  the  ground

�30

Conclusions• Must  respect  ins$tu$onal  structure  – Impossible  to  do  everything  at  once:    open  in  conflict  with  understanding  strategic  reasoning  

– danger  of  strict  requirement  for  “economic  content”  • Fundamental  problem  in  macro  is  lack  of  data  – only  hope  is  ABM  with  microdata  calibra$on  

• Want  different  tools  for  different  jobs  —  diversity  – simple  models  for  understanding  mechanism  – richer  models  for  quan$ta$ve  understanding  

Economics  needs  to  allow  more  diversity!

�31

Future  versions   will  include:

!• Mortgage  markets  • Realis$c  input-­‐output  structure  • Deriva$ve  markets  • Bond  markets  • Shadow  banking  system

�32

Design  philosophy• As  simple  as  possible  (but  no  more)  • Design  model  around  available  data  • Fit  modules  and  agent  behaviors  independently  from  

target  data,  using  several  different  methods:  – micro-­‐data  for  calibra$on  and  tes$ng  – consult  domain  experts  for  behavioral  hypotheses  – adap$ve  op$miza$on  to  cope  with  Lucas  cri$que  – economic  experiments  

• Systema$cally  explore  model  sensi$vi$es  • Plug  and  play  • Standardized  interfaces  • Industrial  code,  sopware  standards,  open  source

�33