advanced financial models
DESCRIPTION
This slide set is a work in progress and is embedded in my Principles of Finance course, which is also a work in progress, that I teach to computer scientists and engineers http://financefortechies.weebly.com/TRANSCRIPT
Advanced Financial Models under construc2on
Learning Objec-ves
¨ Lognormal Distribu-ons ¨ Rela-ons between
¤ Normal & lognormal n Pdfs n Sta-s-cs
¤ Simple and natural log rates
2
Hypotheses and Models
¨ Explana-ons of phenomenon ¤ Hypothesis
n A proposed explana-on for a phenomena
¤ Law n Statement of a cause and effect
without explana-on n Newton’s law of gravity
¤ Theory n A well-‐established explana-on for
a phenomenon n Einstein’s theory of gravity
¨ A model is a mathema-cal or physical representa-on of a phenomenon ¤ The “Bohr atomic model” ¤ Newton’s inverse square law of
gravity
¤ Einstein’s Theory of General Rela-vity
3
221
rmmGF ⋅
⋅=
SPX Daily Ln Rate Histogram: Zoom 4
SPX Daily Ln Rate Histogram: More Zoom 5
Again this histogram includes daily return rates from 1950 <-4.5% should happen less than once in a thousand years, but there have been 31 such days since 1950 or about once every two years -22.9% day should not have happened (Oct 19, 1987)
SPX Daily Ln Rate: August – December 2008 6
SPX Daily Ln Rate: Mean 7
-‐350%
-‐250%
-‐150%
-‐50%
50%
150%
250%
1/5/51 11/9/57 9/13/64 7/19/71 5/23/78 3/27/85 1/30/92 12/4/98 10/8/05
Annualized mean 22 day annualized tailing mean 252 day annualized tailing mean Long term annualized tailing mean
SPX Daily Ln Rate: Mean 8
-‐80%
-‐60%
-‐40%
-‐20%
0%
20%
40%
60%
80%
1/2/90 3/12/92 5/21/94 7/29/96 10/7/98 12/15/00 2/23/03 5/3/05 7/12/07 9/19/09
Zoom in on Annualized mean 252 day annualized tailing mean Long term annualized tailing mean
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1/3/1950 3/22/1958 6/8/1966 8/25/1974 11/11/1982 1/28/1991 4/16/1999 7/3/2007
SPX Daily Ln Rate: Standard Devia-on 9
Annualized standard devia-ons (‘vola-lity’) 22 day annualized trailing vola-lity 252 day annualized trailing vola-lity Long term annualized trailing vola-lity
0%
5%
10%
15%
20%
25%
30%
1/2/1990 9/28/1992 6/25/1995 3/21/1998 12/15/2000 9/11/2003 6/7/2006
SPX Daily Ln Rate: Standard Devia-on 10
Zoom in on Annualized standard devia-ons (‘vola-lity’) 252 day annualized trailing vola-lity Long term annualized trailing vola-lity
SPX Daily Ln Rate: Autocorrela-on Cluster 11
SPX Daily Ln Rate: Autocorrelogram 12
Natural log daily return rates for SPX, v 1950 – 2011 15471 days Rates do look rather uncorrelated
SPX Daily Ln Rate: Autocorrelogram 13
Natural log daily de-‐trended squares of return rates (variance) for SPX, (v-‐u)2 1950 – 2011 15471 days There is some posi-ve autocorrela-on (persistence) Might even be greater persistence over shorter periods
SPX Daily Ln Rate: Histogram of Annualized Daily Variance
14
Histogram of Annualized Daily Variance
SPX: Annual Accumula-on of Daily Returns 15
10,000 annual sums of 252 day (1 year) con-guous return rates randomly selected from 1950 to 2011 This histogram doesn’t look normal at all as the addi-ve CLT would indicate So the rates are not IID/ FV
SPX: Ln Rate Q-‐Q Plot 16
A Q-‐Q plot compares the measured rates to ideal normal rates from measured mean and variance
Natural Log Rate – More Tests
¨ Jarque Bera normality test ¤ JB is a Chi Squared sta-s-c with 2 dof ¤ Normality via chi squared considera-on of s
skew, S, and kurtosis, K, the 3rd and 4th moments of distribu-on which measure asymmetry
17
440,657 4
3)(29.06001.056216
15471
43)(KS
6nJB
22
22
=
⎟⎟⎠
⎞⎜⎜⎝
⎛ −+−=
⎟⎟⎠
⎞⎜⎜⎝
⎛ −+=
JB (χ2
statistic)
If normality is rejected, what is the probability of a rejection error
0.0000 100.00%4.6051 10.00%5.9914 5.00%9.2103 1.00%10.0000 0.67%15.0000 0.06%20.0000 0.00%25.0000 0.00%30.0000 0.00%35.0000 0.00%40.0000 0.00%45.0000 0.00%50.0000 0.00%
So there is ~0% probability of incorrectly rejec-ng the normal hypothesis
Natural Log Rate – Tests For Normality 18
Stock Return Rate Summary
¨ Historical stock return rates, r and v, are characterized by ¤ Leptokurtosis
n Fat or heavy tails: more extreme events than ‘normal’ n More return rates near the mean than ‘normal’
¤ Nega-ve skew n More extreme downside events than upside
¨ Dependence in return rate vola-lity ¤ Rate vola-lity clustering, short term persistence then reversion to mean
¨ Less frequent sampling e.g., weekly and monthly would show some smoothing, but s-ll not normal ¤ However, quarterly or annual sampling would ignore important rate of
return informa-on
19
Lognormal Pdf 20
The lognormal pdf is asymmetric, is not nega-ve, over -me the mean, mode, and median drii further apart, and the distribu-on skews more posi-vely.
Lognormal Pdf 21
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
Mode Median Mean (expected)
22
( )
( )
∞>>⋅⋅
=
−
⋅
−−
x 0
eπ2σx
1σ μ, |x f
1s,uNL~r
σ2μ)(lnx
x
2
2
2
( )
( )
∞>>∞⋅
= ⋅
−−
x -‐
eπ2σ
1σ μ, |x f
s,uN~v
σ2μ)(x
x
2
2
2
u is mean, median, and mode The parameters is the normal pdf above are also the sta-s-cs – the mean and variance
The mean, mode, and median are all different The parameters is the lognormal pdf are the same as for the normal pdf, but they are not the sta-s-cs, not the mean or variance
¨ Why simple returns can’t really be normal ¤ Simple returns are compounded over -me increments, but normal random variables are mul-plied
¤ (1+r)n ¤ u·∙n
23
( )r1lnv +=
...3r
2rr)r1ln(v
)1x( ...3x
2xx)x1ln(
32
32
−+−=+=
−≠−+−=+
Variance of Simple and Log Returns 24
[ ] [ ][ ] [ ]
( )
( )
( )
[ ]( ) ( )
1er1E
1ee
1ee
1ee
ee
ee
xExE
dr1VarrVar
2
2
2
2
2
22
22
222
s2
s
2
2su
s2su2
ssu2
s2us2u2
2
2su
2s2u2
22
2
−⋅+=
−⋅⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛=
−⋅=
−=
−=
⎟⎟⎠
⎞⎜⎜⎝
⎛−=
−=
=+=
⎟⎟⎠
⎞⎜⎜⎝
⎛+
⎟⎟⎠
⎞⎜⎜⎝
⎛+⋅
+⋅
+⋅+⋅
+⋅
+⋅
[ ]( ) ( )( ) ( )
( )( )
( )
1)s 1,(a sd
1ed
d1e
d1lns
1)(a 1 a1
a1d1lns
1ea1
1er1E d
22
s2
2s
22
2
2
22
s2
s22
2
2
2
2
<<<<≈≈
−≈
+≈
+≈
<<≈+
⎟⎟⎠
⎞⎜⎜⎝
⎛
++=
−⋅+=
−⋅+=
( )[ ]
[ ]
[ ][ ] 2
2
22
s2u22
2su
2skukk
2
v
e xE
e xE
e xE
su,NL~
er1X
⋅+⋅
+
⋅+⋅
=
=
=
=+=
Variance of Simple and Log Returns 25
Future Value Factor: 1+r = ev
[ ] ( ) 1eer1var22 ssu2 −⋅=+ +⋅
[ ] *
2
u2s
ueer1E ≡=+
+
[ ] uer1M =+
26
( ) ( )
[ ]
[ ] [ ][ ] ( )[ ]
[ ] ondistributi normal log for Median er1MM[x]
ondistributi lognormal for moment 2 er1E xE
ondistributi normal log for (mean) moment 1 er1E xE
ondistributi normal log for moment k exE
su,NL~ er1
u
nds2u222
st2s
u
th2sk
ukk
2v
2
2
22
=+=
=+=
=+=
=
=+
⋅+⋅
+
⋅+⋅
Central Limit Theorem 27
( )
( )2n
n
n
1ii
n
1ii0n
s,uN~ny
uny
n
v
vSln)Sln()Sln(
→=
=Δ=−
∑
∑
=
=( )
1)x,x(NL~r
g1f)r(1
fr1SS
n1
n
n1
n
1ii
n
n
1ii
0
n
−
+→=⎥⎦
⎤⎢⎣
⎡+
≡+=
∏
∏
=
=
Assume that n is large and r and v are IID/FV
28
( ) ( )
( )
∑∏
∑∏
∑∏
==
==
==
=⎥⎦
⎤⎢⎣
⎡
+=⎥⎦
⎤⎢⎣
⎡
+=⎥⎦
⎤⎢⎣
⎡+
n
1ii
n
1i
v
n
1ii
n
1i
v
n
1ii
n
1ii
veln
r1lneln
r1lnr1ln
i
i
n21
i
vvv0
n
1i
v0n
n210
n
1ii0n
e ...e e S
eS S
)r(1....)r(1)r(1 S
)r(1S S
⋅⋅⋅⋅=
⋅=
+⋅⋅+⋅+⋅=
+⋅=
∏
∏
=
=
( ) ( )
( )
( ) ( )
( ) n210
n
1ii0n
n210
n
1ii0n
v...vvSln
vSln Sln
)rln(1....)rln(1)rln(1Sln
)rln(1Sln Sln
++++=
+=
+++++++=
++=
∑
∑
=
=
Mean Natural Log Return Rates 29
Example: v is distributed uniformly from -‐10% to +20% Average of sums of vi are normal (sum of n rates)
( )2n
1ii
s,uN~n
v∑=
Simple Future Value Factors 30
Example: r is distributed uniformly from -‐10% to +20%
)x,x(NL~)r(1fn1
n
1ii
n1
n ⎥⎦
⎤⎢⎣
⎡+= ∏
=
Simple Future Value Factors 31
( ) )x,x(NL~r1fn
1iin ∏
=
+=
Example: r is distributed uniformly from -‐10% to +20% f is distributed lognormal
We did plot a histogram of natural return rates, v, for the SPX. It did have the general appearance of normality. But a Levy stable seems like a bemer fit, but has disadvantages. However, the typical assump-on in finance is that v is normally distributed which has a number of advantages. One advantage is that the loca-on sta-s-cs are iden-cal – mode, median, and mean – it’s a symmetric
32
v)r1(ln e)r1(
v)r1(ln e)r1(v
iiv
ii
=+=+
=+=+
For stocks or other financial assets, so far there has been no assump-on on the distribu-ons of v and r other than being IID/FV But the rela-onship between r and v has been defined as
( ) ( )2s,uN~r1lnv +=
33
( ) ( )( ) ( ) ( )
( ) ( ) 1s,uNL1e~r
s,uNLe~r1
s,uN~r1lnv
2s,uN
2s,uN
2
2
2
−≡−
≡+
+=
Another advantage is the normal distribu-on scale linearly in -me. The mean driis to the right while the variance increases.
( )2sn,unN~vn ⋅⋅⋅
Another advantage is the normal distribu-on scale linearly in -me. The mean driis to the right while the variance increases. Yet another advantage is the rela-on between the normal and lognormal distribu-ons is similar to the rela-on between the na-ral log rate and simple rate
Therefore the simple rate, r, is lognormal under assump-on that the natural log rate is lognormal
34
35
Natural Log Rate Autocorrelogram 36
Natural log daily absolute return rates for SPX, |v| Daily range 1950 – 2011 15471 days
Common PDFs in Finance
¨ Gaussian / Normal ¤ IID / FV, two parameters ¤ CLT for sums of IID/FV random
variables ¤ Special case Levy stable and ellip-c
distribu-ons
¨ Ellip-c ¤ IID / FV, two parameters ¤ unimodal, no skew, no kurtosis other
than Gaussian case ¤ Linear correla-on defines linear
dependence ¤ Used in MPT and CAPM ¤ Includes Gauss, Cauchy, t-‐distr,
Laplace, symmetric Levy Stable
¨ Lognormal ¤ IID / FV ¤ CLT for products of IID/FV random
variables ¤ Posi-ve ¤ Mode, median, mean non-‐coincident
¨ Levy stable ¤ IID, not generally FV, 4 parameters
¤ Unimodal, skew, kurtosis other than Gaussian case
¤ Central limit theorem for IID and stable but not FV random variables converges to a Levy stable distribu-on
¤ Includes Gaussian, Cauchy, Levy
37
More on Covar & Corre 38
[ ] [ ] [ ] [ ]
yExEyxEy,xCov ⋅−⋅=
Monthly Idealized PDFs From SPX History 39
ln(1+r) = v Normally distributed N(u,s2) u and s2 are normal pdf parameters and sta-s-cs -‐ mean and variance
(1+r)= ev Lognormally distributed NL(u,s2)
Same pdf parameters, but different mean and
variance
Monthly Idealized PDFs From SPX History 40
Future value factor, (1+r) = ev lei shiied by -‐1, r
Return Rate PDFs: Sta-s-cs Increase With Time 41
-‐75% -‐50% -‐25% 0% 25% 50% 75% 100% 125% 150% 175% 200% 225% 250% 275% 300%
Natural log rates (v) are assumed normal. The mean and variance of a normal distribu-on scale linear in -me
The future value factors (1+r) are assumed log normally distributed. The mean and variance do not scale linearly in -me.
42
Future Value Factor: 1+r = ev
[ ] ( ) 1eer1var22 ssu2 −⋅=+ +⋅
[ ] *
2
u2s
ueer1E ≡=+
+
[ ] uer1M =+
43
[ ] [ ] [ ] [ ]
( )
( )
( )
[ ]( ) ( )
1er1E
1ee
1ee
ondistributi normal logof Variance 1ee
ee
ee
xExEr1VarrVar
2
2
2
2
2
22
22
222
s2
s
2
2s
u
s2s
u2
ssu2
s2us2u2
2
2s
u2s2
u2
22
−⋅+=
−⋅⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛=
−⋅=
−=
−=
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛−=
−=+=
⎟⎟⎠
⎞⎜⎜⎝
⎛+
⎟⎟⎠
⎞⎜⎜⎝
⎛+⋅
+⋅
+⋅+⋅
+⋅
+⋅
44
-‐1.25 -‐1.00 -‐0.75 -‐0.50 -‐0.25 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50
Idealized PDFs for 36 Months 45
Forecast 36 month natural log rate of return normally distributed N(36·∙u, 36·∙s2)
Forecast 36 month future value factor lognormally distributed
Forecast 36 month simple rate of return lognormally distributed
Lognormal Distribu-on 46
( ) ( )
[ ]
[ ] [ ][ ] ( )[ ]
[ ] ondistributi normal log for Median er1MM[x]
ondistributi lognormal for moment 2 er1E xE
ondistributi normal log for (mean) moment 1 er1E xE
ondistributi normal log for moment k exE
su,NL~ er1
u
nds2u222
st2s
u
th2sk
ukk
2v
2
2
22
=+=
=+=
=+=
=
=+
⋅+⋅
+
⋅+⋅
( ) ( )
( )
( ) ( )
( ) n210
n
1ii0n
n210
n
1ii0n
v...vvSln
vSln Sln
)rln(1....)rln(1)rln(1Sln
)rln(1Sln Sln
++++=
+=
+++++++=
++=
∑
∑
=
=
SSS
r
)r1(SS
rate return Simple
1i
1iii
i1ii
−
−
−
−=
+⋅=
( ) ( )
)r1ln(
SlnSlnSS
lnv
eSS
rate return log Natural
i
1ii1i
ii
v1ii
i
+=
−=⎟⎟⎠
⎞⎜⎜⎝
⎛=
⋅=
−−
−
Lognormal Distribu-on 47
( ) ( )
[ ]
[ ] [ ][ ] ( )[ ]
[ ] ondistributi normal log for Median er1MM[x]
ondistributi lognormal for moment 2 er1E xE
ondistributi normal log for (mean) moment 1 er1E xE
ondistributi normal log for moment k exE
su,NL~ er1
u
nds2u222
st2s
u
th2sk
ukk
2v
2
2
22
=+=
=+=
=+=
=
=+
⋅+⋅
+
⋅+⋅
GARCH Time Series
¨ Similar to historic vola-lity ¤ Simple condi-onal dependence in the second moment (vola-lity)
n Vola-lity clustering or persistence
¨ The GARCH vola-lity has three contribu-ons ¤ Long term average vola-lity, s2, so there’s a reversion of the mean ¤ Short term dependence on recent square of return rate, v2 ¤ Short term dependence on recent Garch vola-lity, h
¨ To Do n Is there a probability distribu-on? Maybe not n Plot the resul-ng rates and look for fat tails n So it looks good historically, but how can it be used in decision making ?
48
GARCH Time Series
¨ The GARCH(1,1) vola-lity model with the natural log rate process model vola-lity has three contribu-ons
49
( )
0βλ,α,1βαγ
β ,α ,γ :weights
hβvαsβα1
hβvαsγh
zh uv
1i21i
2
1i21i
2i
iii
>
=++
⋅+⋅+⋅−−=
⋅+⋅+⋅=
⋅+=
−−
−−
The Gaussian rate process is vi = u + s ·∙zi s is the (tradi-onal) long term average standard devia-on z is the standard normal random variable h is the Garch variance v is the nat log return rates Example: α = .85 , β = .1 , γ = .05
GARCH Time Series 50
Single simulated GARCH(1,1) vola-lity for 15,461 days
GARCH Time Series 51
GARCH Time Series 52
Adendum: Nat Log & Exp 53
( ) ( ) ( )
y+xyxyxyx
32
32
x
)xln(
e = e e )(e = e
...31x
21x1x)xln(
)1x( ...3x
2xx)x1ln(
x1)xln(
dxd
)yln()xln(yxln
)yln()xln()yxln(x)eln(
)0x( xe
⋅
−−
+−
−−=
−≠−+−=+
=
−=⎟⎟⎠
⎞⎜⎜⎝
⎛
+=⋅
=
>=
⋅
Addendum 54
RatePeriodic mean
Annual mean
Periodic standard deviation
Annual standard deviation
a α
g γ
vi u µ s σ
d = Var(r) = Var(1+r)
ri d δ
Addendum 55
dwσdt2σμ dln(S)2
* ⋅+⋅⎟⎟⎠
⎞⎜⎜⎝
⎛−=( )
( )Tσ
Tσ.5rKSln
d
Tσ
Tσ.5rKSln
d
2*0
2
2*0
1
⋅
⋅⋅−+⎟⎠⎞
⎜⎝⎛
=
⋅
⋅⋅++⎟⎠⎞
⎜⎝⎛
=
[ ] ( ) 1eer1var22 ssu22 −⋅=+=δ +⋅
[ ] *
2
u2s
ueer1E ≡=+
+
( ) ( )[ ]
[ ]
tsz1ii
s,0N
1i
i
v1ii
2i
i1i-‐i
eSS
e~SS
eSS
s,0N~v
vSln Sln
2
i
Δ⋅⋅−
−
−
⋅=
⋅=
+=
tszi
itsz
i1ii
i)rln(1v
ii
ii
i
ii
er
)r(1e
)r(1SS)r(1ee
tszv
)rln(1 v
Δ⋅⋅
Δ⋅⋅
−
+
=
+=
+⋅=
+==
Δ⋅⋅=
+=
Addendum 56
( )[ ] [ ]
( )
( ) ( )
−+−=+
+Δ⋅⋅
+Δ⋅⋅
+Δ⋅⋅+==
−+−=+
++++=
=⋅δ⋅+
=+≡δ
⋅δ⋅+⋅=
⋅δ⋅==
Δ⋅⋅
Δ⋅⋅
3r
2rr)r1ln(
...6
tsz2
tsztsz1ee
3x
2xx)x1ln(
6x
2xx1e
e ΔtZ1
rSDr)(1SD
ΔtZ1SS
ΔtZ?r
3i
2i
ii
3
i
2
ii
tszv
32
32x
tszi
itt
ii
ii
1-‐ii
Not yet ready to related normal and lognormal distribu-ons. Need lognormal sta-s-cs and Ito’s Lemma Normal
Natural log rates Natural log prices
Lognormal Simple rates Future value factors Prices
Levy Stable Distribu-on
¤ Bemer fits historical rates of return n Can model Leptokurtosis and skew
n Constant parameters
n Generalized Central Limit Theorem n Normal distribu-on is a special case n Problems included
n Infinite variance n Variance cant be used as a measure of risk or vola-lity n CAPM, MPT, B-‐S n PDF models not applicable
n Generally no analy-c representa-on ¤ To Do
n Fit data to a distribu-on and graph n Why does FMH without IID invoke this model n How does it relate to power law model (Has an α > 2 ?)
57
Levy Stable Distribu-ons 58
[ ]( )( ) parameter location ,μ
parameter scale 0,cparameter skewness 1,1β
parameter stability (0,2]αParameters
∞∞−∈
∞∈
−∈
∈
undefined otherwise2,α when 0 :kurtosis excessundefined otherwise2,α when 0:skew
infinite otherwise2,α when c2 :variance
undefined otherwise 1,α whenμ :mean2
=
=
=⋅
>
Levy Stable Distribu-ons
¨ DJIA: α = 1.5958 β = -‐.0995 µ = .0002 σ = .0056 ¤ (5/26/1896 – 1/16/2004 daily)
¨ SPX = α = 1.6735 β = .1064 µ = -‐.0002 σ = .0049 ¤ (3/1/1950 – 5/27.2005 daily)
¨ Only three sets of parameters result in closed form ¤ Gaussian
n Actually two of the four parameters are zero (?) or reduced to different 2 ? n Finite variance
¤ Levy ¤ Cauchy
59
[ ]( )( ) parameter location ,μ
parameter scale 0,cparameter skewness 1,1β
parameter stability (0,2]αParameters
∞∞−∈
∞∈
−∈
∈
Power Law 60
Power Law Method
¨ Coopera-on, herding, cri-cality ¤ How Nature Works – Bak ¤ Ubiquity – Buchanan
¨ The ubiquity of scale-‐free behavior and self-‐organiza-on in Nature led Bak, Tang and Wiesenfeld (BTW) to coin the term Self-‐Organized Cri-cality (SOC) to explain the emergence of complexity in dynamical systems with many interac-ng degrees of freedom without the presence of any external agent ; SOC was devised to be a sort of supergeneral theory of complexity.
61
Power Law
¨ Confusion based on Fractal Market Hypothesis: Is it stable or power law?? ¨ Hurst soiware shows a random series to be persistent ??
¨ Hurst exponent ¤ 0.5 is Brownian t1/2 √t
¤ 0 < H < 0.5 : an--‐persistent, mean rever-ng ¤ .5 < H ≤ 1.0 : persistent
¨ Stability parameter ¤ α = 1 / H, example Gaussian: α = 2, H = .5
¨ Correla-on (?) C = 22H-‐1 – 1
¨ Example ¤ SPX: 3/1/1950 – 5/27/2005 daily
n α = 1.6735 β = .1064 µ = -‐.0002 σ = .0049 n H = .5976 C= .1448
¤ SPX: 1/3/1950 – 6/24/2011 daily n H= .562 α= 1.779 C= .090
62
Reference: Nat Log & Exp 63
)1x( ...3x
2xx)x1ln(
x1)xln(
dxd
)yln()xln(yx
ln
)yln()xln()yxln(x)eln(
)0x( xe
32
x
)xln(
−≠−+−=+
=
−=⎟⎟⎠
⎞⎜⎜⎝
⎛
+=⋅
=
>= ( ) ( ) ( )
++++=
⋅
−−
+−
−−=
⋅
6x
2xx1e
e = e e )(e = e
...31x
21x1x)xln(
32x
y+xyxyxyx
32
RatePeriodic mean
Annual mean
Annual standard deviation
Period standard deviation
Rate pdf
a α
g γ
vi u µ s σ Normal
d = SD(r) = SD(1+r)
ri d δ Log normal
Related Concepts
¨ Expected Rate of Return On Equity ¤ CAPM requires that the return rate is normally distributed with a trend ¤ Ordinary least squares
¨ Theore-cal basis for r being an independent random variable ¤ Efficient Market Hypothesis
¨ Theore-cal basis for r being an independent random variable with a trend ¤ Ra-onal Market Hypothesis
64
( ) ( )( )
( )FMFEE
iE1ii
E1ii
rrβr k r zsr1SS
r1SSE
−⋅+==
⋅++⋅=
+⋅=
−
−
Geometric Brownian Mo-on 65
( ) ( )[ ]
[ ]
tsz1ii
s,0N
1i
i
v1ii
2i
i1i-‐i
eSS
e~SS
eSS
s,0N~v
vSln Sln
2
i
Δ⋅⋅−
−
−
⋅=
⋅=
+=
tszi
itsz
i1ii
i)rln(1v
ii
ii
i
ii
er
)r(1e
)r(1SS)r(1ee
tszv
)rln(1 v
Δ⋅⋅
Δ⋅⋅
−
+
=
+=
+⋅=
+==
Δ⋅⋅=
+=
dwσdt2σμ dln(S)2
* ⋅+⋅⎟⎟⎠
⎞⎜⎜⎝
⎛−=( )
( )Tσ
Tσ.5rKSln
d
Tσ
Tσ.5rKSln
d
2*0
2
2*0
1
⋅
⋅⋅−+⎟⎠⎞
⎜⎝⎛
=
⋅
⋅⋅++⎟⎠⎞
⎜⎝⎛
=
[ ] ( ) 1eer1var22 ssu22 −⋅=+=δ +⋅
[ ] *
2
u2s
ueer1E ≡=+
+
Geometric Brownian Mo-on 66
( )[ ] [ ]
( )
( ) ( )...
6tsz
2tsztsz1ee
3x
2xx)x1ln(
6x
2xx1e
e ΔtZ1
rSDr)(1SD
ΔtZ1SS
ΔtZ?r
3
i
2
ii
tszv
32
32x
tszi
itt
ii
ii
1-‐ii
+Δ⋅⋅
+Δ⋅⋅
+Δ⋅⋅+==
−+−=+
++++=
=⋅δ⋅+
=+≡δ
⋅δ⋅+⋅=
⋅δ⋅==
Δ⋅⋅
Δ⋅⋅
Not yet ready to related normal and lognormal distribu-ons. Need lognormal sta-s-cs and Ito’s Lemma Normal
Natural log rates Natural log prices
Lognormal Simple rates Future value factors Prices
u, s µ, σ r, d α, δ g, γ,
Alterna-ves
¨ Fat Tail Models ¤ Power law not exponen-al tails ¤ Leptokurtosis, finite variance ? ¤ Examples
n Student t – no skew n Levy stable – skew
¨ Non IID Models – non-‐sta-onary process ¤ Correla-on in rate vola-lity, but not in rate, so s-ll ‘unpredictable’
ARCH models ¤ Used with normal or other distribu-on
67
Addendum 68
( ) ( )
( ) ( ) ( )SlnSlnSln-‐Sln
eSS
Δtσz Δt2σμ 1SS
Δtσz ΔtμSlnSln
1-‐ii
it
2
ii
i1-‐ii
i1-‐ii
tt
Δtσz Δt2σμ
1tt
t
2
tt
ttt
Δ≠Δ=
=
⎥⎦
⎤⎢⎣
⎡⋅⋅+⋅⎟⎟
⎠
⎞⎜⎜⎝
⎛++⋅=
⋅⋅+⋅+=
⋅⋅+⋅⎟⎟⎠
⎞⎜⎜⎝
⎛+
⋅−
( ) ( ) nszunSlnSln 1i-‐1-‐ni ⋅⋅+⋅+=+( ) ( )[ ]
[ ]
tsztu1ii
s,uN
1i
i
v1ii
2i
i1i-‐i
eSS
e~SS
eSS
s,uN~v
vSln Sln
2
i
Δ⋅⋅+⋅−
−
−
⋅=
⋅=
+=
69
[ ][ ]
[ ]
[ ]2εii1i-‐i
2εii1i-‐i
1i-‐i
2εii1i-‐i
s0,N~ε εSS
s0,IID~ε εSS
SSE s0,~ε εSS
+=
+=
=
+=
( ) ( )[ ]
( ) ( )
( )[ ] ( )
Δtσz 1tt
tt
tt
1i-‐1-‐ni
i
ii
1ii
1ii
eSS
SlnSlnE
tszSlnSln
1,0N~z nszSlnSln
⋅⋅⋅−
+
=
=
Δ⋅⋅+=
⋅⋅+=
−
−
Generally Rate
Periodic mean
Annual mean
Annual standard deviation
Period standard deviation
Rate pdf
a α
g γ
v u µ s σ Normal
d = SD(r) = SD(1+r)
r d δ Log normal
70
[ ][ ]
tΔBzSS
tΔB ,0N~ε
tttΔ tΔ ,0N~ε εSS
i1-‐ii
i
ii1-‐ii
ttt
2t
1i-‐itttt
⋅⋅+=
⋅
−=+=
[ ]
[ ]
mszum1i1mi
sm,umNm
1i
v
2m
1ii
i
2i
eSS
e~e
sm,umN~v
increment d,multiperio
⋅⋅+⋅−−+
⋅⋅
=
=
⋅=
⋅⋅
∏
∑[ ] [ ]
it
1i
i
1ii
eS
eSS
tzt,N~sm,umN~
t
tzttt
tt
22t
µ
Δ⋅σ⋅+Δ⋅µ
⋅=
⋅=
Δ⋅σ⋅+Δ⋅µ=µ
σµ⋅⋅µ
−
−
( )ΔtσzΔtμ
SS
ΔtσzΔtμ1SS
t*
i*
tt 1-‐ii
⋅⋅+⋅=Δ
⋅⋅+⋅+⋅=
ΔtzΔw
tttΔ
SSSΔ
1i-‐i
tt 1-‐ii
⋅=
−=
−=
71
( )
1)r(1g
1fg
)r(1f
fSS
)r(1 SS
n1
n
1iin
n1
nn
n
1iin
n0
n
n
1ii
0
n
−⎟⎟⎠
⎞⎜⎜⎝
⎛+=
−=
+=
=
+=
∏
∏
∏
=
=
=
ns
u
)r(1lns
vSSln
nn
n
1iin
n
1ii
0
n
=
+=
=⎟⎟⎠
⎞⎜⎜⎝
⎛
∑
∑
=
=
Levy Stable Distribu-ons 72
Levy Stable Distribu-ons 73
74
Power Law
¨ Power law with rescaled range ¨ Many natural phenomena modeled
with power law ¨ Nonlinear feedback ¨ Hurst exponent is the slope ¨ Fractal and self similar ¨ Complexity ¨ How can it be used in decision
making? ¨ The rescaled range follows a
power law
75
76
[ ][ ]
tΔBzSS
tΔB ,0N~ε
tttΔ tΔ ,0N~ε εSS
i1-‐ii
i
ii1-‐ii
ttt
2t
1i-‐itttt
⋅⋅+=
⋅
−=+=
Rate based process is Geometric Brownian Motion (GBM)
[ ]
[ ]
mszum1i1mi
sm,umNm
1i
v
2m
1ii
i
2i
eSS
e~e
sm,umN~v
increment d,multiperio
⋅⋅+⋅−−+
⋅⋅
=
=
⋅=
⋅⋅
∏
∑[ ] [ ]
it
1i
i
1ii
eS
eSS
tzt,N~sm,umN~
t
tzttt
tt
22t
µ
Δ⋅σ⋅+Δ⋅µ
⋅=
⋅=
Δ⋅σ⋅+Δ⋅µ=µ
σµ⋅⋅µ
−
−
( )ΔtσzΔtμ
SS
ΔtσzΔtμ1SS
t*
i*
tt 1-‐ii
⋅⋅+⋅=Δ
⋅⋅+⋅+⋅=
ΔtzΔw
tttΔ
SSSΔ
1i-‐i
tt 1-‐ii
⋅=
−=
−=
Appendix: Exponen-als and Natural Logs 77
( )
( )dxdy
y1
dxln(y)d
edxdy
dxed yy
⋅=
⋅=
+++++=
⎟⎠⎞
⎜⎝⎛ +=
∞→
!4x
!3x
!2xx1e
n11lime
432x
n
n
xlndxX1
ea1dxe xaxa
=
⋅=
∫
∫ ⋅⋅
Appendix: Exponen-als and Natural Logs 78
Price as a Stochas-c Diff Eqn 79
( )( )1eSSd
1eSS
eSSS
eSSS
eSS
eSS
dwtd
tztt
tzttt
tzt
t
t
tzt
t
t
tzttt
i
1i
i
1i1i
i
1i
1i
i
1i
i
i
1ii
−⋅=
−⋅=Δ
⋅=+Δ
=+Δ
=
⋅=
⋅σ+⋅µ
Δ⋅σ⋅+Δ⋅µ
Δ⋅σ⋅+Δ⋅µ
Δ⋅σ⋅+Δ⋅µ
Δ⋅σ⋅+Δ⋅µ
Δ⋅σ⋅+Δ⋅µ
−
−−
−
−
−
−
( )SfF =
80
( )[ ] ( )[ ] ... 1eSSF
211eS
SFdt
tFdF
... dSSF
21dS
SFdt
tFdF
2dwtd2
2dwtd
22
2
+−⋅∂
∂⋅+−⋅
∂
∂+
∂
∂=
+∂
∂⋅+
∂
∂+
∂
∂=
⋅σ+⋅µ⋅σ+⋅µ
( )
( )
( )dxdy
y1
dxdy
y1
dxln(y)d
edxyd
dxed
edxdy
dxed
2
y2
2
2
y2
yy
⋅=⋅=
⋅=
⋅=
dxdS
S1
dxdS
S1
dx d
dxdS
S1
dxln(S) d
2 ⋅−=⎟⎠⎞
⎜⎝⎛ ⋅
⋅=
( )
n0
nu0
tμ0t
*
*
n*
nnu
n0
nu0
)a1(SeSeS]E[S
)a1ln(u
)a1ln(nn1u
)a1(lnnu
)a1(e
)a1(SeS
**
*
*
+⋅=⋅=⋅=
+=
+⋅⋅=
+=⋅
+=
+⋅=⋅
⋅⋅
⋅
⋅
¨ Actually, they [power laws] aren’t special at all. They can arise as natural consequences of aggrega-on of high variance data. You know from sta-s-cs that the Central Limit Theorem says distribu-ons of data with limited variability tend to follow the Normal (bell-‐shaped, or Gaussian) curve. There is a less well-‐known version of the theorem that shows aggrega-on of high (or infinite) variance data leads to power laws. Thus, the bell curve is normal for low-‐variance data and the power law curve is normal for high-‐variance data. In many cases, I don’t think anything deeper than that is going on.
81