Column overall title: A Mathematician on Wall Street

Column 1

Option Theory: What I Knew and When I Knew It – Part I

by Edward O. Thorp
Copyright 2001

One of the themes of this column will be how and to what extent markets are
inefficient, and how you might profit from this.
Let’s begin by going back in time to the early days of quantitative finance. Paul
Cootner’s book, The Random Character of Stock Market Prices, M.I.T. Press, 1964,
presented much of the work that had been done on the random walk theory of stock
prices and on the problem of warrant pricing. The warrant valuation problem was
essentially the problem of valuing options and, more generally, derivatives. Progress
was substantial but the Black-Scholes breakthrough would not appear until 1973.
Meanwhile, in 1965 Eugene Fama proposed that markets were well described as
“efficient,” with all-knowing rational participants who acted quickly on their information,
causing securities prices to properly and rapidly adjust to correctly reflect current
knowledge.
I arrived on this scene with a unique perspective. In 1959-60, I had discovered
that the casino game of blackjack could be beaten, and I devised and demonstrated a
mathematical system to do so, based on keeping track of which cards had been played.
Announced in December of 1960 and in January of 1961 (Proc. N.A.S.) and published in
detail in my Beat The Dealer (1962; revised 1966), the system showed that the blackjack
“market” was “inefficient.”
In a similar investigation of other gambling games, I discovered how to beat
roulette by physical prediction (1955-61) and, with Claude Shannon (of Information
Theory fame) built the first wearable computer (1961), whose function was to
successfully predict roulette outcomes. The predictions of the computer gave us the
huge positive expectation of 44%. Shannon and I then used the computer successfully
in Las Vegas to win small sums. The casino gambling “market” had yet another
“inefficiency.” For more, see Thorp (1969, 1998) and
http://wearables.www.media.mit.edu/wearables/. Click “history,” then click “a brief
history of wearable computing timeline.”
I investigated several other gambling games with some additional successes,
and by 1964 I began to consider the greatest gambling game of all time, the stock
market. Whereas I thought of card counting in blackjack as a million dollar idea, my
stock market explorations would lead to a hundred million dollar idea. Here’s a brief
history of what happened.

1964
I spent an intensive summer introducing myself to the markets, investments and
finance.

1964-1965
I discovered, and in some cases rediscovered, curious minor technical patterns
in stock prices.

1965
 At the start of the summer I resumed intensive study of the market and
discovered warrants. I understood at once that there was a qualitative link between the
price behavior of a warrant and its underlying stock and that it could be more or less
quantified and that it could likely be an exploitable market inefficiency. (See Chapter 2,
Beat the Market).

1965-1966
I met economist Sheen Kassouf in the fall of 1965 when we both arrived as
founding faculty members at the new Irvine campus of the University of California.
Kassouf was already hedging warrants and we collaborated, publishing much of our
work in Beat the Market. We understood static hedging, dynamic hedging, and delta
hedging, in particular market neutral delta hedging. We documented a historical market
inefficiency: warrants with two years or less to expiration tended to be very overpriced.
We used a static hedge to simplify historical simulations for the reader. We used
dynamic hedging in our own portfolios, making adjustments in moderate sized steps
both to limit risk and to limit the cost of doing so.
In practice my hedge ratio, or “delta,” was determined both by the macrostructure
of the payoff (wide range of protection against large price changes in either direction)
and the microstructure (market neutrality against small changes). It became increasingly
clear to me as I thought and invested during 1966-1968 that the clearest proof of an
exploitable market inefficiency is to construct a low (ideally zero) risk (hedged) package
that has little covariance with the market, yet produces a substantial “risk adjusted
excess return.”

1967
Beat the Market appeared. Kassouf and I ended our collaboration with the
completion of Beat the Market in late 1966, and went our separate business ways. I had
become familiar with Cootner’s 1964 book and various warrant valuation models based
on integration. I had earlier concluded that using the log normal distribution for stock
price changes and computing the expected terminal value of a warrant, led to a
reasonable candidate for a warrant formula. I learned from reading Cootner that several
people had already attempted this, in various ways. The version I liked, which I give in
Thorp (1969), was, as I learned only in 1975 from Larry Fisher, first derived by Boness.
However I do not see this in the Boness paper that is in Cootner, nor elsewhere in my
copy of Cootner (the original 1964 hardcover), but did note Sprenkle’s related attempt
(Cootner page 466).
“My” formula had a growth rate M for the stock and, in the case where the
warrant or option short sale proceeds are credited to the account at once, a discount
rate d for the present value.
Note: M = m + v 2 2 where m is the log normal drift parameter and v is the
volatility. E ( S (t ) ) = S (0) exp( Mt ) is the expected value of the stock at time t if S (0) is
the initial price.

1967-1968
I puzzled over the two parameters M and d , and speculate that in a risk neutral
world I can set them both equal to r , the riskless rate corresponding to the time until
expiration (which, as it happens, gives the future Black-Scholes formula). I also note
that a continually adjusted “delta neutral” hedged portfolio is riskless and so should have
its value discounted at the rate r. This also suggests the simple idea of just setting M
and d both equal to r. I mention a couple of these 1967 arguments long after the fact in
Thorp (1975).
Occam’s razor and these plausibility arguments suggested to me that if there
was a simple formula, this (the future Black-Scholes formula) is it.
I didn’t see how to prove the formula but I decided to go ahead and use it to
invest, because there was in 1967-68 an abundance of vastly overpriced (in the sense of
Beat the Market) OTC options. I used the formula to sell short the most extremely
overpriced. I had limited capital and margin requirements were unfavorable so I shorted
the options (typically at two to three times fair value) “naked,” i.e. without hedging with
the underlying stock. As it happened, small company stocks were up 84% in 1967 and
36% in 1968 (Ibbotson), so naked shorts of options were a disaster. Amazingly, I ended
up breaking even overall, on about $100,000 worth of about 20 different options sold
short at various times from late ’67 through ’68. The formula has proven itself in action.
I estimated volatility from the values of log( H / L), where H and L are the
monthly highs and lows, using the last twelve months which were easily and quickly
obtained from my library of S&P monthly stock guides. A previously published math
paper (Anderson, 1951?) gives properties of this volatility estimator, including expected
value and standard deviation.
My recollection as I write now is that E log( H / L), the expected value of H / L ,
was SQRT (8 / PI ) ∗ SIGMA where SIGMA is the volatility parameter in the assumed
underlying lognormal distribution. The use of the volatility estimator log( H / L)
prefigures a 1977? paper on this by Parkinson and a 1980 paper by Garman. Garman
combines this and other volatility estimators into a grand overall estimator.

1967-68
Comment: My naked calls sold short were probably the world’s first actual
investment use of the future BS formula.

END – Column 1.
References

[1] Black, Fisher and Myron Scholes. “The Pricing of Options and Corporate
Liabilities.” J. Political Economy 81, 1973, 637-54
[2] Cootner, Paul (editor), The Random Character of Stock Market Prices, M.I.T.
Press, 1964.
[3] Thorp, Edward O. “Fortune’s Formula: A Winning Strategy for Blackjack.” Notices
of the American Mathematical Society, Dec. 1960: 935-936.
[4] _____. “A Favorable Strategy for Twenty-One.” Proceedings of the National
Academy of Sciences, 47, No. 1, (1961) :110-112.
[5] _____. Beat The Dealer. New York: Random House, 1962.
[6] _____. Beat The Dealer, 2nd Ed., New York: Vintage, 1966.
[7] _____. “Optimal Gambling Systems for Favorable Games.” Review of the
International Statistical Institute, 37, 1969, 273-293.
[8] _____. “The Invention of the First Wearable Computer.” Proceedings of the
Second International Symposium on Wearable Computers, Pittsburgh, PA,
October 19-20, 1998.
[9] Thorp, Edward O. and S.T. Kassouf. Beat the Market. New York: Random House,
1967
Option Theory: What I Knew and When I Knew It – Part 2
by Edward O. Thorp
Copyright 2001

In November 1969 l and a partner, Jay Regan, launched what I believe was the
world’s first market neutral hedge fund. We called it Convertible Hedge Associates
(CHA), and later changed its name to Princeton-Newport Partners (PNP). It used
warrants, OTC options and convertible bonds and preferreds, along with the underlying
common stock, to construct delta neutral dynamically adjusted hedges. (Listed options
and publication of the Black-Scholes formula were still almost four years in the future).
Since “the formula” was, to me, highly plausible but not proven, we used in
addition a variety of techniques and screens, all of which the proposed mispriced
security had to pass:
[1] the formula (available for options and warrants, suitably modified for when and to
what extent the economic value of short sale proceeds are actually available;
generally not until expiration; in those days the brokers pocketed it.)
[2] scatter diagrams of prices: derivative versus stock or derivative versus derivative,
over time (e.g. Figure 2.2 of Beat the Market).
[3] cross-sectional scatter plots on standardized coordinate diagrams (at a fixed time,
such as that day’s closing prices) to compare derivatives within a class (e.g.
Figure 10.2 of Beat the Market).

1969-1972
The first market neutral hedge fund, which consisted of a collection of derivatives
hedges, each of which was (dynamically) approximately delta neutral, prospered. See
track records in (Thorp, 71, 75, 00).

Early 1973
The CBOE announced it would soon begin trading exchange-listed options. We
at Convertible Hedge Associates were electrified (figuratively and literally) by the news.
This could facilitate a major expansion of our business. I had an HP 9830A desktop
computer which was easy to program in BASIC, was math user friendly, and drove a pen
plotter with which we drew magnificent color coded graphs.
I had the “integral formula” programmed and drawing option and warrant curves
when, out of the blue I got a letter and an article from someone called Fisher Black. He
said “I am an admirer of your work” and explained that his approach was like Beat the
Market but he and Scholes took another step: they explored the (analytic)
consequences of the no arbitrage principle as applied to our (dynamically adjusted) delta
neutral hedge, noting that such a hedge should then return the (appropriate time period)
riskless rate on net equity invested.
I sat down, programmed in his formula and drew option curves. Shock! The
graphs disagreed with my graphs. It couldn’t be. But then I realized I was graphing the
“short warrants or options, long stock” version of my formula. This version assumed that
the interest from the proceeds from the short sale of the warrant or option was captured
by the broker, not the investor, as was the practice then for the warrants and over the
counter options which I had been trading. But the proceeds from shorting listed options
(only calls on a limited number of large companies were initially available when the
C.B.O.E. opened in 1973) would be credited to the investor on settlement date for the
trade. Thus one needed to pre-multiply by exp( − r (t ∗ − t ) ) to discount the expected
terminal value of the warrant to expected present value. Now the graphs were identical!
1973
What I already had, in fact, was not just the Black-Scholes formula but a more
general pair of formulas, with the Black-Scholes formula as the limiting case. One of
them incorporated a parameter to account for the loss to the broker of some or all, as the
case might be, of the interest earned by the short sale proceeds (SSP interest) on the
warrant (or option) short versus stock long hedge. This family of curves started with the
Black-Scholes curve (all SSP interest available) and moved continuously higher as the
fraction of available SSP interest dropped, with the highest curve being my old warrant
curve, corresponding to no SSP interest available.
The other formula covered the warrant (or option) long versus stock short hedge.
This one parameter family started with the Black-Scholes curve and had successively
lower curves as the economic value of the stock SSP interest available to the investor
was reduced.
The equations for highest and lowest curves are presented in Thorp (1973). That
was written a few weeks after I got the Black-Scholes paper and was immediate
because I already knew these formulas. In the original version of Thorp (1973) I had a
section showing how I had found the Black-Scholes formula by setting M and d equal to
r in the formula for the expected value of the warrant or option (as discussed in the
previous column). But I had to delete this to fit my abstract into the spaced allowed.
The two formulae create a “band” around the Black-Scholes value, within which
the delta neutral hedger cannot expect to achieve the riskless rate. This band widens
further when one adjusts the required pair of stock and option (or warrant) prices to
cover (expected) transactions costs, present and future.
As years passed, industry practice changed with competitive pressures and
investors tended to gain some of the interest from their short sale proceeds, splitting this
economic benefit with their broker-dealer. Currently, in the U.S. some hedge funds and
other institutional investors get an interest credit equal to Fed Funds (a proxy for the
“riskless rate” r ) minus seventy five basis points (0.75% annualized) or better. So the
pair of one parameter families has remained relevant. Yet, even today they have not, as
far as I know, been discussed in the literature. This is curious, given their practical value
for so many users of the Black-Scholes formula.

1973
Planning ahead for the opening of the CBOE, I had prepared a catalog of
standardized call option diagrams (see Beat the Market, chapter 6 for standardized
variables), of (option price)/(exercise price) versus (stock price)/(exercise price). For
stocks which paid no dividends during the life of the call option, for each of a range of r
and v (volatility) pairs there was one set of curves for various times until expiration.
These “universal” Black-Scholes curves covered all cases where our hedge was short
CBOE listed calls (full cash credit at once for SSP) versus the underlying common stock
long. We knew how to use numerical methods to calculate correct values for the option
price in cases where the stock paid dividends during the life of the option, but is was
usually sufficient to use easy approximations which covered most cases and could be
incorporated as a quick correction directly on the graphical plot.
Remember, this was 1973 when computing power was comparatively limited,
scarce and expensive. With market prices continually changing and the number of
options expanding rapidly, plus the need to monitor a substantial list of warrants and
convertibles, graphical short cut methods were valuable in this era. We simply plotted
the latest recent (stock, option) price pairs on the appropriate r, v diagram and looked to
see whether it was far enough above or below the appropriate curve to offer a profitable
hedge. Delta, the hedge ratio, corresponded to the slope of the tangent and could be
immediately read off the picture. We expanded the r, v catalog of diagrams as needed.
Option Theory: What I Knew and When I Knew It – Part 3
by Edward O. Thorp
Copyright 2001

For the reverse case of stock short and options (or warrants) long, we used one
of my formulas extending the Black-Scholes model to draw another catalog of curves,
also correcting for dividends as needed. As computing time allowed, we drew custom
graphs for the more important individual situations. For instance, I have from the fall of
1973 three separate pages of custom curves for hedging Chrysler warrants, accurately
corrected for dividends. One set covered stock long versus warrants short, the second
dealt with stock long versus listed call options short, and the third showed stock short
versus options or warrants long. These latter did not allow for early exercise, so they
were lower bounds only. We knew how to numerically solve the problem for early
exercise but didn’t because our limited computer power was better used elsewhere.

1973
After seeing the Black-Scholes derivation, I explored the power series approach
in 1973-74. After seeing my write-up in Thorp (76), Black told me that they were already
aware of some of the power series ideas in 1969-70.

1974
Using power series, I developed what we dubbed internally as DOP, the
“diversified option portfolio.” This extended and generalized the idea of delta hedging
and measured the risk of arbitrarily complex hedges that were constructed from any or
all of the available derivative securities (e.g. options, warrants, convertibles) on a single
underlying common.
We tried to achieve excess expectation while minimizing the exposure to the
various power series terms. We considered not only delta neutrality but we also reduced
our exposure to what people later called “the Greeks,” and for which they later used
names like gamma, vega, theta, etc. Again, this idea was well ahead of the literature
and, the important point for us, gave us an edge over the practitioners with whom we
competed.

1973 fall
Using the integral method, we easily solved and programmed the numerical
calculation of warrant values for dividend-paying stocks. We had been generating
custom graphs for each listed warrant and approximating the dividend correction. Now
we had it exactly.

1974
We discovered the analytic solution (in terms of multivariate normal distributions)
to the problem of call options on dividend-paying stocks, assuming that early exercise is
not optimal (which is true unless the dividends are “large” or one of them is “close” to the
expiration date). However we preferred our numerical solution because it covered all
cases and was computationally efficient.
Robert Geske discovered and published (1979, 1981) the same analytic solution
in what, if I recall correctly, he called his “compound option model.”

1974 fall
We were told that the CBOE would start trading American puts “soon.” After the
Black-Scholes formula, this was the central unsolved problem in option theory. Because
of the ease and power of the integral model, I was able in an hour to conceive and
outline the solution to the problem for my programmer. It used recursive numerical
integration of the log normal probability density function for the stock, using the
appropriate riskless rate for the expected growth rate and for the discount to present
value, as described earlier. All the boundary conditions were incorporated and the time
and space steps used for the backward integration were chosen fine enough to give the
desired accuracy. This solution was complete: it incorporated dividends and
determined the early exercise region. We drew graphs and printed tables.
At a one-on-one dinner meeting to which Fisher Black invited me on May 14th or
15th, 1975, just prior to the University of Chicago Center for Research in Securities
Prices (CRSP) meeting in Chicago, I brought along my solution to the American put
problem and had placed a folder of graphs on the table to show him. Then he said no
one had solved the problem, and asked what I thought about the “numerical solution to
p.d.e.” approach. While I was giving my view that the approach worked but one had to
be careful in choosing the sizes and relative sizes of the time and space steps (I had
already looked at it and seen how to do it, but chose instead to use the integral method
as “easier.”) I realized I had a fiduciary duty to my investors to keep our secrets, and
quietly put my folder with the world’s first American put curves back in my briefcase.
Schwartz and Parkinson each published solutions in 1977 that were “cousins” to
our integral method version.

1974
In my classes at U.C. Irvine, I taught that there were three roads to option theory:

[1] The integral model, “all powerful” for producing numerical (and some analytical)
solutions.
[2] Stochastic differential equations (Black-Scholes 1973, Merton 1992), the most
elegant and technically demanding approach. Useful for producing analytic
formulas when they existed but they did not always lend themselves as easily to
solving problems numerically.
[3] Power series: very useful for special problems.

I didn’t “teach” a fourth method, the Monte Carlo approach, since it seemed
obvious and, though appropriate for solving various options problems, had much wider
applicability.
Later, Bill Sharpe’s suggestion led the finance world to the development of a fifth
method, the binary model. This is closely related to the integral model, just as discrete
binary random walks in the limit tend to Brownian motion.

1975-1985
We were able to stay ahead using the integral model, then later converting to the
binary model for greater computational speed, but as the financial literature advanced
much of our theoretical edge in option theory slowly vanished. The Black-Scholes
methodology revolutionized finance, “everyone” adopted it, and listed option spreads
narrowed until only the competitors with the lowest costs could still extract excess risk
adjusted returns. However in practice we were able to stay ahead in derivatives trading
through our computer programs and applications, especially in convertible bonds and in
the analysis of an expanding crop of new derivative products.
There is a lot more to this account. I have cartons of rough notes, and research
ideas which I explored, both in derivative theory and practice and in other areas of
finance. Keeping what we discovered secret, while benefiting from ongoing published
academic work, was a major factor in producing some $250 million in profits for
CHA/PNP.

2000
Convertibles and other derivative hedging is still profitable. The derivatives
based market inefficiencies exploited in Beat the Market have expanded vastly in
number and size and account for a significant part of today’s hedge fund industry. A
salient example is the $8 billion Chicago based Citadel Investments with its domestic
and offshore partnerships. Under its principal general partner, Kenneth Griffin, now in
his early thirties, it recently celebrated its tenth anniversary with a 30% annual
compound rate of return for the decade. It’s widely considered to be the most valuable
hedge fund business in the world.
Another concept, first discovered at Princeton-Newport in December 1979 or
January 1980, is the core idea of what is now called statistical arbitrage. The more
primitive “pairs trading” had already been discovered at PNP and used in minor ways in
the late 70s. Based on the approximately Brownian motion structure of stock prices, the
idea has led to a set of techniques for “draining energy” (i.e. money) from the
ceaselessly excessive (see Schiller) fluctuations in stock prices.

Note on the integral model:

The key is the observation that growth and discount rates can all be set to r , the
riskless rate. Cox-Ross later proved in 1976 that this is correct (see Merton, 1992, pp
334ff). As soon as I saw the Black-Scholes proof in 1973, I felt certain that this
consequence of their result applied not only to call options but generally to all derivative
problems using log normal diffusion for the underlying security. I then immediately
implemented this in numerically solving by iterated numerical integration, backward from
the terminal value, in “small” time steps. Fortunately for me, this method was ours alone
to practice from 1973 to 1976, and even after the Cox-Ross proof, we didn’t know of
other practitioners who adopted it.
References

[10] Thorp, Edward O. “Portfolio Choice and the Kelly Criterion.” Proceedings of the
1971 Business and Economics Section of the American Statistical Association,
1971, 215-224. Reprinted in Stochastic Optimization Models in Finance, edited
by W.T. Ziemba, S.L. Brumelle, and R.G. Vickson, Academic Press, 1975, 599-
620.
[11] _____. “Extensions of the Black-Scholes Option Model.” Contributed Papers 39th
Session of the International Statistical Institute, Vienna, Austria, August 1973,
1029-1036.
[12] _____. “Options in Institutional Portfolios, Theory and Practice.” Proceedings,
Seminar on the Analysis of Security Prices, Volume 20, No. 1, May 15-16, 1975,
229-252. Center for Research in Security Prices, Graduate School of Business,
University of Chicago.
[13] _____. “Common Stock Volatilities in Option Formulas.” Proceedings, Seminar
on the Analysis of Security Prices, Vol. 21, No. 1, May 13-14, 1976, 235-276.
Center for Research in Security Prices, Graduate School of Business, University
of Chicago.
[14] _____. “The Kelly Criterion in Blackjack Sports Betting and the Stock Market.”
Finding the Edge: Mathematical Analysis of Casino Games, eds. Olaf Vancura,
Judy A. Cornelius and William R. Eadington, University of Nevada, Reno Bureau
of Business & Economic Res., 2000.
[15] Thorp, Edward O. and S.T. Kassouf. Beat the Market. New York: Random House,
1967.