Complex systems: surprises on the way from micro to macro
J.P. Bouchaud
http://www.cfm.fr
Endogeneous crises?
Financial markets, the economy, many other social phenomena exhibit crises, ruptures, sudden discontinuities that resemble far-from-equilibrium phenomena in complex systems
Accumulating empirical evidence for positive feedback loops, self-reexivity and endogenous crises
Most price jumps appear unrelated to any news at all! Volatility after jumps has the same features as earthquake aftershocks
Market stability is not even an interesting question (Friedman) when feedback loops and instabilities are everywhere!
Financial market turbulence
30
10
20
90
95
100
0 95.0
95.5
96.0
10
50
100
Highly intermittent dynamics, similar to endogeneous noise in complex systems
Complex systems in a nutshell
What is a complex system?
The behaviour of the whole is dicult to guess from the behaviour of its parts
interactions lead to essentially new phenomena
Small causes can lead to large eects (fragility)
cascade, avalanches (cf. bead packs, trac jams)
There are many equilibrium states (often exponentially many in N )
hysteresis, slow dynamics and the impossibility to reach equilibrium in a reasonable time, but rather long periods of stasis separated by crises
Interactions: can be good, can be bad...
Spectacular collective patterns A. Cavagna et al.
Cartoon model for self-referential behaviour I
Binary decision of agent i: Di = 0, 1, inuenced by the aggregate choice Relative benet of the switch Di = 0 Di = 1: v(0 1) = + ; = N 1
i
Di
> 0: cost of leaving the 0 state (e.g.: pay ones taxes, cost of moving to solar heating, etc.) > 0: incentive due to either social/peer pressure, true benets due to reduced costs, increased usefulness (phones)
Choice Theory: non-deterministic (non-rational) choices: ev P (0 1) = ; v 1+e P (1 0) = 1 1 + ev
Collective equilibrium states
The Ising/Weidlich model (Curie-Weiss; Brock-Durlauf)
When < 4 (strong irrationality, weak collective eects):
a unique solution decribing a mixed population, constantly evolving between the two choices
When > 4 (strong collective eects)
two polarized solutions; in equilibrium the population should jump discontinously from 0 for < 2 to 1 for > 2.
Beyond equilibrium: Dynamics
In the polarized phase > 4:
The time needed to cross the barrier spontaneously is given by:
N 2 c exp 2 2 2
Unless is really small or really large, this never happens in a large population! The tragedy of the commons: everybody knows that = 1 is a better social optimum, but it is individually too costly for anyone to make the move
Solutions to escape from bad minimum: Enforce 0 (e.g. make the cost of cheating high), nucleation, synchronization (instead of independent decisions),...
Cartoon model for self-referential behaviour II
Now the agents are both heterogeneous and inuenced by the behaviour of others, and for simplicity the update rule: Di(t) = 1 (1 + sign [i (t) + (t)]) , 2
personal opinion, propensity or utility i heterogeneous with rms public information (price, technology level, news, zeitgeist) (t) (for illustration purposes, smooth in time) social pressure or imitation eects (t) (as above)
The RFIM: a unifying framework for many phenomena
Cartoon model of self-referential behaviour II
Breakdown of Representative Agent & Spontaneous discontinuities
Cartoon model of self-referential behaviour II
< c: personal choices dominate, smooth demand curve
> c: herding dominates, strong distortion/amplication of the fundamental demand curve: discontinuities appear at the macro level imitations induced panic/crashes/mistrust
Hysteresis in and out of the crisis
c: avalanche dynamics (power-law distribution of sizes)
c : More dispersed opinions avoid polarisation and stabilizes the system
Cartoon model of self-referential behaviour II
1.0 0.2 0.0 0.8 Normalized sound intensity 0.2 0.4 0.6 0.6 0.8 5 0 5
0.4
0.2
Sharp Fast Slow
0.0
0 Time (sec)
End of clappings
Application: spontaneous evaporation of trust
20 10
-10
-20
-30
-40
10000
20000
30000
40000
50000
Number of trust bonds and average trust as a function of time I trust you because he trusts you because I trust you As trust builds up, the system becomes more fragile cf. Battiston et al., Kirman et al.
Conclusion, lessons for agent based modelling
Surprises on the way from micro to macro: interactions can generate discontinuities at the macro level, even in a smoothly evolving world Equilibrium analysis is not enough: one can be dynamically stuck in the wrong minimum, or in fact evolve forever in a rough landscape
rich dynamics with transient stasis, punctuated by microcrises mimics (r)evolution/regime shifts
Increased eciency generically means more instabilities
Critical fragility to external perturbations and small changes of parameters small local shocks generically trigger large systemic eects
Rough landscape
Rather than:
