statistical methods: from physics to financev) k. ito and r. stratonovich (~1950-60): calculus with...
TRANSCRIPT
Statistical Methods: from Physics to Finance
Luca DoriaTRIUMF, 2011
Palazzo Mezzanotte, Milano (Italy)
The PIENU Experiment
Precise Measurement of the Branching Ratio: B= e
e
- TRIUMF Based: pion beam thanks to the 500 MeV Cyclotron - Precision Experiment- Uncover Physics beyond the Standard Model
Jun 2011 Luca Doria, TRIUMF 3
The intersections between Physics and Finance/Economy are many!A very personal (and limited) selection of topics:
I) Introduction and Short History II) Some Physics Problems and Probability Distributions - Heavy Tails: Multiple Scattering - Gaussians: Heat Flow and Diffusion
- Brownian Motion
III) Stochastic Calculus IV) Stable Distributions V) Stocks VI) Derivates: Options VII) Summary & Conclusions
Jun 2011 Luca Doria, TRIUMF 4
Is there any logic behind Economics or the Stock Market?
Jun 2011 Luca Doria, TRIUMF 5
Is there any logic behind Economics or the Stock Market?
Jun 2011 Luca Doria, TRIUMF 6
Is there any logic behind Economics or the Stock Market?
Jun 2011 Luca Doria, TRIUMF 7
Historical Overview
Louis Bachelier (1870-1946)The Theory of Speculation
I) R. Brown (1773-1858): Observation of random motion in pollen samples II) L. Bachelier (1870-1946): First attempt to model stocks movements III) A. Einstein (1905): First model of the brownian motion IV) Mandelbrot and Pareto Distributions V) K. Ito and R. Stratonovich (~1950-60): Calculus with random variables VI) Black, Merton, Scholes: Stochastic model for options (~1973) 1997 Nobel Prize in Economics
Now:“Econophysics”: tries to apply physics methods to Finance and Economy:
- Stochastic Processes- Statistical Physics- Agent-based Models- Statistical Analysis- Feynman's Path Integrals
K. Ito (1915-2008)
R. Black, M. ScholesB. Mandelbrot (1924-2010)
Jun 2011 Luca Doria, TRIUMF 8
PART 1:
PHYSICS MODELSand
PROBABILITY DISTRIBUTIONS:EXPONENTIALS vs POWER LAWS
Jun 2011 Luca Doria, TRIUMF 9
Exponentials and Power-Laws Distributions
Probability Distribution: p(x)
Probability:
Moments:
Paxb=∫a
bp xdx
k=∫ x k pxdx
Jun 2011 Luca Doria, TRIUMF 10
Exponentials and Power-Laws Distributions
“Tails”
Exponential (e.g. Gaussian) Power-Law (e.g. Lorentzian)
Probability Distribution: p(x)
Probability:
Moments:
Paxb=∫a
bp xdx
p x= 1
2 2e− x−2
2 2
k=∫ x k pxdx
- All the moments are finite- Central Limit Theorem- Limit of many distributions
Jun 2011 Luca Doria, TRIUMF 11
Exponentials and Power-Laws Distributions
“Tails”
Exponential (e.g. Gaussian) Power-Law (e.g. Lorentzian)
Probability Distribution: p(x)
Probability:
Moments:
Paxb=∫a
bp xdx
p x= 1
2 2e− x−2
2 2
k=∫ x k pxdx
p x = 1
x−22
- All the moments are finite- Central Limit Theorem- Limit of many distributions
- Undefined moments k>0- CLT cannot hold
Jun 2011 Luca Doria, TRIUMF 12
Multiple Scattering
pdf ∝1
0=13.6MeV c p
z x /X0[10.038 ln x /X0]
Used e.g. in:- Detector Physics and Simulation- Medical Physics
Two physics mechanisms: - EM Force - Strong Force
MaterialCharged Particle
θ
Jun 2011 Luca Doria, TRIUMF 13
Heat Transfer and Diffusion
q=−k∇ uFourier's Law
Q=c PuThermodynamics:
If a change in internal energy in a material is given only by heat flux across the boundaries in a space/time region:
c P∫x− xx x
[u , t t −u ,t− t ]d=cP∫x− xx x
∫t−ttt ∂ u
∂d d
k∫t− tt t[∂ u∂ xx x ,−
∂ u∂ xx− x ,]d =k∫x− x
x x
∫t− tt t ∂2u
∂2 d d
∂u∂ t−k
∂2u
∂ x2=0
By energy conservation:
x− xx x ; t− tt t
Jun 2011 Luca Doria, TRIUMF 14
Solution of the Heat Equation
∂u∂ t−k
∂2u
∂ x2=0∂u∂ tx , t −k∇ 2
u x , t =0
∂c∂ tx ,t −k∇2
c x ,t =0
Green's Function (Propagator):
General Solution:
Gx ,t =1
4 kte− x2
4kt
u x , t =∫G x− y ,t f y , t=0dy
Temperature
Concentration
Jun 2011 Luca Doria, TRIUMF 15
Why Physicists are interested in power-law distributions?
Psx=asPx
Random events usually give rise to exponential tails: understood.If there are power law tails, something interesting is happening:
1) Critical State?During phase transitions, correlation functions have power-law behavior.This means e.g. that fluctuations are happening at all scales (see 2)).
2) Scale-free systemPower-law distributions are scale free:
Px =C x−
P sx=C sx−=C s− x−=a s Px
Only one candidate distribution:
Partial proof:
Jun 2011 Luca Doria, TRIUMF 16
Brownian Motion
R. Brown (1773-1858): Observation of random motion in pollen samples
A. Einstein (1905): First quantitative explanation ofthe brownian motion
⟨ x2⟩=2Dt
D= RTN
16 k r
(For spherical particles)
AE estimate : ~6μm / minute @ 17C, N=1023
“… It is hoped that some enquirer may succeed shortly in solving the problem suggested here, which is so important in connection with the theory of heat.”Berne, May 1905 (Received, 11 May 1905).
N= 1
⟨ x2 ⟩RT3 k P
Opportunity to estimate N !
∂u∂ t−D ∂
2u
∂ x2=0
Jun 2011 Luca Doria, TRIUMF 17
A Simple Model for the Brownian Motion: Symmetric Random Steps
t
x
t
h
p N1 t ,ih= 12p N t ,i−1h1
2p N t ,i1h
p N1 t ,ih h2
2∂x2 p N t , ihp N t ,ih∂t p N t , ih t
∂t p t , x=2
2∂x2 p t , x
⟨ x ⟩=0
Var x=⟨ x2⟩−⟨ x ⟩2= 12h2 1
2−h2=h2
t=N t⇒Nh2= th2
t
2 t⇒ h= t
If we want the variance at t being:≈ ≈
p t , x = 1
2t− se
− x−x 02
22 t−s
Jun 2011 Luca Doria, TRIUMF 18
Physics: Summary
I) Random processes often encountered in physics (Probability/Distributions).
II) Connection to the diffusion (or heat) equation.
III) Not all random processes have gaussian tails !
IV) “Fat” tails are not always easy to explain: associated with “complex” systems, phase transitions, critical states, scale invariance, …
In general: something interesting for a physicist !
V) More formal methods needed to treat stochastic processes.
Jun 2011 Luca Doria, TRIUMF 19
PART 2:
STOCHASTIC PROCESSES
Jun 2011 Luca Doria, TRIUMF 20
Brownian Processes
A standard brownian process B_{t}, t>0 is a continuous stochasticprocess in t, with the following properties:
I) B_{0}=0 with probability 1
II) The increments are independent, i.e.: if 0<t0<...tk, then Bt1-Bt0, .... , Btk-Btk-1 are independent.
III) For every s,t, with 0 < s < t, the increments Bt-Bs are distributed like N(0,t-s)
In a non-standard process, N(0,σ2(t-s))
Jun 2011 Luca Doria, TRIUMF 21
Stochastic Processes and Stochastic Calculus
X t =X t0∫t0
tF ,dt∫t 0
tG ,dW
dX t=F t , .dtG t , .dW t
It is not possible to define derivatives in a stochastic process (jumps!)It is possible to define the integral : the Ito (Stratonovich) integral:
General Stochastic Process:
Stochastic Differential:
Jun 2011 Luca Doria, TRIUMF 22
The Ito/Stratonovich Integral
Ito Integral:
Stratonovich Integral:
Riemann-Stieltjes Integral:
∫0tf t dg t =lim n∞∑i=1
nf i g t i1−g ti
i=t i
i=12 ti1t i
If g is smooth, the limit is independent on the choice of τ.It is not true for stochastic processes!
Example:
Example:
∫0tW t dW t=
12W t
2−12W 0t
∫0tW t dW t=
12W t
2−12W 0
Jun 2011 Luca Doria, TRIUMF 23
The Ito's Formula
dY t ={∂tU t , X t F t ∂ X U t , X t 12G2 t ∂X
2 U t , X t}dtG t ∂ X U t , X t dW t
Y t=U t , X t
dX t=F t , .dtG t , .dW tStochastic Process:
...a function of it:
The Differential:
Ito's formula tells us how to calculate derivatives of composite functions. The result is different wrt the classical calculus.
Jun 2011 Luca Doria, TRIUMF 24
The Ito's Formula
dY t ={∂tU t , X t F t ∂ X U t , X t 12G2 t ∂X
2 U t , X t}dtG t ∂ X U t , X t dW t
Y t=U t , X t
dX t=F t , .dtG t , .dW tStochastic Process:
...a function of it:
The Differential:
Ito's formula tells us how to calculate derivatives of composite functions. The result is different wrt the classical calculus.
Jun 2011 Luca Doria, TRIUMF 25
Stochastic Processes
dX t=F t , X t dtG t , X t dW t
dX t=2dW t
dX t= X t dt2X t dW t
dX t=−X t dt2 X t dW t
dX t=− X t dt2dW t Ornstein-Uhlenbeck (Fluctuation around zero)
“Return to the mean” process
Geometric Process
Simple Brownian Motion (L. Bachelier Model)
dX t=dt2dW t Brownian Motion with Drift
Jun 2011 Luca Doria, TRIUMF 26
Summary
I) There is a powerful formalism for treating stochastic processes
II) Used in many scientific areas: - Condensed Matter Physics - Meteorology/Oceanography - Chemistry - Connections with Quantum Mechanics / QFT - Turbulence / Hydrodynamics - ....
III) Financial time series look like random walks: is it possible to apply these ideas to finance or economy?
Jun 2011 Luca Doria, TRIUMF 27
PART 3:
STOCK MARKET
Jun 2011 Luca Doria, TRIUMF 28
Financial Time Series
I) Can we describe financial time series with stochastic processes?
II) If yes, what are the properties of such a process?
III) In the market, not only stocks are exchanged: derivates. Models?
IV) Not only Finance: Economic Systems
Any relation with physics models?
Jun 2011 Luca Doria, TRIUMF 29
Stock Market Time Chart
Goldman-Sachs Group, Inc, NYSE
0 200 400 600 800 1000 1200 1400 1600 1800
0
50
100
150
200
250
300
~today
Jan,1 2005
2007
S(t)
t(days)
Jun 2011 Luca Doria, TRIUMF 30
From Bachelier to Mandelbrot
Jun 2011 Luca Doria, TRIUMF 31
Observables
Nowadays: huge database of (high frequency) financial data. Opportunity to study (“data analysis”) the market and try to “understand” it. What are the important quantities to study? No reference units!
Y t
Z t =Y t t −Y t
Rt =Y t t −Y t
Y t
S t =ln Y t t −ln Y t
Price
Price Change
Return
Log-Change
Jun 2011 Luca Doria, TRIUMF 32
Features:- Not Gaussian! It is not a Brownian Motion.
- Leptokurtic: heavy tails
- Almost no skewness
Distribution of the increments (returns)
Stock: Goldman-Sachs Group, Inc, NYSE
i=∫ x−x0i p xdxK=
4 4−3 0
S=3≈0
Jun 2011 Luca Doria, TRIUMF 33
Correlation Functions
N-point Functions: ⟨ x t ⟩=∫ dx x f x , t
⟨ x t1 x t 2⟩=∫dx1dx2 x1 x 2 f x1 , t1 , x 2 , t 2
⟨ x t1 xt 2... xt n⟩
P [ x t ]=P [ x t]
C t 1 , t 2=⟨ x t 1 x t 2⟩−t 1 t 2=R−2 =t 2−t 1
S f =∫ℜ Re−2i f d =2 2c
12 f tc 2
Stationarity:
Autocorrelation:
Power Spectrum:S f ≈ 1
f 2
S f ≈ 1
f N 0N2
Brownian
Memory: M=∫0∞R d
Finite
Infinite
Undetermined
Jun 2011 Luca Doria, TRIUMF 34
Correlation Functions (Single Stock)
Autocorrelation Function (Coca-Cola Company):Very fast decay (<1day).Arbitrage difficult.No memory
Spectral Density: (Coca-Cola Company):
Compatible with a random walk
S f ≈ 1
f 2
From: Stanley, Mantegna
Jun 2011 Luca Doria, TRIUMF 35
Correlation Functions (Market Index)
Autocorrelation Function (S&P 500):Very fast decay (Corr time <4 min.).Arbitrage difficult.No memory
Spectral Density: (S&P 500):
Compatible with a random walk
S f ≈ 1
f 2
From: Stanley, Mantegna
Jun 2011 Luca Doria, TRIUMF 36
Higher Order Correlations
- Pairwise independence does not imply independent random variables !- Look for high-order functions of the price change: e.g. the VOLATILITY.
LONG RANGE CORRELATION!Power Law with exponent ~0.3
Jun 2011 Luca Doria, TRIUMF 37
Volatility: Summary
I) No “average” volatility existsII) Volatility is difficult to estimate (in the future)III) Clustering (Intermittency?).IV) Memory.V) ARCH / GARCH Models
Jun 2011 Luca Doria, TRIUMF 38
Levy Stable Distributions
Definition: x is a stable random variable if for every positive A,B, exist C,D so that:
Where x1 and x2 are copies of x.
Ax1Bx2 ≈ CxD
Theorem: The sum of independent, identically distributed random variables converge to a stable distribution as the number of iid variables tends to infinity.
Stable (Levy) Distributions
lnk =i k− ⌈ k ⌉ [1−i k
⌈k ⌈tan /2] ≠1
i k− ⌈k ⌉[1i k⌈ k ⌈2 / tan ⌉k ⌈ ] =1
S , ,
02 0 ∈ℜ ∈−1,1Stability Index Scale Factor Shift Factor Asymmetry Factor
Jun 2011 Luca Doria, TRIUMF 39
Power Law (Fat) Tails
limx∞ P Xx≈1sin /2
⌈ x ⌈1≈ 1
⌈ x ⌉1
Stable Distributions have fat tails:
Some known Analytic Forms:
=2
=1 =1
=12
=1
Gaussian
Landau
Levy
Lorentz=1 =0
Jun 2011 Luca Doria, TRIUMF 40
The “stylized” Facts
I) Fat TailsLarge returns follow a fat-tailed distribution P(r) ~ r-α . Generally, α>2, therefore the variance is finite. This questions the use of Levy stable distributions. Truncated Levy pdfs have been proposed, among the many other proposals. The issue is still debated.
II) Asymptotic NormalityWhen the time scale of the returns is increased, the distribution tends to the normal one. Conclusion: the shape of the distribution changes with the time scale and therefore the underlying stochastic process has a non trivial temporal structure.
III) No linear autocorrelations (efficient market, no arbitrage)The autocorrelation falls quickly to 0: after ~15min it is practically =0. There is no memory effect: support of the “efficient market hypothesis”. Over longer timescales, (months) there are some (non conclusive) hints of autocorrelation.
IV) Clustered VolatilityIt is evident that the volatility of the market is not constant in time and there is autocorrelation: high volatility periods cluster in time and large changes are followed by large changes.
Jun 2011 Luca Doria, TRIUMF 41
Stochastic Processes in Finance: Conclusions
Data:- Today, thanks to computers, we have a very large dataset of financial data to study.- It was not possible at the times of Bachelier!
Stochastic Process:- It is not a simple Brownian Motion (Bachelier Model)- Exhibits heavy (fat) tails- Very short correlation time (No memory)- Volatility has longer correlation (Memory?)- …
Models:- Brownian Motion (Not accurate, possibility of negative prices)- Lognormal Process (Better, but not fully accurate)- Ornstein-Uhlenbeck Process- Levy, Truncated Levy, Tsallis, … (Better heavy tails, but debated)- ARCH / GARCH , .... (Powerful, simple: worthed the 2003 Nobel Price!)
Jun 2011 Luca Doria, TRIUMF 42
The Black-Merton-Scholes Model
Black, F. och M. Scholes, 1973, "The Pricing of Options and Corporate Liabilities", Journal of Political Economy, Vol. 81, pp. 637-654.
Black, F., 1989, "How We came Up with the Option Formula", The Journal of Portfolio Management, Vol. 15, pp. 4-8
Hull, J.C., 1997, Options, Futures and Other Derivates, 3rd edition, Prentice Hall
Merton, R.C., 1973, "Theory of Rational Option Pricing", Bell Journal of Economics and Management Science, Vol. 4, pp. 141-183.
Robert C. Merton Myron S. Scholes
The Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel 1997 was awarded jointly to Robert C. Merton and Myron S. Scholes "for a new method to determine the value of derivatives "
Jun 2011 Luca Doria, TRIUMF 43
Derivates: The Options
Definition:
In finance, an option is a derivate financial instrument that establishes a contract between two parties concerning the buying or selling of an asset at a reference price.
“Put” (“Call”) European Options:
The owner of the option has the right (but no obligation) to sell (buy) a certain
good at a certain time for a certain price.
Other jargon terms:
Underlying (Asset): A stock for a simple option (can be any good).
Strike Price: Price for buying or selling the underlying.
Expiration Date: When the option might be used.
American Options: The option can be used at any time before expiration.
Exotic Options: All other kind of options (e.g. asian options, ...).
Jun 2011 Luca Doria, TRIUMF 44
The Black-Merton-Scholes Option Pricing ModelHypoteses: I) The short-term interest rate r is known and constant and it is possible to ask for money at that rate.
II) The strike price X is known and constant
III) The stock price S follows a geometric stochastic process
Where W is a Wiener process, μ is the expected return, σ the volatility.
IV) The stock does not pay dividends
V) No transaction costs and no limits to the short-selling
VI) There is no arbitrage
dS t= S t dt2S t dW t
Jun 2011 Luca Doria, TRIUMF 45
The Black-Merton-Scholes Option Pricing Model
dV=dS− 1dc
dc=[∂ c∂ t 12 ∂2c∂ S 2 2S 2 ]dt ∂ c∂ S S dW
dV=−1 [ ∂ c∂ t 12 ∂2c∂ S 2 2 S2]dt
−1 [ ∂c∂ t 12 ∂2c∂ S 2 2S 2]dt=S− 1 crdt
∂c∂ t=rc−rS ∂c
∂ S− 12 ∂
2c
∂ S 2 2 S 2
Build a portfolio made of stocks and options: V =S− 1c
Consider its variation:
Apply Ito's Lemma to c(t,S):
Substituting:
Equivalence to a risk-free portfolio:(no arbitrage)
After some algebra:
Jun 2011 Luca Doria, TRIUMF 46
Solution of the BMS Equation
c S , t =S N d 2−X e−r T− tN d 2
d 1=ln S / X r 1
2 2T−t
T−t
d 2=d1−T−t
N d i=1,2=∫−∞d i 1
2e−x2 /2dx
The solution of the BSM equation is:
Boundary condition at t=T:(for a call option)
c=max S−K ,0
Jun 2011 Luca Doria, TRIUMF 47
Physics Interpretation of the BMS Equation
∂t c12 2S2∂SS crS ∂S c−rc=0
Diffusion Term
Convection Term
Reaction Term
Example :Dispersion of a pollutant in a river:
- Dispersion : Diffusion Term- Water Flow: Convection Term- Absorption (e.g. by sand) : Reaction Term
Jun 2011 Luca Doria, TRIUMF 48
The BMS equation IS the heat equation !
cS , t ∂t c12 2S2∂SS crS ∂S c−rc=0
u x , ∂u=∂xx u
S=ex = 12 2T−t k= 2r
2
Substitution:
is solution of the BMS equation
is solution of the heat equation
cS , t =exp [−12 k−1 x− 12 k12]u x ,
Jun 2011 Luca Doria, TRIUMF 49
BMS and Montecarlo (N=10): The “Greeks”
= ∂ c∂T
=∂ c∂ S
=∂ c∂ S
=∂ c∂ S
V = ∂ c∂
c=c S
Jun 2011 Luca Doria, TRIUMF 50
BMS and Montecarlo (N-=100): The “Greeks”
c=c S
=∂ c∂ S
V = ∂ c∂
= ∂ c∂T
Jun 2011 Luca Doria, TRIUMF 51
BMS and Montecarlo (N=1000): The “Greeks”
c=c S
=∂ c∂ S
V = ∂ c∂
= ∂ c∂T
Jun 2011 Luca Doria, TRIUMF 52
Summary
I) Nowadays, statistical analysis of financial time series possible:- Probability Distributions (Returns, Rates, …) : Heavy-tailed distributions
- Time correlations, cross-correlations. - Spectral densities - …
II) Modeling of the time series with stochastic models: - Brownian motion (Log-normal process) - (Truncated) Levy processes - Mixed processes (jump-diffusion,...)
III) Derivates: - BMS Model and all the possible generalizations, Delta-Hedging - PDE methods (also numerical) - Montecarlo methods - Feynman Path Integrals !!
Jun 2011 Luca Doria, TRIUMF 53
Conclusion
Do really a physics theory of finance (economics) exists?
(My) answer is: NO:
Often there are only some analogies and the economic system do not resemble closely any physics system (yet).There is no “econophysics” theory, like quantum mechanics or general relativity: it is more like a collection of methods, many of them borrowed from or used by physicists.
BUT:Physicists have skills often very useful if applied in this field:
I) Strong mathematical backgroundII) Strong computational backgroundIII) Construction of modelsIV) Statistical analysis of (huge) datasetsV) ....
Jun 2011 Luca Doria, TRIUMF 54
Physics Models
Stochastic CalculusStatistics
Financial Time SeriesDerivatesEconomic Systems
Brownian MotionDiffusionTurbulenceStatistical Mechanics.....
(Auto)-correlation functionsLognormal ProcessIto's Formula....
Heat Equation“Fat” TailsMC methods.....
Jun 2011 Luca Doria, TRIUMF 55
Ettore Majorana (1906 – 1938 - ?)
Il valore delle Leggi Statistiche nella Fisica e nelle Scienze Sociali(Unpublished)
The Value of the Statistical Laws in Physics and in the Social Sciences
“It is therefore important that the principles of quantum mechanics brought us to recognize the the statistical character of the ultimate laws of the elementary processes.”
“The study of the relationships, true or supposed, between physics and the other sciences, had always been of big interest because of the special influence physics had in modern times on the scientific thinking.”
My “free” translation!
Jun 2011 Luca Doria, TRIUMF 56
Grazie!Tank Yo!
Merci!