Eigenvector stability:

Random Matrix Theory

and Financial Applications

J.P Bouchaud
with: R. Allez, M. Potters

http://www.cfm.fr
Portfolio theory: Basics
• Portfolio weights wi, Asset returns Xit

• If expected/predicted gains are gi then the expected gain of

the portfolio is
X
G= wi g i
i

• Let risk be defined as: variance of the portfolio returns

(maybe not a good definition !)
2
X
R = wiσiCij σj wj
ij

where σi2 is the variance of asset i, and

Cij is the correlation matrix.

J.-P. Bouchaud
Markowitz Optimization
• Find the portfolio with maximum expected return for a given
risk or equivalently, minimum risk for a given return (G)

• In matrix notation:
C−1g
wC = G T −1
g C g
where all gains are measured with respect to the risk-free
rate and σi = 1 (absorbed in gi).

• Note: in the presence of non-linear contraints, e.g.

X
|wi| ≤ A
i
an NP complete, “spin-glass” problem! (see [JPB,Galluccio,Potters])

J.-P. Bouchaud
Markowitz Optimization

• More explicitely:

λ−1 (λ−1
X X
w∝ α ( Ψα · w ) Ψα = g+ α − 1) (Ψα · w) Ψα
α α

• Compared to the naive allocation w ∝ g:

– Eigenvectors with λ ≫ 1 are projected out

– Eigenvectors with λ ≪ 1 are overallocated

• Very important for “stat. arb.” strategies

J.-P. Bouchaud
Empirical Correlation Matrix

• Empirical Equal-Time Correlation Matrix E

t t
1 X Xi Xj
Eij =
T t σiσj
Order N 2 quantities estimated with N T datapoints.

When T < N , E is not even invertible.

Typically: N = 500 − 1000; T = 500 − 2500

J.-P. Bouchaud
Risk of Optimized Portfolios

• “In-sample” risk
2 T 1
Rin = wE EwE =
gT E−1g

• True minimal risk

2 T Cw = 1
Rtrue = wC C
gT C−1g

• “Out-of-sample” risk

2 T gT E−1CE−1g
Rout = wE CwE =
(gT E−1g)2

J.-P. Bouchaud
Risk of Optimized Portfolios

• Let E be a noisy, unbiased estimator of C. Using convexity

arguments, and for large matrices:
2 2 2
Rin ≤ Rtrue ≤ Rout

• If C has some time dependence, one expects an even worse

underestimation

J.-P. Bouchaud
In Sample vs. Out of Sample

Return 150

100

50 Raw in-sample
Cleaned in-sample
Cleaned out-of-sample
Raw out-of-sample

0
0 10 20 30
Risk

J.-P. Bouchaud
Possible Ensembles (stationary case)
• Null hypothesis Wishart ensemble:
′
hXitXjt i = σi σj δij δtt′
Constant volatilities and X with a finite second moment

• General Wishart ensemble:

′
hXit Xjt i = σiσj Cij δtt′
Constant volatilities and X with a finite second moment

• Elliptic Ensemble
t t′
hXi Xj i = s σiσj Cij δtt′
Random common volatility, with a certain P (s)
(Ex: Student)

J.-P. Bouchaud
Null hypothesis C = I
• Goal: understand the eigenvalue density of empirical corre-
lation matrices when q = N/T = O(1)

t = (X t X t )/T
• Eij is a sum of (rotationally invariant) matrices Eij i j

• Free random matrix theory: R-transform are additive →

q
4λq − (λ + q − 1)2 √ 2 √
ρE (λ) = λ ∈ [(1 − q) , (1 + q)2]
2πλq
[Marcenko-Pastur] (1967) (and many rediscoveries)

• Any eigenvalue beyond the Marcenko-Pastur band can be

deemed to contain some information (but see below)

J.-P. Bouchaud
Null hypothesis C = I

• Remark 1: −GE (0) = hλ−1iE = (1 − q)−1, allowing to com-

pute the different risks:
Rin Rin
Rtrue = √ ; Rout =
1−q 1−q

• Remark 2: One can extend the calculation to EMA estima-

tors [Potters, Kondor, Pafka]:

Et+1 = (1 − ε)Et + εX tX t

J.-P. Bouchaud
General C Case
• The general case for C cannot be directly written as a sum
of “Blue” functions.

• Solution using different techniques (replicas, diagrams, S-

transforms):
1
Z
GE (z) = dλ ρC (λ) ,
z − λ(1 − q + qzGE (z))

• Remark 1: −GE (0) = (1 − q)−1 independently of C

• Remark 2: One should work from ρC −→ GE and postulate

a parametric form for ρC (λ), i.e.:
µA
ρC (λ) = 1+µ
Θ(λ − λmin)
(λ − λ0)

J.-P. Bouchaud
Empirical Correlation Matrix

1.5 8
Data
Dressed power law (µ=2) 6
Raw power law (µ=2)
4
Marcenko-Pastur

κ
2

1 0
-2
0 250 500
ρ(λ)

rank

0.5

0
0 1 2 3 4 5
λ

J.-P. Bouchaud
Eigenvalue cleaning

Return 150

100

50 Raw in-sample
Cleaned in-sample
Cleaned out-of-sample
Raw out-of-sample

0
0 10 20 30
Risk

J.-P. Bouchaud
What about eigenvectors?

• Up to now, most results using RMT focus on eigenvalues

• What about eigenvectors? What natural null-hypothesis?

• Are eigen-directions stable in time?

• Important source of risk for market/sector neutral portfolios:

a sudden/gradual rotation of the top eigenvectors!

• ..a little movie...

J.-P. Bouchaud
What about eigenvectors?
• Correlation matrices need a certain time T to be measured

• Even if the “true” C is fixed, its empirical determination

fluctuates:
Et = C + noise

• What is the dynamics of the empirical eigenvectors induced

by measurement noise?

• Can one detect a genuine evolution of these eigenvectors

beyond noise effects?

J.-P. Bouchaud
What about eigenvectors?

• More generally, can one say something about the eigenvec-

tors of randomly perturbed matrices:

H = H0 + ǫH1
where H0 is deterministic or random (e.g. GOE) and H1
random.

J.-P. Bouchaud
Eigenvectors exchange
• An issue: upon pseudo-collisions of eigenvectors, eigenvalues
exchange

• Example: 2 × 2 matrices
H11 = a, H22 = a + ǫ, H21 = H12 = c, −→
s
ǫ 2 ǫ2
λ± ≈ǫ→0 a + ± c +
2 4

• Let c vary: quasi-crossing for c → 0, with an exchange of the

top eigenvector: (1, −1) → (1, 1)

• For large matrices, these exchanges are extremely numerous

→ labelling problem

J.-P. Bouchaud
Subspace stability
• An idea: follow the subspace spanned by P -eigenvectors:
′ ′ ′
|ψk+1i, |ψk+2i, . . . |ψk+P i −→ |ψk+1 i, |ψk+2 i, . . . |ψk+P i

• Form the P × P matrix of scalar products:

′
Gij = hψk+i|ψk+j i

• The determinant of this matrix is insensitive to label per-

mutations and is a measure of the overlap between the two
P -dimensional subspaces

– Q = P1 ln | det G| is a measure of how well the first subspace

can be approximated by the second

J.-P. Bouchaud
Null hypothesis

• Note: if P is large, Q can be “accidentally” large

• One can compute Q exactly in the limit P → ∞, N → ∞,

with fixed p = P/N :

• Final result:([Wachter] (1980); [Laloux,Miceli,Potters,JPB])

Z 1
Q= ds ln s ρ(s)
0
with:
q
1 s2(4p(1 − p) − s2)+
ρ(s) = 2
.
p πs(1 − s )

J.-P. Bouchaud
Intermezzo
• Non equal time correlation matrices
t+τ
1 X Xit Xj
τ
Eij =
T t σiσj
N × N but not symmetrical: ‘leader-lagger’ relations

• General rectangular correlation matrices

T
1 X
Gαi = Yαt Xit
T t=1
N ‘input’ factors X; M ‘output’ factors Y

– Example: Yαt = Xjt+τ , N = M

J.-P. Bouchaud
Intermezzo: Singular values
• Singular values: Square root of the non zero eigenvalues
of GGT or GT G, with associated eigenvectors ukα and vik →
1 ≥ s1 > s2 > ...s(M,N )− ≥ 0

• Interpretation: k = 1: best linear combination of input vari-

ables with weights vi1, to optimally predict the linear com-
bination of output variables with weights u1α, with a cross-
correlation = s1.

• s1: measure of the predictive power of the set of Xs with

respect to Y s

• Other singular values: orthogonal, less predictive, linear com-

binations

J.-P. Bouchaud
Benchmark: no cross-correlations
• Null hypothesis: No correlations between Xs and Y s:

Gtrue ≡ 0

• But arbitrary correlations among Xs, CX , and Y s, CY , are

possible

• Consider exact normalized principal components for the sam-

ple variables Xs and Y s:
1
X̂it = √ Uij Xjt ; Ŷαt = ...
X
λi j

and define Ĝ = Ŷ X̂ T .

J.-P. Bouchaud
Benchmark: Random SVD
• Final result:([Wachter] (1980); [Laloux,Miceli,Potters,JPB])
q
+
(s2 − γ−)(γ+ − s2)
ρ(s) = (m + n − 1) δ(s − 1) +
πs(1 − s2)
with
q
γ± = n + m − 2mn ± 2 mn(1 − n)(1 − m), 0 ≤ γ± ≤ 1

• Analogue of the Marcenko-Pastur result for rectangular cor-

relation matrices

• Many applications; finance, econometrics (‘large’ models),

genomics, etc.

• Same problem as subspace stability: T −→ N , n = m −→ p

J.-P. Bouchaud
Sectorial Inflation vs. Economic indicators

0.8

0.6
Data
P(s’<s)

Benchmark

0.4

0.2

0
0.2 0.4 0.6 0.8 1
s

N = 50, M = 16, T = 265

J.-P. Bouchaud
Back to eigenvectors: perturbation theory

• Consider a randomly perturbed matrix:

H = H0 + ǫH1

• Perturbation theory to second order in ǫ yields:

ǫ2 hψi|H1|ψj i 2
!
X X
|det(G)| = 1 − .
2 i∈{k+1,...,k+P } j6∈{k+1,...,k+P } λi − λj

J.-P. Bouchaud
The case of correlation matrices
• Consider the empirical correlation matrix:
T
1 X
E=C+η η= (X t X t − C)
T t=1

• The noise η is correlated as:

D E 1
ηij ηkl = (Cik Cjl + Cil Cjk )
T

• from which one derives:

 
P N λiλj
1 
| det(G)|1/P
D E X X
≈1− 2
.
2T P i=1 j=P +1 (λi − λj )

(and a similar equation for eigenvalues)

J.-P. Bouchaud
Stability of eigenvalues: Correlations

Eigenvalues clearly change: well known correlation crises

J.-P. Bouchaud
Stability of eigenspaces: Correlations

8 meaningful eigenvectors

J.-P. Bouchaud
Stability of eigenspaces: Correlations

P =5

J.-P. Bouchaud
The case of correlation matrices
• Empirical results show a faster decorrelation → real dynamics
of the eigenvectors

• The case of the top eigenvector, in the limit λ1 ≫ λ2, and

for EMA:

– An Ornstein-Uhlenbeck process on the unit sphere around

θ=0

– Explicit solution for the full distribution P (θ) and time

correlations

– det G = cos(θ − θ ′)

• Full characterisation of the dynamics for arbitrary P ? (Ran-

dom rotation of a solid body)

J.-P. Bouchaud