Economische Wetenschap na Tinbergen in een complexe wereld · RelatiemetTinbergen...
Transcript of Economische Wetenschap na Tinbergen in een complexe wereld · RelatiemetTinbergen...
Economische Wetenschap na Tinbergenin een complexe wereld
Cars Hommes
Universiteit van Amsterdam
Alumni LezingUvA Kringen Amsterdams Economen & Andragologie
UvA, 17 Mei, 2018
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 1 / 40
Relatie met TinbergenJ. Tinbergen, Zeitschrift für Nationalökonomie, April 30, 1930
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 2 / 40
Relatie met Tinbergen
Common featuresmathematical modelsempirical relevancepolicy analysis
New features:behavioural modelingcomplex systems approach
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 3 / 40
Main Themes of the Talk
Economy as a Complex Evolving Systemsimple behavioral agent-based models with heterogeneousagentsEmpirical Validation
stock markethousing market
Laboratory Macro ExperimentsPolicy implications: how to manage complex socio-economicsystems?
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 4 / 40
Introduction
Literature
Hommes, C.H. (2013), Reflexivity, expectations feedback and almos self-fulfillingequilibria: economic theory, empirical evidence and laboratory experiments, Journalof Economic Methodology special issue on Reflexivity and Economics: George Soros’Theory of Reflexivity and the Methodology of Economic Science, 406-419.Battiston, S., Farmer, J.D., Flache, A., Garlaschelli, D., Haldane, A.G., Heesterbeek,H., Hommes, C.H., Jaeger, C., May, R. and Scheffer, M. (2016), Complexity theoryand financial regulation. Economic policy needs interdisciplinary network analysisand behavioral modeling, Science Vol. 351, 6275, 818-819.Hommes, C.H., (2013), Behavioral Rationality and Heterogeneous Expectations inComplex Economic Systems, Cambridge.
Behavioral Rationality and Heterogeneous Expectations
in Complex Economic Systems
Cars Hommes
“Nosto ea facin ulput veros del utem zzrit duisseq uamet, si esse vent am et euis
nonsectet ercidunt prat, consecte min eleniam zzrit essecte feugue vel iusto odo
coreetu eraesequat ulla feui ea feuguer cillam zzrilla ad doluptat ad te facillamet,
quat do consenim exer ipsummolore delis nulluptat. Lutat. Feugait ulla cor
sequam, sequisl ullamcore feu feugiamet, velit aliqui blaorer ostrud dit non ut at
ex et lum eugiate volore faccum nim estie velit dolore magniscinit alit lum ex et,
quat.”
SOMEBODY, somewhere
“Nosto ea facin ulput veros del utem zzrit duisseq uamet, si esse vent am et euis
nonsectet ercidunt prat, consecte min eleniam zzrit essecte feugue vel iusto odo
coreetu eraesequat ulla feui ea feuguer cillam zzrilla ad doluptat ad te facillamet,
quat do consenim exer ipsummolore delis nulluptat. Lutat. Feugait ulla cor
sequam, sequisl ullamcore feu feugiamet, velit aliqui blaorer ostrud dit non ut at
ex et lum eugiate volore faccum nim estie velit dolore magniscinit alit lum ex et,
quat.”
SOMEBODY, somewhere
“Nosto ea facin ulput veros del utem zzrit duisseq uamet, si esse vent am et euis
nonsectet ercidunt prat, consecte min eleniam zzrit essecte feugue vel iusto odo
coreetu eraesequat ulla feui ea feuguer cillam zzrilla ad doluptat ad te facillamet,
quat do consenim exer ipsummolore delis nulluptat. Lutat. Feugait ulla cor
sequam, sequisl ullamcore feu feugiamet, velit aliqui blaorer ostrud dit non ut at
ex et lum eugiate volore faccum nim estie velit dolore magniscinit alit lum ex et,
quat.”
SOMEBODY, somewhere
Cover designed by Hart McLeod Ltd
Behavioral Rationality and Heterogeneous Expectations
in Complex Econom
ic Systems
Hom
mes
9781
1070
1929
4 H
OM
ME
S -
BE
HA
VIO
RA
L R
ATI
ON
ALI
TY
AN
D H
ETE
RO
GE
NE
OU
S E
XP
EC
TATI
ON
S
C M
Y K
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 5 / 40
Introduction
Traditional Rational View in Economics
representative, average agent, who is perfectly rationalexpectations are model consistentFriedman hypothesis: “irrational agents will lose money andwill be driven out the market by rational agents”simple (linear), stable model, driven by exogenous randomnews about fundamentals (crisis ≡ large shock)prices reflect economic fundamentals (market efficiency)Lucas: macroeconomic policy should be based on rationalexpectations
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 6 / 40
Introduction
Alternative Complexity ApproachBehavioural Agent-Based Models
heterogeneous agents, heterogeneous beliefsmarket psychology, herding behavior (Keynes (1936))bounded rationality (Simon (1957))markets as complex adaptive, nonlinear evolutionary systemsinteractions of agents’ individual (micro) decision rules createemergent aggregate (macro) structure explaining observedstylized factsnonlinear and critical transitions:small changes at micro-level may lead to large and irreversiblechanges at macro levelpolicy: avoid “dark corners” and undesirable critical transitions
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 7 / 40
Introduction
Examples of Complex Systems
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 8 / 40
Introduction
Challenges for Complexity Agent-based Approach
how to model non-rational agents?in economics the “particles can think”need a theory of adaptive behaviour and learning‘wilderness’ of bounded rationalitymany degrees of freedom for heterogeneitywhat exactly causes the outcome in a (large) computational HAM
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 9 / 40
Introduction
How to Discipline Bounded Rationality?
stylized agent-based modelsbehavioral rationality –behavioral consistency:simple heuristics that work reasonably wellevolutionary selection (‘survival of the fittest’) andreinforcement learninglaboratory experiments to study group behaviour:simultaneous test of individual decision rules and aggregate macrobehavior
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 10 / 40
simple ABM
Simple 2-type Behavioural Switching ModelBrock-Hommes, 1997,1998
two types financial market model:fundamentalists believe price will return to RE benchmarkchartists believe price will follow past trends
reinforcement learning / evolutionary selection:investors gradually switch to better performing strategy
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 11 / 40
simple ABM
Rational Fundamental Benchmark
standard asset pricing equilibrium:(common beliefs on future dividends Et[yt+1])
pt =1
REt[pt+1 + yt+1], R = 1 + r
strong positive feedback: 1/R ≈ 1
unique rational fundamental solution p∗t :(discounted sum expected future cash flow)
p∗t =Et[yt+1]
R+Et[yt+2]
R2+ · · ·
For special case of IID dividends, with Et[yt+1] = y:
p∗ =y
R− 1=y
r
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 12 / 40
simple ABM
Model in deviations from fundamental
deviation from fundamental
xt = pt − p∗
homogenous benchmark model in deviations:
xt =1
REtxt+1
Behavioural model with heterogeneous beliefs:(in deviations from RE-fundamental)
xt =1
R
H∑h=1
nhtEhtxt+1
strong positive feedback: weak mean reversion
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 13 / 40
simple ABM
Evolutionary selection of strategiesBrock and Hommes, Econometrica 1997
evolutionary selection or reinforcement learning:more successful strategies attract more followers
fractions of belief types are gradually updated in each period:(discrete choice model with asynchronous updating)
nht = δnh,t−1 + (1 − δ)eβUh,t−1
Zt−1
where Zt−1 is normalization factor.Uht fitness measure (e.g. realized profits)β is intensity of choice.δ asynchronous updating
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 14 / 40
simple ABM
Bubble and Crash Dynamics due to switchingbetween Fundamentalists versus ChartistsBrock and Hommes, JEDC 1998
Fundamentalists can not drive out chartistsdriven by short run profitability
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 15 / 40
simple ABM
Empirical Validation Two type model
Two trader types, with forecasting rules
f1t = φ1xt−1, 0 ≤ φ1 < 1 fundamentalistsf2t = φ2xt−1, φ2 > 1, trend extrapolators
xt =1
R[n1tφ1xt−1 + (1 − n1t)φ2xt−1] + εt
φt =ntφ1 + (1 − nt)φ2
Rmarket sentiment
φt < 1: mean reversion;φt > 1: explosive, trend following
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 16 / 40
simple ABM
2-type Behavioural Switching ModelHommes and in’t Veld, 2017
excess volatility: the stock market fluctuates much more than therational fundamental price p∗ based on dividends
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 17 / 40
simple ABM
Bubbles and Crashesdue to time-varying Market Sentiment
fractionfundamentalists
market sentiment
temporary bubbles and crashes: triggered by fundamentals andstrongly amplified by trend-following behaviour
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 18 / 40
simple ABM
Bubbles and Crashes in Housing Marketsjoint with DNB in NWO Comlexity program
Time
pric
e in
dex
1970 1980 1990 2000 2010
100
120
140
160
real house pricefundamental real house price
US
Time
pric
e in
dex
1970 1980 1990 2000 2010
100
150
200
300
real house pricefundamental real house price
NL
Time
pric
e in
dex
1970 1980 1990 2000 2010
100
200
300
500
real house pricefundamental real house price
JP
1
Outline of the NWO strategic theme
Dynamics of complex systems
Netherlands Organisation for Scientific Research
Complexity
Time
X
1970 1980 1990 2000 2010
−0.
50.
00.
51.
0
US
Time
X
1970 1980 1990 2000 2010
−0.
50.
00.
51.
0
NL
Time
X
1970 1980 1990 2000 2010−
0.5
0.0
0.5
1.0
JP
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 19 / 40
simple ABM
Persistent Bubbles and Crashes in Housing MarketsBolt, Demertzis, Diks, Hommes and van der Leij, 2017
−0.
10.
00.
10.
20.
3
X
0.0
0.2
0.4
0.6
0.8
1.0
n 1
1970 1980 1990 2000 2010
0.90
0.95
1.00
1.05
Time
AR
(1)
US
−0.
20.
00.
20.
40.
6
X
0.0
0.2
0.4
0.6
0.8
1.0
n 1
1970 1980 1990 2000 2010
0.94
0.98
1.02
1.06
Time
AR
(1)
NL
−0.
20.
00.
20.
4
X
0.0
0.2
0.4
0.6
0.8
1.0
n 1
1970 1980 1990 2000 2010
0.95
1.00
1.05
Time
AR
(1)
JP
1
Outline of the NWO strategic theme
Dynamics of complex systems
Netherlands Organisation for Scientific Research
Complexity
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 20 / 40
Experiments
Laboratory Experiments in Macro and Finance
laboratory test for (simple) complex systems;study individual (micro) as well as aggregate (macro) behaviorin controlled laboratory environmentempirical foundation for individual decision rules foragent-based models ABMs to discipline wilderness of boundedrationalitylaboratory test for policy analysis;test policies in more realistic controlled lab environment
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 21 / 40
Experiments
Learning to Forecasts Laboratory ExperimentsRepeated Keynesian Beauty Contest Game
individuals only have to forecast prices, ceteris paribus,with all other behavior computerized by theory
price depends on average forecast: pt = f(pet+1)
Round Prediction Real value
1 33,70 50,232 33,70 56,633 37,00 65,324 40,10 65,005 43,50 66,126 50,00 64,537 48,35 58,358 38,70 42,359 30,10 40,01
10 28,25
Total Earnings Remainingearnings: this period: time:
10357 1298 00
What is your prediction Prediction:this period?
Your prediction mustbe between 0 and 100
0102030405060708090
100
1 6 11 16 21 26 31 36 41 46
prediction
real number
Round
Number
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 22 / 40
Experiments
Financial Market Experimental Setting
asset pricing experiment (with/without robot trader)two-period ahead
positive feedback
mean dividend y = 3 and interest rate r = 0.05 are knownrational fundamental price pf = y/r = 60 not known(but can be computed)
Repeated Keynes’ Beauty Contest
pt =1
1 + r
((1 − nt)
pet+1,1 + · · · + pet+1,6
6+ nt p
f + y + εt
)
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 23 / 40
Experiments
Examples Simulation Benchmark Models
anchor and adjustment trend-following rule
pet+1 =60 + pt−1
2+ (pt−1 − pt−2)
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 24 / 40
Experiments
Financial Market Experiment (with Robot Trader)6 different markets
40
45
50
55
60
65
70
Pric
e
Group 2
fundamental price experimental price
40
45
50
55
60
65
70
Group 5
fundamental price experimental price
40
45
50
55
60
65
70
Pric
e
Group 1
fundamental price experimental price
40
45
50
55
60
65
70
Group 6
fundamental price experimental price
10 20 30 40 50 60 70 80 90
0 10 20 30 40 50
Pric
e
Group 4
fundamental price experimental price
40
45
50
55
60
65
70
0 10 20 30 40 50
Group 7
fundamental price experimental price
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 25 / 40
Experiments
Financial Market ExperimentStrong coordination of individual forecasts and errors
35
45
55
65
0 10 20 30 40 50
Pred
ictio
ns
45
55
65
Pric
e
Group 2
-2 0 2
35
45
55
65
0 10 20 30 40 50
Pred
ictio
ns
45
55
65
Pric
e
Group 1
-5 0 5
10 30 50 70 90
0 10 20 30 40 50
Pred
ictio
ns
10
30
50
70
90
Pric
e
Group 4
-30 0
30
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 26 / 40
Experiments
Summary Results Asset Pricing Experiment
Results are inconsistent with rational, fundamental forecasting
One would like to explain:three qualitatively different patters
(almost) monotonic convergence
constant oscillations
damping oscillations
coordination of agents in their predictions
no homogeneous expectations model fits theseexperimentsneed heterogeneous expectations model
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 27 / 40
Experiments
Examples of Individual Predictions and Switching
48
50
52
54
56
58
60
0 10 20 30 40 50
Pric
e, P
red
icti
on
s
Group 2, participants 1 and 5
prediction 5prediction 1
price
52
54
56
58
60
62
64
66
0 10 20 30 40 50
Group 6, participant 1
prediction price
48
50
52
54
56
58
60
62
64
66
0 10 20 30 40 50
Pric
e, P
red
icti
on
Time
Group 1, participant 3
prediction price
40
45
50
55
60
65
70
75
0 10 20 30 40 50
Time
Group 7, participant 3
prediction price
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 28 / 40
Experiments
Heterogeneous ExpectationsHeuristics Switching ModelAnufriev and Hommes, AEJ:Micro 2012
agents choose from a number of simple forecasting heuristics
performance based reinforcement learning:(extension of Brock and Hommes, Econometrica 1997)agents evaluate the performances of all heuristics, and tend toswitch to more successful rules; impacts are evolving over time
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 29 / 40
Experiments
Heuristic Switching Model: four forecasting heuristicsAnufriev and Hommes, AEJ:Micro 2012
adaptive expectations rule, [w = 0.65]
ADA pe1,t+1 = 0.65 pt−1 + 0.35 pe1,t
weak trend-following rule, [γ = 0.4]
WTR pe2,t+1 = pt−1 + 0.4 (pt−1 − pt−2)
strong trend-following rule, [γ = 1.3]
STR pe3,t+1 = pt−1 + 1.3 (pt−1 − pt−2)
anchoring and adjustment heuristic with learnable anchor
LAA pe4,t+1 = 12
(pavt−1 + pt−1
)+ (pt−1 − pt−2)
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 30 / 40
Experiments
Switching Model fitted to Experimental Data
35
45
55
65
0 10 20 30 40 50
Pred
ictio
ns
ADA WTR STR LAA 45
55
65
Pric
e
Group 5
simulation experiment
-2
0
2
35
45
55
65
0 10 20 30 40 50
Pred
ictio
ns
ADA WTR STR LAA 45
55
65
Pric
e
Group 6
simulation experiment
-5
0
5 45
55
65
75
0 10 20 30 40 50
Pred
ictio
ns
ADA WTR STR LAA
45
55
65
75
Pric
e
Group 7
simulation experiment
-10
0
10
0
0.2
0.4
0.6
0.8
1
0 10 20 30 40 50
Fractions of 4 rules in the simulation for Group 5
ADA WTR STR LAA
0
0.2
0.4
0.6
0.8
1
0 10 20 30 40 50
Fractions of 4 rules in the simulation for Group 2
ADA WTR STR LAA
0
0.2
0.4
0.6
0.8
1
0 10 20 30 40 50
Fractions of 4 rules in the simulation for Group 7
ADA WTR STR LAA
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 31 / 40
Experiments
More Financial Market Experiments
without RE robot trader large group size (n = 26)
0
200
400
600
800
1000
0 10 20 30 40 50
gr 1gr 2gr 3gr 4gr 5gr 6
Does positive feedback cause instability?
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 32 / 40
Experiments
Positive versus Negative Feedback ExperimentsRepeated Beauty Contest Games
negative feedback (strategic substitute environment)
pt = 60 − 20
21[
6∑h=1
1
6peht] − 60] + εt
positive feedback (strategic complementarity environment)
pt = 60 +20
21[
6∑h=1
1
6peht − 60] + εt
common feature: same RE equilibrium 60
only difference: sign in the slope of linear map +0.95 vs −0.95
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 33 / 40
Experiments
Feedback Mappings in LtFE
negative feedback positive feedback
20 40 60 80 100 120Prediction
20406080100120Price
20 40 60 80 100 120Prediction
20406080100120Price
pt = 60 − 2021
(pet − 60
)+ εt pt = 60 + 20
21
(pet − 60
)+ εt
My main concern with macroeconomics:rational expectations ignores almost self-fulfilling equilibria inpositive feedback systems
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 34 / 40
Experiments
Negative vs. Positive Feedback ExperimentsPrices, Individual Predictions and Errors; (Heemeijer et al., JEDC 2009)
0
20
40
60
80
100
0 10 20 30 40 50
Pred
ictio
ns
20
40
60
80
Pric
e
NEGATIVE
-3 0 3
0
20
40
60
80
100
0 10 20 30 40 50
Pred
ictio
ns
20
40
60
80
Pric
e
POSITIVE
-3 0 3
Positive Feedback: coordination on “wrong” non-RE price;coordination on almost self-fulfilling equilibria
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 35 / 40
Experiments
Prices in Experiments with Positive/Negative Feedback(7/6 groups)
negative feedback positive feedback
20
40
60
80
0 10 20 30 40 50
Pri
ce
Time
20
40
60
80
0 10 20 30 40 50Pr
ice
Time
convergence to REalmost self-fulfilling
equilibria
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 36 / 40
Experiments
Positive vs Negative Feedback; Small ShocksHeuristics Switching Model Simulations
prices strategy frequencies
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
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
positive feedback: trend-followers amplify fluctuationsCars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 37 / 40
Experiments
Managing Positive Feedback through Negative FB PolicyHousing Market Experiments, Bao and Hommes, 2017
no FB policylarge bubble
(λ = 0.95; r = 5%)
weak negativeFB policy
oscillations(λ = 0.85; r = 18%)
strong negativeFB policystable
(λ = 0.71; r = 40%)
0 10 20 30 40 500
100
200
300
400
500
600
700
800
900
1000
Period
Market N2
Market PriceSimulation
0 10 20 30 40 500
20
40
60
80
100
120
Period
Market L1
Market PriceSimulation
0 10 20 30 40 500
20
40
60
80
100
120
Period
Market H1
Market PriceSimulation
adding negative FB stabilizes complex positive FB system
Note: policy under RE: do not interfereCars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 38 / 40
Science, February 19, 2016Policy Dashboard based on Complexity Theory
INSIGHTS | PERSPECTIVES
sciencemag.org SCIENCE818 19 FEBRUARY 2016 • VOL 351 ISSUE 6275
By Stefano Battiston,1* J. Doyne Farmer,2,3
Andreas Flache,4 Diego Garlaschelli,5
Andrew G. Haldane,6 Hans Heesterbeek,7
Cars Hommes,8,9† Carlo Jaeger,10,11,12
Robert May,13 Marten Scheffer14
Traditional economic theory could not
explain, much less predict, the near
collapse of the financial system and its
long-lasting effects on the global econ-
omy. Since the 2008 crisis, there has
been increasing interest in using ideas
from complexity theory to make sense of eco-
nomic and financial markets. Concepts, such
as tipping points, networks, contagion, feed-
back, and resilience have entered the finan-
cial and regulatory lexicon, but
actual use of complexity models
and results remains at an early
stage. Recent insights and techniques offer
potential for better monitoring and manage-
ment of highly interconnected economic and
financial systems and, thus, may help antici-
pate and manage future crises.
TIPPING POINTS, WARNING SIGNALS. Fi-
nancial markets have historically exhibited
sudden and largely unforeseen collapses, at
a systemic scale. Such “phase transitions”
may in some cases have been triggered by
unpredictable stochastic events. More of-
ten, however, there have been endogenous
underlying processes at work. Analyses of
complex systems ranging from the climate
to ecosystems reveal that, before a major
transition, there is often a gradual and un-
noticed loss of resilience. This makes the sys-
tem brittle: A small disruption can trigger a
domino effect that propagates through the
system and propels it into a crisis state.
Recent research has revealed generic em-
pirical quantitative indicators of resilience
that may be used across complex systems to
detect tipping points. Markers include rising
correlation between nodes in a network and
rising temporal correlation, variance, and
skewedness of fluctuation patterns. These
indicators were first predicted mathemati-
cally and subsequently demonstrated experi-
mentally in real complex systems, including
living systems (1). A recent study of the
Dutch interbank network (2) showed that
standard analysis using a homogeneous net-
work model could only lead to late detection
of the 2008 crisis, although a more realistic
and heterogeneous network model could
identify an early warning signal 3 years be-
fore the crisis (see the first chart) (Fig. 1).
Ecologists have developed tools to quan-
tify the stability, robustness, and resilience
of food webs and have shown how these
depend on the topology of the network and
the strengths of interactions (3). Epidemi-
ologists have tools to gauge the potential for
events to propagate in systems of interacting
entities, to identify superspreaders and core
groups relevant to infection persistence, and
to design strategies to prevent or limit the
spread of contagion (4).
Extrapolating results from the natural
sciences to economics and finance presents
challenges. For instance, publication of an
early warning signal will change behav-
ior and affect future dynamics [the Lucas
critique (5)]. But this does not affect the
case where indicators are known only to
regulators or when the goal is to build bet-
ter network barriers to slow the spread of
contagion.
TOO CENTRAL TO FAIL. Network effects
matter to financial-economic stability be-
cause shock amplification may occur via
strong cascading effects. For example, the
Bank of International Settlements recently
developed a framework drawing on data on
the interconnectedness between banks to
gauge the systemic risk posed to the finan-
cial network by Global Systemically Impor-
tant Banks. Recent research on contagion in
financial networks has shown that network
topology and positions of banks matter; the
global financial network may collapse even
when individual banks appear safe (6). Cap-
turing these effects is essential for quanti-
fying stress on individual banks and for
looking at systemic risk for the network as
a whole. Despite on-going efforts, these ef-
fects are unlikely to be routinely considered
anytime soon.
Information asymmetry within a net-
work—e.g. where a bank does not know
about troubled assets of other banks—can
be problematic. The banking network typi-
cally displays a core-periphery structure,
with a core consisting of a relatively small
number of large, densely interconnected
banks that are not very diverse in terms of
business and risk models. This implies that
core banks’ defaults tend to be highly cor-
related. That, in turn, can generate a collec-
tive moral hazard problem (i.e., players take
on more risk, because others will bear the
costs in case of default), as banks recognize
that they are likely to be supported by the
authorities in situations of distress, the like-
lihood amplifies their incentives to herd in
the first place.
Estimating systemic risk relies on granu-
lar data on the financial network. Unfortu-
nately, business interactions between banks
are often hidden because of confidentiality
issues. Tools being developed to reconstruct
networks from partial information and to
estimate systemic risk (7) suggest that pub-
licly available bank information does not al-
low reliable estimation of systemic risk. The
estimate would improve greatly if banks
publicly reported the number of connec-
tions with other banks, even without dis-
closing their identity.
In addition to data, understanding the ef-
fects of interconnections also relies on in-
tegrative quantitative metrics and concepts
that reveal important network aspects, such
as systemic repercussions of the failure of
individual nodes. For example, DebtRank,
which measures the systemic importance
of individual institutions in a financial net-
work (8), shows that the issue of too-central-
to-fail may be even more important than
too-big-to-fail.
COMPLEX SYSTEMS
Complexity theory and financial regulationEconomic policy needs interdisciplinary network analysis and behavioral modeling
“…policies and financial regulation…are successful in stabilizing experimental macroeconomic systems’’
1Department of Banking and Finance, University of Zurich, 8032 Zürich, Switzerland. 2Institute for New Economic Thinking, Oxford Martin School, and Mathematical Institute, University of Oxford, Oxford OX1 2JD, UK. 3Santa Fe Institute, Santa Fe, NM 87501, USA. 4Department of Sociology, University of Groningen, 9712 TG Groningen, Netherlands. 5Lorentz Institute for Theoretical Physics, University of Leiden, 2311 EZ Leiden, Netherlands. 6Bank of England, London, EC2R 8AH, UK. 7Faculty of Veterinary Medicine, University of Utrecht, 3512 JE Utrecht, Netherlands. 8Amsterdam School of Economics, University of Amsterdam, 1018 WB Amsterdam, Netherlands. 9Tinbergen Institute, 1082 MS Amsterdam, Netherlands. 10Beijing Normal University, 100875 Beijing, China. 11Potsdam University, 14469 Potsdam, Germany. 12Global Climate Forum 10178 Berlin, Germany. 13Department of Zoology, University of Oxford, Oxford OX1 2JD, UK. 14Environmental Sciences, Wageningen University 6708 PB Wageningen, Netherlands. *Authors are in allphabetical order. †Corresponding author. E-mail: [email protected]
POLICY
SCIENCE sciencemag.org 19 FEBRUARY 2016 • VOL 351 ISSUE 6275 819
ILL
US
TR
AT
ION
: C
. S
MIT
H/SCIENCE
AGENTS AND BEHAVIOR. Agent-based
models (ABMs) are computer models in
which the behavior of agents and their in-
teractions are explicitly represented as de-
cision rules mapping agents’ observations
onto actions. Although ABMs are less well
established in analyzing financial-economic
systems than in, e.g., traffic control, epide-
miology, or battlefield conflict analyses, they
have produced promising results. Axtell (9)
developed a simple ABM that explains more
than three dozen empirical properties of
firm formation without recourse to external
shocks. ABMs provide a good explanation
for why the volatility of prices is clustered
and time-varying (10) and have been used
to test systemic risk implications of reforms
developed by the Basel Committee on Bank-
ing Supervision, which show how dynami-
cally changing risk limits can lead to booms
and busts in prices (11, 12). ABMs of market
dynamics can be linked with ABM work on
opinion dynamics in the social sciences (13)
to understand how propagation of opinions
through social networks affects emergent
macro behavior, which is crucial to manag-
ing the stability and resilience of socioeco-
nomic systems.
Laboratory experiments with human
subjects can provide empirical validation
of individual decision rules of agents, their
interactions, and emergent macro behav-
ior. Recent experiments studying behavior
of a group of individuals in the laboratory
show that economic systems may deviate
significantly from rational efficient equi-
librium at both individual and aggregate
levels (14). This generic feature of positive
feedback systems leads to persistent devia-
tions of prices from equilibrium and emer-
gence of speculation-driven bubbles and
crashes, strongly amplified by coordination
on trend-following and herding behavior
(15). There is strong empirical evidence of
these behaviors in financial markets in prac-
tice, and these controlled laboratory experi-
ments provide more detailed understanding
of mechanisms, causality, and conditions for
emergence of macro phenomena.
A simple behavioral model, with agents
gradually switching to better performing
heuristics, explains individual, as well as
emergent, macro behavior in these laboratory
economies. The experiments also provide
a general mechanism for managing social
contagion in such systems. For example,
monetary and fiscal policies and financial
regulation, designed to weaken positive feed-
back, are successful in stabilizing experimen-
tal macroeconomic systems when properly
calibrated (16). Complexity theory provides
mathematical understanding of these effects.
POLICY DASHBOARD. It is an opportune
time for academic economists, complexity
scientists, social scientists, ecologists, epi-
demiologists, and researchers at financial
institutions to join forces to develop tools
from complexity theory, as a complement
to existing economic modeling approaches
(17). One ambitious option would be an on-
line, financial-economic dashboard that in-
tegrates data, methods, and indicators. This
might monitor and stress-test the global so-
cioeconomic and financial system in some-
thing close to real time, in a way similar to
what is done with other complex systems,
such as weather systems or social networks.
The funding required for essential policy-
relevant and fundamental interdisciplinary
progress in these areas would be trivial com-
pared with the costs of systemic financial
failures or the collapse of the global finan-
cial-economic system. ■
REFERENCES AND NOTES
1. M. Scheffer et al., Science 338, 344 (2012). 2. T. Squartini et al., Sci. Rep. 3, 3357 (2013). 3. R. M. May et al., Nature 451, 893 (2008). 4. H. Heesterbeek et al., Science 347, aaa4339 (2015). 5. R. E. Lucas Jr., Carnegie-Rochester Conf. Ser. Public Policy
1, 19 (1976). 6. S. Battiston et al., J. Econ. Dynam. Control 36, 1121 (2012). 7. G. Cimini et al., Sci. Rep. 5, 15758 (2015). 8. S. Battiston et al., Sci. Rep. 2, 541 (2012). 9. R. Axtell, “Endogenous dynamics of multi-agent firms”
(Working paper version 1.5, Univ. of Oxford, 2014); www.css.gmu.edu/~axtell/Rob/Research/Pages/Firms.html.
10. B. LeBaron, in Handbook of Computational Economics, vol.2, Agent-Based Computational Economics, L. Tesfatsion, and K. L. Judd, Eds. (North-Holland, Amsterdam, 2006), pp. 1187–1233.
11. S. Thurner et al., Quant. Finan. 12, 695 (2012). 12. C. Aymanns, J. D. Farmer, J. Econ. Dyn. Control 50, 155
(2015). 13. A. Flache, M. W. Macy, J. Conflict Resolut. 55, 970 (2011). 14. T. Bao, C. Hommes, T. Makarewicz, “Bubble formation
and (in)efficient markets in learning-to-forecast and –optimize experiments” (TI 2015-107/II Working paper, Tinbergen Institute, Amsterdam http://papers.tinbergen.nl/15107.pdf.
15. C. H. Hommes, Behavioral Rationality and Heterogeneous Expectations in Complex Economic Systems (Cambridge Univ. Press, Cambridge, 2013).
16. T. Bao, C. H. Hommes, “When speculators meet constructors: Positive and negative feedback in experi-mental housing markets” CeNDEF Working paper 15-10, University of Amsterdam, Netherlands, 2015)]; http://bit.ly/WP15-10.
17. A. G. Haldane, “On microscopes and telescopes,” Workshop on Socio-Economic Complexity, Lorentz Center, Leiden, 23 to 27 March 2015 (Bank of England, London, 2015); http://bit.ly/1VIJlvX.
ACKNOWLEDGMENTS
We acknowledge financial support from the Netherlands Institute of Advanced Studies in the Humanities and Social Sciences, the Netherlands Organisation for Scientific Research, the Lorentz Center, and the Tinbergen Institute.
10.1126/science.aad0299
16
14
12
10
8
6
4
2
0
-2
-4
-6
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
Homogeneous
Heterogeneous
z-score
Pre-crisis Crisis
Normal z-score range
Early-warning signals of the 2008 crisis in the Dutch interbank network. The figure portrays a temporal
analysis of two loops, pairs of banks that are at the same time debtor and creditor to each other. Although the
raw number of two loops is not very informative about possible ongoing structural changes, its comparison
with a random network model benchmark is. A z-score represents the number of standard deviations by which
the number of two loops in the real network deviates from its expected value in the model. Small magnitude
z-scores (<3) indicate approximate consistency with the model, whereas larger magnitudes indicate statistically
significant deviations. Two different random network models were used: a homogeneous network with the
same total number of links as in the real network (top) and a heterogeneous network where every bank has the
same number of connections as in the real network (bottom). The homogeneous model, often used in standard
analyses, highlights only a late and abrupt structural change (2008). The more realistic heterogeneous model
also identifies a gradual, early-warning “precrisis” phase (2005–2007). [Modified from (2)]
Experiments
Concluding Remarks
New alternative complexity, agent-based approach is on its way...... depends much on new generation of students!policy analysis: identify and model nonlinearities to getgrip on complexity
Thank you very much!
1
Outline of the NWO strategic theme
Dynamics of complex systems
Netherlands Organisation for Scientific Research
Complexity
Published by:Netherlands Organisation for Scientific Research (NWO)
Visiting address:Laan van Nieuwe Oost-Indië 300The Hague
Postal address:P.O. Box 934602509 AL The HagueThe Netherlands
T: +31 (0)70 344 07 09E: [email protected] W: www.nwo.nl/complexity
The Hague, September 2014
This publication is supported by the
Royal Netherlands Academy of Arts and
Sciences (KNAW) Programme “Over Grenzen”.
Grip on ComplexityHow Manageable are Complex Systems?
Directions for future complexity research
Physical Sciences
Netherlands Organisation for Scientific Research
Cars Hommes (UvA) UvA Alumni Lezing 17 Mei, 2018 40 / 40