# Optimal bet-sizing and the Kelly Criterion

Mark Richardson, 24th October 2012

## Contents

(Chebfun example statistics/KellyCriterion.m)

LW = 'LineWidth'; MS = 'MarkerSize'; format short


## 1. Theoretical setup

Suppose you have a fixed amount of money to invest, say £100, and you are made aware of a particular investment opportunity which may be entered into an unlimited number of times, each time on identical terms. The rules of the wager are as follows. Your total balance grows or shrinks over time depending on the accumulative outcomes of your bets. You may bet any amount of your balance on each wager. With probability $p$, your stake is trebled and returned to you; with probability $1-p$, your stake is lost. Two questions follow:

1) Do you want to play this game?

2) If so, how much do you want to wager?

Regarding the first of these questions: Notwithstanding any philosophical or other objections, the answer is, clearly, that it depends on $p$. In fact, we can see quite easily that since the payoff odds of this wager sit at $2:1$ (i.e., you can win two units for every one unit wagered), then if $p \leq 1/3$ then we would certainly not wish to play since the expected value of the game is nonpositive. So let us assume that $p > 1/3$ in order that the answer to the first question is "yes". In fact, for definiteness, let's say $p = 1/2$.

An answer to the second question was first given in 1956 by a Bell Labs scientist named John Kelly [1]. (Kelly, incidentally was working at the time on problems in the fledgling discipline of information theory together with eminent scientists such as Claude Shannon of the famous Nyquist-Shannon sampling theorem.) Kelly's solution to the bet-sizing problem became known as the "Kelly Criterion". It was used to great effect later on by, amongst others, the MIT Blackjack card-counting team. It is now a mainstay of any self-respecting first course on basic investment principles.

To set up this problem, we consider a quantity $G$ called the "logarithmic growth rate of capital". Suppose that our starting amount is $C_0$, and that $C_n$ is the amount we have accumulated after the $n^{th}$ play of the game. Then, we define

$$G := \lim_{n \to \infty} \frac{1}{n} \log \left( \frac{C_n}{C_0} \right) .$$

Each time, we shall seek to wager a fixed proportion $0 \leq f \leq 1$ of our capital. Suppose that $w$ is the number of wins and $l$ is the number of losses resulting from $n$ wagers. Note that $n = w + l$. Suppose also that our game is generalised such that the payoff odds of the game are $a:1$ for some $a > 0$, so that we win $a$ units with probability $p$, and lose $1$ unit with probability $1-p$. Then we have

$$C_n = (1+af)^w (1-f)^l C_0 .$$

Thus,

$$G = \lim_{n \to \infty}\left( \frac{w}{n}\log(1+af) + \frac{l}{n}\log(1-f) \right),$$

so that, with probability one, we obtain

$$G(f) = p\log(1+af) + (1-p)\log(1-f).$$

This is the logarithm of the expected increase in our capital per bet. Our task now is to attempt to maximise this quantity over $f \in [0,1]$. Working through some basic calculus (i.e., differentiating and computing the root of the resulting linear function), we find that $G$ is maximised at

$$f = \frac{ap - (1-p)}{a} .$$

This is the Kelly Criterion. For our problem above, we had $a = 2$, $p = 1/2$. Using the Kelly Criterion formula in this case gives us $f = 1/4$. In other words, the optimal strategy is to wager one quarter of our bankroll each time we play.

Let's look at the expected capital growth function $\exp(G(f))$ for this example. Note that we use a truncated interval due to the logarithmic singularity at $x=1$.

p = 0.5; a = 2;
G = @(f) p*log(1+a*f) + (1-p)*log(1-f);
GG = chebfun(G,[0 0.999]);
plot(exp(GG),LW,2), grid on
[Gmax,f] = max(GG);
hold on, plot(f,exp(Gmax),'.r',MS,20), hold off
title('Expected rate of capital growth per bet')
xlabel('f'), ylabel('exp(G(f))'), ylim([0 1.2])


The computed optimal fraction agrees with the Kelly formula:

disp(f)

    0.2500



Using this strategy over the long term, we can therefore expect, on average, our capital to be multiplied by the following quantity per bet:

disp(exp(Gmax))

    1.0607



With these game parameters, betting anything less than a quarter of our wealth is too conservative, and betting any more is too aggressive. We also note additionally that we expect to make money over the long term so long as $G(f)$ is positive. The crucial interval for which this is the case can be trivially determined using roots:

roots(GG)

ans =

0.0000
0.5000



## 2. Numerical approach

Numerical methods were not really need above; they just provided a useful way for us to check the theory. However, should we wish to attack more complicated problems, numerical methods, and in particular Chebfun, will be of value. Indeed, the binomial problem considered above is really far too simplistic to be of much use in practical applications.

The following example involving staggered payoffs is still artifical, but is hopefully closer to a situation one might encounter in practice.

We have a bet with six possible outcomes. Associated with each outcomes are six probabilies $p_j$, which sum to 1, and some associated payoffs $a_j$. Suppose that the probabilities and payoffs are as follows:

p1 = 0.4; p2 = 0.21; p3 = 0.26; p4 = 0.1; p5 = 0.02; p6 = 0.01;
a1 = -1 ; a2 = 0;    a3 = 1.2;  a4 = 1.3; a5 = 1.4;  a6 = 10;


We can then follow the same setup as before by defining:

G = @(f) p1*log(1+a1*f) + p2*log(1+a2*f) + p3*log(1+a3*f) + ...
p4*log(1+a4*f) + p5*log(1+a5*f) + p6*log(1+a6*f);


Now the global optimisation step is taken care of numerically. This gives us the following:

GG = chebfun(G,[0 0.999]);
plot(exp(GG),LW,2), grid on
[Gmax,f] = max(GG);
hold on, plot(f,exp(Gmax),'.r',MS,20), hold off
title('Expected rate of capital growth per bet')
xlabel('f'), ylabel('exp(G(f))'), ylim([0 1.2])


And the correct proportion to bet in this case is therefore:

disp(f)

    0.1201



Finally, we should expect to lose our wealth if we bet any more than the following fraction:

r = roots(GG); disp(r(2))

    0.2487



We examined this problem in the somewhat idealised setting for which accurate and a priori knowledge of the probabilities and corresponding payoffs was assumed. However, such information is rarely available in practice; typically proababilities are estimated from past data. For this reason, it is common to take a more conservative approach and bet less than the Kelly fraction -- "half-Kelly" is typical. This approach also has the additional benefit of reducing the volatility (read annualised standard-deviation of logarithmic returns) which can result from betting the full Kelly fraction.

References:

[1] J. L. Kelly, Jr, A New Interpretation of Information Rate, Bell Systems Technical Journal, 35, (1956), 917–-926

[2] Unknown author: http://www.elem.com/~btilly/kelly-criterion/