Lecture 5:

Mean-Reversion

Marco Avellaneda
G63.2936.001

Spring Semester 2009

 Stationarity/ Non Stationarity

Definition: a stochastic process is stationary if

∀m, ∀(t1 ,..., t m ), ∀A ∈ R n

{( ) } {( ) }
Pr . X t1 , X t2 ,..., X tm ∈ A = Pr . X t1 + h , X t2 + h ,..., X tm + h ∈ A

A stationary process is a process that is ``statistically invariant under

translations’’

Examples: the Ornstein-Uhlembeck process is stationary, Brownian motion

is not.
 The Ornstein-Uhlenbeck process

dX t = κ (m − X t )dt + σdWt

( )m + σ ∫ e
t
−κ (t − s ) − κ (t − s ) − κ (t − u )
Xt = e X s + 1− e dWu
s

t
X t = m + σ ∫ e −κ (t − s )η (s )ds, η (s ) = Gaussian white noise
−∞

Exponentially-weighted moving average of uncorrelated Gaussian

random variables.
 Statistics of the OU process

t +h
η (s )ds ⋅ ∫ e −k (t + h − s ')η (s')ds'
− k (t − s )
t
X t X t+h = σ 2
∫ e
−∞ −∞

t +h
e − k (t − s )e − k (t + h − s ' )δ (s − s')dsds '
t
=σ 2
∫ ∫
− ∞ −∞
t
( ) − k (t + h − s )
=σ 2
∫ e e −k t −s
ds
−∞

e ∫e ( −2k t − s )
t
=σ 2 − kh
ds
−∞

σ 2 e − kh
=
2k

σ2
X t+h − X t
2
=
k
(1 − e ) − kh Structure Function
 Random Walk, Fractional BM
X t = σWt , Wt = Brownian motion Brownian motion (non-stationary)
X t+h − X t = σ 2h X t+h X t = t Structure fn grows linearly
2

t
η (s )ds
Xt = σ ∫ (1 + t − s ) p
p > 1/ 2 Geometric Brownian motion
−∞

∞
σ2 du
h 2 p −1 ∫1 u p (1 + u ) p
X t X t+h =
h

 σ2
 1/ 2 < p < 1
 h 2 p −1
 σ 2 ln( h) Correlations decay like power-laws
X t X t+h ≈ p =1
 h (large h)
 σ2
p >1
 hp
 Autoregressive Models

X 1 , X 2 ,..., X n ,...

X n +1 = a0 + a1 X n + ... + am X n − m +1 + σν n+1 , ν i ~ N (0,1)

 Xn   a1 ... am   a0  σ 
       
Yn =  ... , A =  1 0 ... , A 0 =  0 , Σ =  0 
X   ... 1 0   ...   ... 
 n − m +1       

Yn+1 = A 0 + AYn + Σν n +1 AR-n model corresponds to a

vector AR-1 model
 Structure function: SPY Jan 1996-Jan 2009
Use log prices as time series. Structure function with lags 1 day to 2 yrs

0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
1 27 53 79 105 131 157 183 209 235 261 287 313 339 365 391 417 443 469 495

SPY is highly non stationary, as shown in the chart.

Look for mean-reversion in relative value, i.e. in terms of two or more assets.
 Structure function log (SLB/OIH)
Data: Apr 2006 to Feb 2009
0.014

0.012

0.01

0.008
S(x)

0.006

0.004

0.002

0
1 28 55 82 109 136 163 190 217 244 271 298 325 352 379 406 433 460 487

Saturation days

OIH: Oil Services ETF, SLB: Schlumberger-Doll Research

 Structure Function: long-short
equal dollar weighted SLB-OIH
0.025

0.02

0.015
S(x)

0.01

0.005

0
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484

days

Pn +1 = Pn × (1 + Rslb − Roih ), X n = ln Pn
 Structure Function for Beta-Neutral
long-short portfolio SLB-Beta*OIH
0.035
0.03
0.025
0.02
S(x)

0.015
0.01
0.005
0
1 36 71 106 141 176 211 246 281 316 351 386 421 456 491
days

Pn +1 = Pn × (1 + Rslb − β 60 d ⋅ Roih ), X n = ln Pn
 Structure Function log (GENZ/IBB)

0.07
0.06
0.05
0.04
S(x)

0.03
0.02
0.01
0
1 35 69 103 137 171 205 239 273 307 341 375 409 443 477
days
 Structure function ln (DNA/GENZ)
0.18
0.16
0.14
0.12
0.1
S(x)

0.08
0.06
0.04
0.02
0
1 29 57 85 113 141 169 197 225 253 281 309 337 365 393 421 449 477

days

DNA: Genentech Inc. Mean-reversion: large negative

GENZ; Genzyme Corp. curvature here.
 Structure Fn for Beta-Neutral GENZ-DNA
Spread
0.14
0.12

0.1
0.08
S(x)

0.06

0.04
0.02
0
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484

days

Poor reversion for the beta adjusted pair. Beta is low ~ 0.30
Systematic Approach for looking for
MR in Equities
Look for stock returns devoid of explanatory factors, and analyze the
corresponding residuals as stochastic processes.

m
Rt = ∑ β k Fkt + ε t Econometric factor model
k =1

t
Xt = X 0 + ∑εs View residuals as increments of a
s =1
process that will be estimated

dS (t ) m dP (t )
= ∑ β k k + dX (t )
Continuous-time model for evolution
S (t ) k =1 Pk (t ) of stock price
 Interpretation of the model
The factors are either

A. eigenportfolios corresponding to significant eigenvalues

of the market

B. industry ETF, or portfolios of ETFs

Questions of interest:

Can residuals be fitted to (increments of) OU processes or

other MR processes?

If so, what is the typical correlation time-scale?

Estimation of Ornstein-Uhlenbeck models
t + ∆t
( )
X t + ∆t = e −k∆t X t + m 1 − e −k∆t + σ ∫ e − k (t − s )
dWs
t

  − − 2 k∆t

X n +1 = aX n + b +ν n +1 {ν n } i.i.d. N  0, σ 2 
1 e
 

  2k 

a = SLOPE(( X n−l ,..., X n ); ( X n−l −1 ,..., X n −1 )),

b = INTERCEPT (( X n −l ,..., X n ); ( X n −l −1 ,..., X n −1 ))
1 1 b STDEV ( X n+1 − aX n − b ) 1 1
k = ln , m= , σ= 2 ln 
∆t  a  1− a 1− a 2 ∆t  a 
 Portfolio Strategy
Q1 , Q2 ,...., QN $ invested in different stocks (long or short)
S1 , S 2 ,....., S N dividend - adjusted prices

N
dSi  N 
dΠ = ∑ Qi −  ∑ Qi rdt (neglect transaction costs)
i =1 S i  i =1 
N
 m dPk   N 
= ∑ Qi  ∑ β ik + dX i  −  ∑ Qi rdt
i =1  k =1 Pk   i =1 
N
 N
m
 dPk  N 
= ∑ Qi dX i + ∑  ∑ Qi β ik  −  ∑ Qi rdt
i =1 k =1  i =1  Pk  i =1 

∑Q β
i =1
i ik : net dollar - beta exposure along factor k

 N 
 ∑ Qi  : net dollar exposure of portfolio
 i =1 
 Market-Neutral Portfolio

Assume dX i = ki (m − X i )dt + σ i dWi {dWi }iN=1 uncorrelated

N
 N 
dΠ = ∑ Qi dX i −  ∑ Qi rdt
i =1  i =1 
N
 N 
= ∑ Qi (ki (m − X i )dt + σ i dWi ) −  ∑ Qi  rdt
i =1  i =1 
N N
= ∑ Qi (ki (m − X i ) − r )dt + ∑ Qiσ i dWi
i =1 i =1

∴
N
E (dΠ | X ) = ∑ Qi (ki (m − X i ) − r )dt
i =1
N
Var (dΠ | X ) = ∑ Q 2 iσ 2 i dt
i =1
 Mean-Variance Optimal Portfolio
 1 2 2 µi
max Q  ∑ Qi µi − ∑ Qi σi  ∴ Qi = λ
 i 2 λ i  σ i2
(if r = 0, or ∑ Qi = 0)
ki2 (m − X i ) ki (m − X i )
2
dΠ = λ ∑ dt + λ ∑ dWi
i σ i2 i σi
ki 2 k m − Xi
dΠ = λ ∑ ξ i dt + λ ∑ i ξi dWi ξi = 2 ki
i 2 i 2 σi

λN  ∑ λ2 N  ∑
 k   k 
i
 i

dΠ = i
 dt ; (dΠ )2 − dΠ
2
= 
i
dt
2  N  2  N 
   

 ∑ ki 
N   Nk
Annualized Sharpe Ratio = ⋅ i =
2  N  2
 
 Statistics on the Estimated OU
Parameters
ETF Abs(Alpha) Beta Kappa Reversion days EquiVol Abs(m)
HHH 0.20% 0.69 38 7 4% 3.3%
IYR 0.11% 0.90 39 6 2% 1.8%
IYT 0.18% 0.97 41 6 4% 3.0%
RKH 0.10% 0.98 39 6 2% 1.7%
RTH 0.17% 1.02 39 6 3% 2.7%
SMH 0.19% 1.01 40 6 4% 3.2%
UTH 0.09% 0.81 42 6 2% 1.4%
XLF 0.11% 0.83 42 6 2% 1.8%
XLI 0.15% 1.15 42 6 3% 2.4%
XLK 0.17% 1.03 42 6 3% 2.7%
XLP 0.12% 1.01 42 6 2% 2.0%
XLV 0.14% 1.05 38 7 3% 2.5%
XLY 0.16% 1.03 39 6 3% 2.5%
Total 0.15% 0.96 40 6 3% 2.4%

Average over 2006-2007

Mean reversion days: how long does it take
to converge?

T_{days}=252/k

Days
Max 30
75 % 11.4
Median 7.5
25 % 4.9
Min 0.5
Fast days 36% Fast days : Percentage of faster mean
reversion than 7 days