The Kelly Growth Criterion
Transcript of The Kelly Growth Criterion
![Page 1: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/1.jpg)
The Kelly Growth Criterion
Niels Wesselhöt
Wolfgang K. Härdle
International Research Training Group 1792Ladislaus von Bortkiewicz Chair of StatisticsHumboldtUniversität zu Berlin
http://irtg1792.hu-berlin.dehttp://lvb.wiwi.hu-berlin.de
![Page 2: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/2.jpg)
Motivation 1-1
Portfolio choice
Playing Blackjack
Figure 1: 'Ed' Thorp Figure 2: Matlab GUI
![Page 3: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/3.jpg)
Motivation 1-2
Portfolio choice
Wealth for discrete returns Xi ∈ Rk
Wn(f ) = W0
n∏i=1
1 +k∑
j=1
fjXj ,i
(1)
I W0 ∈ R+ starting wealth
I k ∈ N+ assets with index jI n ∈ N+ periods with index i
How to chose fraction vector f ∈ Rk?
![Page 4: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/4.jpg)
Motivation 1-3
Managing Portfolio Risks
Two main strands
1. Mean-Variance approach: Markowitz (1952), Tobin (1958),Sharpe (1964) and Lintner (1965)
2. Kelly growth-optimum approach: Kelly (1956), Breiman(1961) and Thorp (1971)
Leo Breiman on BBI:
![Page 5: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/5.jpg)
Outline
1. Motivation X
2. Bernoulli - Kelly (1956)
3. Gaussian - Thorp (2006)
4. General i.i.d. - Breiman (1961)
5. Appendix
![Page 6: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/6.jpg)
Bernoulli - Kelly (1956) 2-1
Arithmetic mean maximization
Consider n favorable Bernoulli games with probability1
2< p ≤ 1 (q = 1− p) and outcome X = 1 (−1)
For P(X = 1) = p = 1, investor bets everything, f = 1
Wn = W02n (2)
Uncertainty - maximizing the expectation of wealth impliesf = 1
E(Wn) = W0 +n∑
i=1
(p − q)E (fWn−1) , (3)
Leads to ruin asymptotically
P (Wn ≤ 0) = Plimn→∞
(1− pn)→ 1 (4)
![Page 7: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/7.jpg)
Bernoulli - Kelly (1956) 2-2
Minimizing risk of ruin
Alternative: minimize the probability of ruin
For f = 0P (Wn ≤ 0) = 0 (5)
Minimum ruin strategy leads also to the minimization of theexpected prots as no investment takes place
![Page 8: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/8.jpg)
Bernoulli - Kelly (1956) 2-3
Geometric mean maximization
Gambler bets a fraction of his wealth with m games won
Wn = W0(1 + f )m(1− f )n−m (6)
Exponential rate of growth per trial(log of the geometric mean)
Gn(f ) = log
(Wn
W0
) 1n
= log
(1 + f )mn (1− f )
n−mn
(7)
=(mn
)log(1 + f ) +
(n −m
n
)log(1− f ) (8)
![Page 9: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/9.jpg)
Bernoulli - Kelly (1956) 2-4
Geometric mean maximization
By Borel's law of large numbers
E Gn(f ) = g(f ) = p · log(1 + f ) + q · log(1− f ) (9)
Maximizing g(f ) w.r.t. f :
g ′(f ) =
(p
1 + f
)−(
q
1− f
)=
p − q − f
(1 + f )(1− f )
= 0
(10)
≺ f = f ∗ = p − q, p ≥ q > 0 (11)
Second derivative according to f
g ′′(f ) = −
p
(1 + f )2
−
q
(1− f )2
< 0 (12)
![Page 10: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/10.jpg)
Bernoulli - Kelly (1956) 2-5
Closed form for Bernoulli trials
Growth optimal fraction, under Bernoulli trials:
f ∗ = p − q (13)
Maximizes the expected value of the logarithm of capital ateach trial
g(f ∗) = p · log(1 + p − q) + q · log(1− p − q) (14)
= p · log(p) + q · log(q) + log(2) > 0 (15)
A link to information theory
![Page 11: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/11.jpg)
Bernoulli - Kelly (1956) 2-6
Bernoulli example, p = 0.6
Exponential rate of asset growth for binary channel with p=0.6
0 0.5 1
f
-1.5
-1
-0.5
0
0 0.2
f
0
0.005
0.01
0.015
0.02
Figure 3: Bernoulli Exponential growth rate g(f)
![Page 12: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/12.jpg)
Bernoulli - Kelly (1956) 2-7
Bernoulli
-41
-3
-2
1
-1
p
0.5
0
f
0.5
0 0-1
-0.5
0
0.5
1
Figure 4: Bernoulli - Exponential growth rate g(f,p)
![Page 13: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/13.jpg)
Gaussian - Thorp (2006) 3-1
Gaussian (One-dimensional)
X ∼ F with E(X ) = µ and Var(X ) = σ2
Return of the risk free asset r > 0
Wealth given investment fractions and restriction∑k
j=1fj = 1
W (f ) = W0 1 + (1− f )r + fX (16)
= W0 1 + r + f (X − r) (17)
![Page 14: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/14.jpg)
Gaussian - Thorp (2006) 3-2
Gaussian (One-dimensional)
Maximize
g(f ) = E logWn(f ) = E G (f ) = E log Wn(f )/W0(18)
Wealth after n periods
Wn(f ) = W0
n∏i=1
1 + r + f (Xi − r) (19)
Taylor expansion of
E
[log
Wn(f )
W0
]= E
[n∑
i=1
log 1 + r + f (Xi − r)
](20)
![Page 15: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/15.jpg)
Gaussian - Thorp (2006) 3-3
Gaussian (One-dimensional)
Given log(1 + x) = x − x2
2+ · · ·
log 1 + r + f (X − r) = r + f (X − r)− r + f (X − r)2
2+ · · ·
(21)
≈ r + f (X − r)− X 2f 2
2(22)
Taking sum and expectation
E
[n∑
i=1
log 1 + r + f (Xi − r)
]≈ r + f (µn − r)− σ2nf
2
2
(23)
Myopia: taking∑n
i=1Xi has no impact on the solution
![Page 16: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/16.jpg)
Gaussian - Thorp (2006) 3-4
Gaussian (One-dimensional)
Result of the Taylor expansion
g(f ) = r + f (µ− r)− σ2f 2/2 +O(n−1/2). (24)
For n −→∞, O(n−1/2) −→ 0
g∞(f ) = r + f (µ− r)− σ2f 2/2. (25)
Dierentiating g(f ) according to f
∂g∞(f )
∂f= µ− r − σ2f = 0 ≺ f ∗ =
µ− r
σ2= σ−1MPR (26)
Betting the optimal fraction f ∗ leads to growth rate
g∞(f ∗) =(µ− r)2
2σ2+ r . (27)
g∞(f ) is parabolic around f ∗ with range 0 ≤ f ∗ ≤ 2f ∗
![Page 17: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/17.jpg)
Gaussian - Thorp (2006) 3-5
Gaussian - µ = 0.03, σ = 0.15, r = 0.01
0 0.5 1 1.5 2 2.5
f
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
Figure 5: Gaussian approximation - Exponential growth rate g(f)
![Page 18: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/18.jpg)
Gaussian - Thorp (2006) 3-6
Gaussian (Multi-dimensional)
X ∼ N(µ,Σ) and risk free rate r > 0
Wn(f ) = W0
1 + r + f >(X − r)
(28)
Taking logarithm and expectations on both sides leads viaTaylor series to
g(f ) = E
log(1 + r) +
1
1 + r(µ− 1r)>f − 1
2(1 + r)2f >Σf
(29)
From quadratic optimization (Härdle and Simar, 2015)
f ∗ = Σ−1(µ− 1r) (30)
g∞(f ∗) = r + f ∗>Σf ∗/2 (31)
![Page 19: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/19.jpg)
Gaussian - Thorp (2006) 3-7
Gaussian -µ = [0.03 0.08], σ = [0.15 0.15], ρ = 0, r = 0.01
03
0.05
32
f2
2
0.1
f1
1 1
0.15
0 00
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Figure 6: Gaussian approximation - Exponential growth rate g(f)
![Page 20: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/20.jpg)
General i.i.d. - Breiman (1961) 4-1
General i.i.d.
Asymptotic dominance (in terms of wealth) of the Kellystrategy in a general i.i.d. setting in discrete time
Figure 7: Warren Buf-
fett Figure 8: Matlab GUI
![Page 21: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/21.jpg)
General i.i.d. - Breiman (1961) 4-2
General i.i.d.
Investment strategy Λ =
fi ,j · · · fn,j...
. . ....
fi ,k · · · fn,k
= [fi · · · fn]
I investment fractions fi from time i to n ∈ N+
I opportunities j to k ∈ N+
Security price vector pi =
pi ,j...pi ,k
Return per unit invested xi =
pi,j
pi−1,j
...pi,k
pi−1,k
.
![Page 22: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/22.jpg)
General i.i.d. - Breiman (1961) 4-3
Discrete i.i.d. setting
Wealth of the investor in period n
Wn(fn) = Wn−1(fn−1)f >n xn
(32)
Wn(fn) increases exponentially
Log-optimal fraction through growth rate maximization ateach trial
f ∗ = argmaxf ∈Rk
E log(Wn) (33)
![Page 23: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/23.jpg)
General i.i.d. - Breiman (1961) 4-4
Asymptotic outperformance
Theorem
Myopic log-optimal strategy Λ∗ = [f ∗ · · · f ∗] Signicantly dierent strategy Λ
E logWn(Λ∗) − E logWn(Λ) −→ ∞, (34)
Kelly investor dominates asymptotically
limn→∞
Wn(Λ∗)
Wn(Λ)a.s.−→∞ (35)
Leo Breiman on BBI:
![Page 24: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/24.jpg)
General i.i.d. - Breiman (1961) 4-5
Minimize time to reach goal g
Theorem
Let N(g) be the smallest n, such that Wi ≥ g , g > 0
If equation (34) holds,
∃α ≥ 0 |= Λ, g (36)
such that
E N∗(g) − E N(g) ≤ α, (37)
|= - independent of Λ∗ asymptotically minimizes the time to reach goal g
![Page 25: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/25.jpg)
General i.i.d. - Breiman (1961) 4-6
Time invariance
Theorem
Given a xed set of opportunities the strategy is
I xed fraction
I independent of the number of trials n
Λ∗ = [f ∗1 · · · f ∗n ], f ∗1 = · · · = f ∗n (38)
![Page 26: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/26.jpg)
General i.i.d. - Breiman (1961) 4-7
Bernoulli revisited
Theorem
Two investors with equal initial endowment, investment
fractions f1 and f2
For exponential growth rates
Gn(f1) > Gn(f2) (39)
the Kelly bet dominates asymptotically
limn→∞
Wn(f1)
Wn(f2)a.s.−→∞ (40)
![Page 27: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/27.jpg)
General i.i.d. - Breiman (1961) 4-8
Bernoulli revisited
Proof.
Dierence in exponential growth rates Gn(f ) = log
Wn(f )W0
1n
log
Wn(f1)
W0
1n
− log
Wn(f2)
W0
1n
= log
Wn(f1)
Wn(f2)
1n
(41)
by Borel strong law of large numbers
P
[limn→∞
log
Wn(f1)
Wn(f2)
1n
]> 0
a.s.−→ 1. (42)
![Page 28: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/28.jpg)
General i.i.d. - Breiman (1961) 4-9
Bernoulli revisited
Proof.
For ω ∈ Ω, there exists N(ω) such that for n ≥ N(ω),
W0 exp nG (f1) >W0 exp nG (f2) (43)
Wn(f1) >Wn(f2) (44)
Asymptotically
limn→∞
Wn(f1)
Wn(f2)a.s.−→∞ (45)
![Page 29: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/29.jpg)
Utility functions - Thorp (1971) 5-1
Utility functions
Three types of utility theories: Thorp (1971)
I Descriptive utility - empirical data and mathematical tting
I Predictive utility - derives utility functions out of hypotheses
I Normative utility - describe the behavior to achieve a certain
goal
The logarithmic utility function is used in a normative way
![Page 30: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/30.jpg)
Utility functions - Thorp (1971) 5-2
Conclusion
Comparison of risk management theoriesI Markowitz-approach
• arithmetic mean-variance ecient• maximizing single period returns• rests on two moments
I Kelly-approach
• geometric mean-variance ecient• maximize geometric rate of multi-period returns• utilizes the whole distribution
![Page 31: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/31.jpg)
Information theory 6-1
Information
Closed form for Bernoulli trials
Self-information (uncertainty) of outcome x
i(x) = − log P(x) = log1
P(x)(46)
i(x) = 0, for P(x) = 1 (47)
i(x) > 1, for P(x) < 1 (48)
Example: For a fair coin, the change of P (x = tail) = 0.5
i(x) = − log2(1/0.5) = 1 bit
![Page 32: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/32.jpg)
Information theory 6-2
Information
0 0.2 0.4 0.6 0.8 1
Probability p
0
2
4
6
8
10
Info
rmat
ion
in b
it
Figure 9: Self information of an outcome given probability p
![Page 33: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/33.jpg)
Information theory 6-3
Entropy
Entropy as expectation of self-informations (averageuncertainty), given outcomes X = X1, . . . ,Xn
H(X ) = E I (X ) = −E log P(X ) (49)
= −∑x
P(x) log2 P(x) ≥ 0 (50)
For two outcomes and p = q = 0.5
H(X ) = − (p log2 p + q log2 q)
= − (1/2 log2 1/2 + 1/2 log2 1/2) = 1 bit
![Page 34: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/34.jpg)
Information theory 6-4
Entropy
0 0.2 0.4 0.6 0.8 1
Probability p
0
0.2
0.4
0.6
0.8
1E
ntro
py H
(X)
Figure 10: Entropy for two outcomes given probability p (1-p)
![Page 35: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/35.jpg)
Information theory 6-5
Entropy
Joint entropy
H(X ,Y ) = −E log P(X ,Y ) (51)
= −∑x ,y
P(x , y) log P(x , y) (52)
Conditional entropy
H(X | Y ) = −E log P(X | Y ) (53)
= −∑x ,y
P(x | y) log P(x | y) (54)
![Page 36: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/36.jpg)
Information theory 6-6
Noisy binary channel
Figure 11: Noisy binary channel
![Page 37: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/37.jpg)
Information theory 6-7
Mutual information
Mutual information
I (X ;Y ) = H(X )− H(X | Y ) (55)
= E
log
P(X | Y )
P(X )
(56)
For the binary symmetric channel
I (X ;Y ) =∑x
∑y
P(x , y) logP(x , y)
P(x)P(y)(57)
= q log(2q) + p log(2p) (58)
= p log p + q log q + log(2) (59)
![Page 38: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/38.jpg)
Information theory 6-8
Mutual information
0 0.2 0.4 0.6 0.8 1
Probability p
0
0.2
0.4
0.6
0.8
1
Mut
ual i
nfor
mat
ion
I(X
;Y)
Figure 12: Mutual Information for a binary channel
![Page 39: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/39.jpg)
Information theory 6-9
Mutual information
Figure 13: Relation of Entropy and Mutual Information
![Page 40: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/40.jpg)
Information theory 6-10
A link to information theory
I (X ;Y ) - mutual information
I highest possible rate of information transmission in the
presented channel
I also called the channel's information carrying capacity or rate
of transmission
Equivalence to equation (14)
I (X ;Y ) = g(f ∗) (60)
Closed form for Bernoulli trials
![Page 41: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/41.jpg)
Information theory 6-11
A link to estimation theory
Relative entropy or Kullback-Leibler divergence
D (P(x) || Q(x)) = −E
log
P(x)
Q(x)
(61)
=∑x
P(x) logP(x)
Q(x)≥ 0 (62)
Relation to mutual information
I (X ;Y ) = D P(x , y) || P(x)P(y) (63)
![Page 42: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/42.jpg)
The Kelly Growth Criterion
Niels Wesselhöt
Wolfgang K. Härdle
International Research Training Group 1792Ladislaus von Bortkiewicz Chair of StatisticsHumboldtUniversität zu Berlin
http://irtg1792.hu-berlin.dehttp://lvb.wiwi.hu-berlin.de
![Page 43: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/43.jpg)
References 7-1
For Further Reading
J. KellyA new interpretation of information rate
Bell System Technology Journal, 35, 1956
L. BreimanOptimal gambling system for favorable games
Proceedings of the 4th Berkeley Symposium on Mathematics,Statistics and Probability, 1, 1961
E. O. ThorpPortfolio choice and the Kelly criterion
Proceedings of the Business and Economics Section of theAmerican Statistical Association, 1971
![Page 44: The Kelly Growth Criterion](https://reader034.fdocuments.net/reader034/viewer/2022042122/62583065397b6346463931b5/html5/thumbnails/44.jpg)
References 7-2
For Further Reading
R. RollEvidence on the growth optimum model
The Journal of Finance, 1973
L. C. MacLean, W. T. Ziemba and G. BlazenkoGrowth versus Security in Dynamic Investment Analysis
Management Science, 38(11), 1992
E. O. ThorpThe Kelly criterion in Blackjack, Sports betting and the Stock
Market
Handbook of Asset and Liability Management, 2006