Lecture 2: The Kelly criterion for favorable games:

stock market investing for individuals

David Aldous

August 29, 2017

 Most adults drive/own a car
Few adults work in the auto industry.
By analogy
Most (middle class) adults will have savings/investments
Few adults work in the banking/finance industries.
In my list of 100 contexts where we perceive chance

(22) Risk and reward in equity ownership

refers to the investor’s viewpoint;

(81) Short-term fluctuations of equity prices, exchange rates etc

refers to the finance professional’s viewpoint.
Over the last 30 years there has been a huge increase in the use of
sophisticated math models/algorithms in finance, and many
“mathematical sciences” majors seek to go into careers in finance. There
are many introductory textbooks, such as Capinski - Zastawniak
Mathematics for Finance: An Introduction to Financial Engineering. and
college courses such as IEOR221. But these represent first steps toward a
professional career; not really relevant to today’s lecture.

Today’s topic is “the stock market from the typical investor’s viewpoint”.
Best treatment is Malkiel’s classic book A Random Walk Down Wall
Street. But instead of summarizing that book, I will focus on one aspect,
and do a little math.
There are many different theories/viewpoints about the stock market,
none of which is the whole truth. So don’t believe that anything you read
is the whole truth.
A conceptual academic view: financial markets are mostly about
moving risk from those who don’t want it to those who are willing
to be paid to take over the risk.
The rationalist view: today’s stock price reflects consensus
discounted future profits, plus a risk premium. This “explains”
randomness mathematically (martingale theory). Textbook math
theory starts by assuming some over-simplified random model
without wondering how randomness arises.
Many “psychological” theories say stock prices can stay out of
alignment with “true value” for many years – cycles of sectors
becoming fashionable/unfashionable.
“Fundamental analysis” – see the Decal course Introduction to Wall
Street – seeks to assess “true value” better than the market. The
efficient market hypothesis says this is not practical.
The “just a casino” view emphasizes the fact than most trading is
from one owner of existing shares to another, rather than raising
new capital for a business to start or grow.
Our starting point in this lecture . . . . . .
The future behavior of the stock market will be statistically similar
to the past behavior in some respects but will be different in other
respects – and we can’t tell which.
This is true but not helpful! So go in one of two directions.
1 Devise your investment strategy under the assumption that “the
future will be statistically similar to the past”, recognizing this isn’t
exactly true.
2 Decide (by yourself or advice from others) to believe that the future
will be different from what the market consensus implies in certain
specific ways, and base your strategy on that belief.
Wall St makes money mostly by (2), selling advice or speculating with
their own money.
I’m not going to discuss whether you should invest in the stock market at
all. If you choose to do so, here’s the academic viewpoint.
The default choice is (something like) a S&P index fund, available
with very low expenses.
As a matter of logic, because most investments are made via
professional managers, their average gross return must be about the
same as the market average, so the actual return to an individual
investor must be on average be less than the market average,
because managers charge fees and expenses.
There’s overwhelming empirical evidence (next slide, and course
project: survey the literature) that individual investors on average
do even worse, typically by going in and out of the market, or
switching investments.
In class I ask students
Suppose you invest $1,000 today in the stock market, more
precisely in an S&P500 index fund. What do you guess the
investment will be worth in 10 years?
Converting their answers to annual percentage growth, the spread of
their guesses is shown here.

Table: Student guesses annual S&P500 growth next 10 years, and actual
growth rate over previous 3 calendar years

date 25th percentile median 75th percentile previous

9/17/2008 -0.5% 5.4% 8.4% 6.6%
9/6/2011 -3.5% 4.1% 6.1% - 5.0%
9/3/2014 4.2% 7.8% 11.2% 13.6%
1/20/2016 4.8% 8.2% 14.9% 12.7%
8/24/2017 3.4% 7.2% 19.6% 9.0%

This is roughly consistent with the often-remarked observation that

amateur investors tend to base their expectations on the previous several
years performance.
[show million dollar bet]
In comparing our default (index fund) with more sensible possible
alternatives, the issue is measuring risk and reward in the stock market.
Here is the “anchor” data for this lecture.

[show IFA page]

Implicit in the figure is that we measure

reward = long-term growth rate
risk = SD of percentage change each year.
Here is a first issue that arises.
it is very hard to pin down a credible and useful number for the
historical long-term average growth rate of stock market
investments.
Over the 50 years 1967-2017 the total return (including dividends) from
the S&P500 index rose at (geometric) average rate
10.28% . This is from IFA – let’s check another source.
[show other source].
Aside from the (rather minor) point that we are using a particular index
to represent the market what could possibly be wrong with using this
figure? Well,
it ignores expenses
it is sensitive to choice of start and end dates; starting in 1950
would make the figure noticeably higher, whereas ending in 2009
would make it noticeably lower.
to interpret the figure we need to compare it to some alternative
investment, by convention some “risk-free” investment.
it ignores inflation
it ignores taxes.
Here are two graphs which give very different impressions of long-term
stock market performance.
[show S&P index – click on max]
[show inflation-adjusted S&P]
In comparing our default (index fund) with more sensible possible
alternatives, the issue is measuring risk and reward in the stock market.
Regarding reward, as said above
it is very hard to pin down a credible and useful number for the
historical long-term average growth rate of stock market
investments.
But a typical conclusion is
Over the very long run, the U.S. stock market has had an
inflation-adjusted annualized return rate of between six and
seven percent.
-----------------------------------
It is not obvious how to measure – assign a numerical value to – the risk
of an investment strategy. If I lend you $100 today and you promise to
pay me back $200 in 10 years then my “reward” is 7.0% growth rate but
my “risk” is that you don’t pay me back – and I don’t know that
probability. A convention in the stock market context is to interpret risk
as variability, and measure it as SD of annual percentage returns.
(Project: literature on alternative measures of risk?).
[Show link 2 again – annual returns]
IFA and similar sites start out by trying to assess the individual’s
subjective risk tolerance using a questionnaire. The site then suggests
one of a range of 21 portfolios, represented on the slightly curved line in
the figure. The horizontal axis shows standard deviation of annual return,
(3% to 16%), and the vertical axis shows mean annual return (6% to
13%). Of course this must be historical data, in this case over the last 50
years. Notwithstanding the standard “past performance does not
guarantee future results” legal disclaimer, the intended implication is that
it is reasonable to expect similar performance in future.

Questions: (a) How does this relate to any theory?

(b) Should you believe this predicts the actual future if you invested?
(c) Does the reward/risk curve continue upwards further?

Answers: (a) The Kelly criterion says something like this curve must
happen for different “good” investment strategies.
(b) Even if future is statistically similar to past, any algorithm will
“overfit” and be less accurate at predicting the future than the past.
(c) No.
Expectation and gambling.
Recalling some basic mathematical setup, write P(·) for probability and
E[·] for expectation. Regarding gambling, any bet has (to the gambler)
some random profit X (a loss being a negative profit), and we say that
an available bet is (to the gambler)
favorable if E[X ] > 0
unfavorable if E[X ] < 0
and fair if E[X ] = 0.
Note the word fair here has a specific meaning. In everyday language, the
rules of team sports are fair in the sense of being the same for both
teams, so the better team is more likely to win. For 1 unit bet on team
B, that is a bet where you gain some amount b units if B wins but lose
the 1 unit if B loses,

E[profit] = bp − (1 − p); p = P(B wins)

and so to make the bet is fair we must have b = (1 − p)/p. (Confusingly,

mathematicians sometimes say “fair game” to mean each player has
chance 1/2 to win, but this is sloppy language).
Several issues hidden beneath this terminology should be noted. Outside
of games we usually don’t know probabilities, so we may not know
whether a bet is favorable, aside from the common sense principle that
most bets offered to us will be unfavorable to us.

One of these days in your travels, a guy is going to show you a brand-new
deck of cards on which the seal is not yet broken. Then this guy is going
to offer to bet you that he can make the jack of spades jump out of this
brand-new deck of cards and squirt cider in your ear. But, son, do not
accept this bet, because as sure as you stand there, you’re going to wind
up with an ear full of cider.
[spoken by Sky Masterson in Guys and Dolls, 1955].
The terminology (fair, favorable, unfavorable) comes from the law of large
numbers fact that if one could repeat the same bet with the same stake
independently, then in the long run one would make money on a favorable
bet but lose money on an unfavorable bet. Such “long run” arguments
ignore the issues of (rational or irrational) risk aversion and utility theory,
which will be discussed later. In essence, we are imagining settings where
your possible gains or losses are small, in your own perception.
Unfavorable bets.
Roughly speaking, there are two contexts in which we often encounter
unfavorable bets. One concerns most activities we call gambling, e.g. at
a casino, and the other concerns insurance. Regarding the former,
mathematicians often say ridiculous things such as
Gambling against the house at a casino is foolish, because the
odds are against you and in the long run you will lose money.
What’s wrong is the because. Saying
Spending a day at Disneyland is foolish, because you will leave
with less money than you started with
is ridiculous, because people go to Disneyland for entertainment, and
know they have to pay for entertainment. And the first quote is equally
ridiculous. Casino gamblers may have irrational ideas about chance and
luck, but in the U.S. they typically regard it as entertainment with a
chance of winning, not as a plan to make money. So it’s worth being
more careful and saying
 Gambling against the house at a casino and expecting to
make money is foolish, because the odds are against you and
in the long run you will lose money.
The second context is that buying insurance is mathematically similar to
placing an unfavorable bet – your expected gain in negative, because the
insurance company wants to cover its costs and make a profit. But the
whole point of buying insurance is risk aversion, so this needs to be
treated in the setting of utility theory and psychology of probability (a
later Lecture).
 So where can I find a favorable bet?
The wiseacre answer “start your own casino or insurance company” is not
so practical, but a variant of the latter is. For those who can, following
the advice
increase your insurance deductibles to the maximum you can
comfortably afford to lose
is a favorable bet, likely to save you money over a lifetime.
In this lecture we consider investing in the stock market as
mathematically similar to making a sequence of favorable bets (and
letting your winnings ride). Exactly why one could consider this a
favorable bet could be debated endlessly – standard economic theory
asserts that investors need to be rewarded for taking risk rather than
using alternative risk-free investments, while empiricists observe that, in
countries without anti-capitalist revolutions, the historical performance of
stock markets actually has been better than those alternatives.
 long term versus short term.
In everyday language, a job which will only last six months is a short
term job; someone who has worked for a company for fifteen years is a
long term employee. Joining a softball team for a summer is a short term
commitment; raising children is a long term commitment. We judge
these matters relative to human lifetime; long term means some
noticeable fraction of a lifetime.

Table: Effect of 7% interest, compounded annually.

year 0 4 8 12 16 20
simple interest 1000 1,280 1,560 1,840 2,120 2,400
compound interest 1000 1,311 1,718 2,252 2,952 3,870
One of several possible notions of long term in financial matters is “the
time span over which compounding has a noticeable effect”. Rather
arbitrarily interpreting “noticeable effect” as “10% more” and taking the
7% interest rate, this suggests taking 8 years as the cut-off for long term.
Being about 10% of a human lifetime, this fortuitously matches
reasonably well the “noticeable fraction of a lifetime” criterion above.
And indeed in matters pertaining to individuals, financial or otherwise,
most writers use a cut-off between 5 and 10 years for “long term”.
 The mathematical theme of this lecture is the nature of
compounding when gains and losses are unpredictable.
The relevant arithmetic is multiplication not addition: a 20% gain
followed by a 20% loss combine to a 4% loss, because
1.2 × 0.8 = 0.96.
Let’s move on to some mathematics . . . . . .
First let us make explicit the type of model used implicitly above. A
“return” x = 0.2 or x = −0.2 in a year means a 20% gain or a 20% loss.

The IID model.

Write Xi for the return in year i. Suppose the (Xi ) are IID random
variables. Then the value Yn of your investment at the end of year n is
n
Y
Yn = Y0 (1 + Xi ) (1)
i=1

where Y0 is your initial investment.

Conceptually, we are assuming the future is statistically the same as the
past, and assuming independence over different time periods.
To analyze this model we take logs and divide by n:
n
X
n−1 log Yn = n−1 log Y0 + n−1 log(1 + Xi )
i=1

and the law of large numbers says that as n → ∞ the right side
converges to E[log(1 + X )]. We want to compare this to an investment
with a non-random return of r . For such an investment (interest rate r ,
compounded annually) we would have Yn = Y0 (1 + r )n and therefore
n−1 log Yn → log(1 + r ). Matching the two cases gives the conclusion

In the IID model, the long term growth rate is

exp(E[log(1 + X )]) − 1.
The formula looks strange, because to compare with the IID annual
model we are working with the equivalent “compounded annually”
interest rate. It is mathematically nicer to use instead the “compounded
instantaneously” interest rate, which becomes just E[log(1 + X )].
The main impact of this result is that what matters “in the long term”
about the random return X is not precisely the mean E[X ], but rather its
“multiplicative” analog E[log(1 + X )]. Let us note, but set aside for a
while, the points
Is the model realistic for stock market investing?
The phrase long term here refers to the applicability of the law of
averages as an approximation to finite time behavior – this is a third
meaning of the phrase, logically quite distinct from the two previous
meanings.
Instead we focus on the conceptual point is that there are many
investment possibilities, that is ways to allocate money to different risky
or safe assets. Write α for a portfolio, that is a way of investing given
proportions of your fortune in different assets.

We can now jump to our first key mathematical point.

 The Kelly Criterion. Suppose you have a range of possible
investment portfolios α, which will produce return Xα , with
known distribution. Then (assuming the IID model) the
portfolio α that maximizes the long term growth rate is the
portfolio that maximizes E[log(1 + Xα ]). So choose that
portfolio.
The math here is just STAT134, but the implications are rather subtle, as
we will see from playing with some hypothetical, very simple models.

Mathematics of the Kelly criterion: one risky and one safe asset.
Suppose there is both a “risky” (random return) asset (a “stock”, more
realistically a S&P500 index fund) and a risk-free alternative investment
(a “bond”) that pays a fixed interest rate r .
Suppose we choose some number 0 ≤ p ≤ 1 and at the start of each year
we invest a proportion p of our total “investment portfolio” in the stock
market, and the remaining proportion 1 − p in the bond. In this case our
return in a year is
X ∗ = pX + (1 − p)r
where X is the return on the stock. The long term growth rate is now a
function of p: in the continuous setting

growth(p) = E[log(1 + pX + (1 − p)r )]. (2)

The Kelly criterion says: choose p to maximize growth(p). Let’s see two
examples. In the first X is large, and we end up with p small; in the
second X will be small, and we end up with large p. In these two
examples we take the time unit to be 1 day instead of 1 year (which
doesn’t affect math formulas).
 Example: pure gambling.
Imagine a hypothetical bet which is slightly favorable. Suppose each day
we can place a bet of any size s; we will either gain s (with probability
0.5 + δ) or lose s (with probability 0.5 − δ), independently for different
days (here δ is assumed small). Take r = 0 for the moment. What
proportion p of our portfolio do we want to bet each day?
 growth rate
2δ 2

2δ p 4δ
Here, for small δ,

E[log(1 + pX )] = ( 12 + δ) log(1 + p) + ( 21 − δ) log(1 − p)

≈ ( 12 + δ)(p − p 2 /2) + ( 12 − δ)(−p − p 2 /2)
= 2δp − p 2 /2.

Thus the asymptotic growth rate is approximately the quadratic function

of p
G (p) = 2δp − p 2 /2 (3)
shown in the Figure. The Kelly criterion says to choose p ≈ 2δ and then
your long term growth rate will be ≈ 2δ 2 .
Now recall that we simplified by taking r = 0; when r > 0, the fact that
a proportion 1 − p ≈ 1 of the portfolio not put at risk each day can earn
interest, brings up the optimal growth rate to r + 2δ 2 ; the quantity 2δ 2
represents the extra growth one can get by exploiting the favorable
gambling opportunity.

To give a more concrete mental picture, suppose δ = 1%. The model

matches either of the two following hypothetical scenarios.

(a) To attract customers, a casino offers (once a day) an opportunity to

make a roulette-type bet with a 51% chance of winning.
(b) You have done a statistical analysis of day-to-day correlations in some
corner of the stock market and have convinced yourself that a certain
strategy (buying a portfolio at the start of a day, and selling it at the
end) replicates the kind of favorable bet in (a).
In either scenario, the quantity 2δ 2 = 2/10, 000 is the “extra” long term
growth rate available by taking advantage of the risky opportunity. Note
this growth rate is much smaller than 2% “expected gain” on one bet.
On the other hand we are working “per day”, and in the stock market
case there are about 250 days in a year, so the growth rate becomes
about 5% per year; recalling this is “5% above the risk-free interest
rate”, it seems a rewarding outcome. But if δ were instead 0.5% then the
extra growth rate becomes 1 14 %, and (taking into account transaction
costs and our work) the strategy hardly seems worth the effort.

Implicit in the Figure (back 2 pages) is a fact that at first strikes

everyone as counter-intuitive. The curve goes negative when p increases
above approximately 4δ. So even though it is a favorable game, if you are
too greedy then you will lose in the long run!
This setting was artificially simple; here is a first step towards a more
realistic setting.
Example: what proportion of your portfolio to put into the “stock”?
As before, suppose there is both a “risky” (random return) asset (a
“stock”, more realistically a S&P500 index fund) and a risk-free
alternative investment (a “bond”) that pays a fixed interest rate r . Now
we imagine typical values of the return X as ±0.2., on a yearly
time-scale. As before we choose some number 0 ≤ p ≤ 1 and at the start
of each year we invest a proportion p of our total “investment portfolio”
in the stock market, and the remaining proportion 1 − p in the bond. We
know that the long term growth rate is a function of p:

growth(p) = E[log(1 + pX + (1 − p)r )].

There is a nicer algebraic way of dealing with the interest rate r . Set

X = r + (1 + r )X ∗

and interpret X ∗ = (X − r )/(1 + r ) as “return relative to interest rate”.

Then a couple of lines of algebra let us rewrite the formula as

growth(p) = (1 + r ) exp(E[log(1 + pX ∗ )]) − 1 (4)

and the optimization problem now doesn’t involve any r .

If we imagine the stock market on a daily time-scale and suppose
changes X ∗ are small, with mean µ and variance σ 2 , then we can use the
series approximation

log(1 + pX ∗ ) ≈ pX ∗ − 21 (pX ∗ )2

to calculate

E[log(1 + pX ∗ )] ≈ pµ − 12 p 2 (µ2 + σ 2 ) ≈ pµ − 21 p 2 σ 2

(the latter because µ and σ 2 are in practice of the same order, so µ2 is of

smaller order than σ 2 ). So the Kelly criterion says: choose p to maximize
pµ − 12 p 2 σ 2 , that is choose

p = µ/σ 2 . (5)

This is another remarkable formula, and let us discuss some of its

mathematical implications.
1. The formula is (as it should be) time-scale free. That is, writing
2 2
µday , µyear , σday , σyear for the means and variances over a day and a
N-day year, then (because compounding has negligible effect over a year)
2 2
µyear ≈ Nµday and σyear ≈ Nσday , so we get the same value for µ/σ 2
whether we work in days or years.
2. Even though we introduced the setup by stating that 0 ≤ p ≤ 1, the
model and its analysis make sense outside that range. Economic theory
and experience both say that the case µ < 0 doesn’t happen (investors
are risk averse and so would buy no stock; this would cause the current
price of stock to drop), but if it did then the formula p = µ/σ 2 < 0 says
that not only should be invest 100% of our wealth in the bond, but also
we should “sell short” (i.e. borrow) stock and invest the proceeds in the
bond.
3. More interesting is the case p > 1. Typical values given for the
S&P500 index (as noted earlier, stating meaningful historical values is
much harder than one might think) are (interest-rate-adjusted) µ = 5.6%
and σ = 20%, in which case the Kelly criterion says to invest a
proportion p = 140% of your wealth in the stock market, i.e. to borrow
money (at fixed interest rate r ) and invest your own and the borrowed
money in the stock market.
What about the not-so-long term?
We started with the multiplicative model, which assumes that returns in
different time periods are IID. This is not too realistic, but the general
idea behind the Kelly criterion works without any such assumption, as we
now explain.
Going back to basics, the idea
to invest successfully in the stock market, you need to know
whether the market is going to go up or go down
is just wrong. Theory says you just need to know the probability
distribution of a future return. So suppose (a very big SUPPOSE, in
practice!) at the beginning of each year you could correctly assess the
probability distribution of the stock return over the coming year, then you
can use the Kelly criterion (2) to make your asset allocation. The fact
that the distribution, and hence your asset allocation, would be different
in different years doesn’t make any difference – this strategy is still
optimal for long-term growth.
 The second key insight from mathematics
The numbers for growth rates that come out of the formula of course
depend on the distributions of each next year’s returns, but there’s one
aspect which is “universal”. In any situation where there are sensible
risky investments, following the Kelly strategy means that you accept a
short-term risk which is always of the same format:
40% chance that at some time your wealth will drop to only
40% of what you started with.
The magical feature of this formula is that the percents always match: so
there is a 10% chance that at some time your wealth will drop to only
10% of what you started with.

The math here is from the theory of diffusions or stochastic calculus; a

little too hard to explain here.
 x% chance that at some time your wealth will drop to only x%
of what you started with.
For an individual investor, it is perfectly OK to be uncomfortable with
this level of medium-term risk and to be be less aggressive (in investment
jargon) by using a partial Kelly strategy, that is using some smaller value
of
p = proportion of your assets invested in stocks
than given by the Kelly criterion. Theory predicts you will thereby get
slower long-term growth but with less short-term volatility.

A memorable quote
The Kelly strategy marks the boundary between aggressive and
insane investing.
How one might expect this theory (based on assuming known true
probability distributions for the future, and on seeking to optimize
long-term growth rate) to relate to the actual stock market is not
obvious, but one can certainly look at what the actual percentages have
been.

The next figure shows the historical distribution (based on hypothetical

purchase of S&P500 index on first day of each year 1950-2009 and on
subsequent monthly closing data) of the minimum future value of a 100
investment.
0 50 60 70 80 90 100

This is obviously very different from the “Kelly” prediction of a flat

histogram over [0, 100]. This data is not adjusted for inflation or for
comparison with a risk-free investment, and such adjustments (course
project?) would make the histogram flatter, but still not close to the
Kelly prediction. We mentioned before that over the historical long term
it has been more profitable to borrow to invest more than 100% of your
assets in the market. Both observations reflect the fact that the stock
market fluctuates less than would the fortune of a Kelly-optimizing
speculator.
 Wrap up
The math I’ve shown is still a long way away from explaining the IFA
graphic. They have many different “asset classes”, each represented by
an index fund [show IFA]. We want to construct a portfolio by weighting
over each asset class (note we still avoid individual stocks). Any possible
portfolio has (historical data) some average and some SD of annual
return. We just choose a range of portfolios that maximize average for
given SD (this exploits past pattern of correlations as well as averages
and SDs).
Note that WolframAlpha will do this kind of thing for you [show].
The standard textbook math is Modern portfolio theory which is
essentially the short-term Normal approximation for price fluctuations.
The reason I emphasize Kelly instead is that (in principle) you could
model “Black Swans” (rare severe shocks) or at least use actual historic
annual return distributions rather than assuming log-Normal.
I’m not going to discuss whether you should invest in the stock market at
all. If you choose to do so, here’s the academic viewpoint.
The default choice is (something like) a S&P index fund, available
with very low expenses.
-------------------------------
If you really want to do better then there are three vaguely plausible ways
none of which I recommend.
Borrow money, if you can find a low interest rate.
Diversified Kelley-optinmized portfolio, like IFA.
Long term market timing such as the Shiller PE ratio [show]
 Further reading
The two books most related to our “Kelly criterion” are
William Poundstone Fortune’s Formula: a history of precisely this topic
Aaron Brown Red-Blooded Risk: The Secret History of Wall Street: his
experience of 1980s risk management in finance.
A memorable quote from Brown: Kelly enables you to get rich
exponentially slowly.
Any book by Robert Schiller.
For a little more math see the paper Good and bad properties of the
Kelly criterion.