Inefficient Markets VI - HEC

Inefficient Markets VI
Damien Challet
[email protected]
October 22, 2014
Damien Challet
Inefficient Markets VI
Previous episodes
Adaptive Complex Market Hypothesis
Price returns: interesting noise?
Price returns: not chaos
Simple agent-based models
fundamentalists
trend followers
not playing
feedback from gains to strategy use
Interaction
Phase transitions: DANGER
Damien Challet
Inefficient Markets VI
Fractals
Mandelbrot (1924-2010)
fractal object: invariant under change of scale (zoom, de-zoom)
(romanesco)
Damien Challet
Inefficient Markets VI
Fractals
Mandelbrot set
Damien Challet
Inefficient Markets VI
Critical point: fractal
At critical point:
fractal spatial properties
≡ self-similar system
≡ scale invariance
≡ system same at all scales
Damien Challet
Inefficient Markets VI
Scale invariance and power-laws
Measure of something over self-similar system
g(x)
Change of scale by factor a:
g(ax) = h(a)g(x)
Hypothesis: g(1) = 1 =⇒ g(a) = h(a)
g(xy) = g(x)g(y)
Only solution:
g(x) = x−α
Damien Challet
Inefficient Markets VI
Benford’s law
Astronomer Newcomb (1881)
Logarithm tables: first pages are the most dirty
Measures and proposes Pdigit (d)
Damien Challet
Inefficient Markets VI
Benford’s law
Astronomer Newcomb (1881)
Logarithm tables: first pages are the most dirty
Proposes Pdigit (d)
Physicist Benford (1936)
law valid in variety of data sources
numbers in newspapers
length of rivers
taxation forms
accounting books
NOT phone books
Hill (1995)
Random numbers from a random set of probability distributions
→ Benford’s law
Damien Challet
Inefficient Markets VI
Benford’s law
First digit law
Pdigit (d is the first digit)
1
Pdigit (d) = log10 1 +
d
Pdigit (d) independent from unit (meters, feet, etc)
P(x) ∝ 1x , xmax
R d+1
Pdigit (d) =
P(x)
R xdmax
dxP(x)
1
Damien Challet
1
= log10 1 +
d
Inefficient Markets VI
Complex networks: how much complex?
Benchmark (zero model): Erdös-Renyi (ER) random graphs
k: number of links, P(k) ∝ exp(−k/p) (no hubs)
Complex networks
k: number of links, P(k) ∝ k−α (hubs)
Clustering
Assortativity
Damien Challet
Inefficient Markets VI
Complex networks: clustering
open triple
closed triple
The friend of my friend is also my friend
clustering coefficient CC =
3x number of triangles
number of links
Small world: CC CCER = c/N → 0
Damien Challet
Inefficient Markets VI
Complex networks: assortativity
Assortative: hubs link to hubs
Intuitively: E(knearest neighbours |k) = ∑k0 k0 P(k0 |k)
E(knearest neighbours |k) grows with k in assortative networks
Damien Challet
Inefficient Markets VI
Processes on networks
Transfer
Roads
Airplane routes
Electric wiring
Information
Internet wires
Emails
Money
Contagion
sentiment
risk
disease
panic
default
Damien Challet
Inefficient Markets VI
Network: systemic risk
E.g.: links between businesses
Caldarelli (2007)
Contagion
sentiment
risk
disease
panic
default
Cascading defaults
Damien Challet
Inefficient Markets VI
Epidemic propagation on heterogeneous networks
Vespgnagni et al (1999): SIS/SIR
Site i susceptible → infected → susceptible (SIS)
Site i susceptible → infected → removed (SIR)
If power-law link distribution, disease propagate much more
easily
Optimal immunisation strategy: hubs
Application
(computer) virus
rumors
financial defaults
Damien Challet
Inefficient Markets VI
Systemic crisis of banking systems (TRENDY)
Overnight network of lending
Bank i lends ci→j to bank j
Directed network
Assume bank k defaults
Default propagation?
If many banks affected , systemic risk
What about many banks defaulting at the same time for the same
reason (cf. Greece)?
Damien Challet
Inefficient Markets VI
Binary interactions and global sentiment: framework
2-by-2 interaction
Opinion of agent i: si ∈ {−1, 0, 1}
Agent j influences agent i: link Jij 6= 0
J: network of influences
Effective influence: Jij sj
Total influence on agent i: ∑j Jij sj
Total influence of agent i: si ∑j Jji
Damien Challet
Inefficient Markets VI
Interactions: framework
Convention: Jij > 0 ⇐⇒ i likes to agree with j
Frustration if Jij si sj < 0
Total frustration
H = − ∑ Jij si sj
ij
Central principle
Dynamics of si should tend to minimise H
Damien Challet
Inefficient Markets VI
Market sentiment dynamics: example
Michard-Bouchaud (2005)
N agents
si ∈ {−1, 1}
Opinion
Personal a priori towards si = +1 φi
heterogeneity
Global pressure
F(t)
Social pressure
∑j Jij sj (t)
Frustration of i
−si [F + φi + ∑j Jij sj (t)]
Damien Challet
Inefficient Markets VI
Opinion dynamics
Michard-Bouchaud (2005)
Dynamics
!
si (t + 1) = sign
F(t) + φi + ∑ Jij sj (t)
j
Total frustration
H = −NF ∑ si − ∑ φi si − ∑ Jij si sj
i
i
ij
a.k.a random-field ({φi }) Ising ({si }) model (RFIM)
Damien Challet
Inefficient Markets VI
Opinion shifts: absence of social influence
Jij = 0:
(
+1
si = sign (F + φi ) =
−1
φi > −F
φi < −F
Global opinion
O(t) =
if R(x) =
Rx
−∞ P(φ )dφ
O0 = −
1
si (t)
N∑
i
and N 1
R(−F) + (1 − R(−F)) = 1 − 2R(−F)
| {z }
|
{z
}
fraction of - fraction of +
Damien Challet
Inefficient Markets VI
Opinion shifts: social influence
Jij = J/N: complete graph, no distance, mean-field model
(
+1 φi > −F − JO
si (t + 1) = sign (F(t) + φi + JO(t)) =
−1 φi<−F−JO
O = 1 − 2R(−F − JO)
Small imitation J 1: R(−F − JO) ' R(−F) − P(−F)OJ
O'
Damien Challet
O0
1 − 2P(−F)J
Inefficient Markets VI
Abrupt opinion shifts
Response function: reaction to small perturbation
dO
dF
maximum at maximum of P(φ )
P(φ ) ∼ N(φ0 , σ ): max P(φ ) ∝ 1/σ
J small enough: no divergence
J = Jc ∝ 1/ max P(φ ) ∝ σ : divergence of
Damien Challet
dO
dF
Inefficient Markets VI
Abrupt opinion shifts
J small enough: no divergence
J = Jc ∝ σ : divergence of
dO
dF
J > Jc : O = 1 − 2R(−F − JO) has three solutions, hysteresis
Damien Challet
Inefficient Markets VI
Opinion shifts
Near critical point J . Jc
1
dO
=
G
dF
Jc − J
F − Fc (J)
(Jc − J)3/2
G : known exactly, indep
from P(φ )
G (0) = cst < ∞
G (x → ∞) ∝ x−2/3
Damien Challet
Inefficient Markets VI
Test for presence of social interaction
Near critical point J . Jc
dO
1
=
G
dF
Jc − J
max
dO
dF
1
Jc −J
dO
dF large
F − Fc (J)
(Jc − J)3/2
=h∝
near its max,
in window w ∝ (Jc − J)3/2
h ∝ w−2/3
Without interaction:
dO
dF
= 2P(−F): e.g. Gaussian
h ∝ w−1
Damien Challet
Inefficient Markets VI
Example: birth rates
Damien Challet
Inefficient Markets VI
Example: cell phone
Damien Challet
Inefficient Markets VI
Example: hand clapping
Damien Challet
Inefficient Markets VI
Validation I
Damien Challet
Inefficient Markets VI
Validation II
Damien Challet
Inefficient Markets VI
RFIM and finance
Iori (1999): RFIM + thresholds
si (t) ∈ {−1, 0, +1}
signal
Yi (t) = ∑ Jij sj (t) + ε(t) + νi (t)
j
where ε(t) ∼ N(0, σε ), νi (t) ∼ N(0, σν )
thresholds θi ∼ N(0, σθ )


+1
si (t + 1) = −1


0
Yi (t) > θi (t)
Yi (t) < −θi (t)
otherwise
Random ε(t) sweeps over critical point
Damien Challet
Inefficient Markets VI
RFIM and finance
Iori (1999): RFIM + thresholds
P(ξ ) unbiased: max at ξ = 0
Random ε(t) sweeps over critical point if J large enough
Damien Challet
Inefficient Markets VI
Summary
Phase transitions:
scale-invariance
extreme sensitivity
opinion dynamics
Binary models: si = ±1
Interaction on networks
Damien Challet
Inefficient Markets VI