Entropic Derivative Security Valuation

1
Michael Stutzer
Professor of Finance and Director

Burridge Center for Securities Analysis and Valuation

University of Colorado, Boulder, CO 80309

1 Mathematical Framework

Let x denote a (vector) of random variable(s), β a (vector) of parameter(s), and f (x, β) a (column

vector) of real-valued function(s). Let EP [f (x, β)] denote its expectation computed with probability

measure P . For a specific value of β, define the set of probability measures:

(1) P(β) ≡ {P : EP [f (x, β)] = 0}

which additionally are absolutely continuous with respect to a distinguished measure µ that is

determined by the application. Selection of a particular probability measure in P(β) is called a

linear inverse problem. Many applications of entropy make use of the following solution (when it

exists):

Z
(2) min D(P kµ) ≡ min log(dP/dµ)dP s.t. EP [f (x, β)] = 0.
P ∈P (β) P

R
In (2), D(P kµ) = log(dP/dµ)dP is the Kullback-Leibler measure of discrepancy between the

measure P and the distinguished measure µ, a.k.a. the relative entropy. When µ is a discrete
PH
probability measure with H possible values, D(P kµ) = h=1 Ph log(Ph /µh ), and it can be shown

1
that D ≥ 0, with equality when and only when Ph = µh , h = 1, . . . , H. When in addition

µh ≡ 1/H, i.e. the uniform distribution, the constrained minimization of D(P kµ) is equivalent to

maximization of the Shannon entropy − 2

The solution to (2) is well-known [5, sec.3(A)] to have the following Gibbs Canonical or Esscher

Transformed density:

e−γ(β) f (x,β)
0
dP (β)
(3) = .
dµ Eµ [e−γ(β)0f (x,β)]

To compute the coefficient vector γ(β) in (3), solve the following problem :

0
(4) γ(β) = arg max I(β, γ) ≡ − log Eµ [e−γ f (x,β)]
γ

Finally, D(P (β)kµ) = I(β, γ(β)) defined above. This numerical value has a frequentist interpreta-

tion from the large deviations theory of IID processes that is quite useful in asset pricing model

parameter estimation [11] and optimal portfolio choice [14].

2 Application to Derivative Security Valuation

Consider the simplest problem of option pricing: value a European call option, written on a single

underlying stock that pays no dividends, whose price at expiration T -periods ahead is denoted

x(T ). The riskless, continuously compounded gross interest rate r is constant between now and

expiration. Under the familiar assumptions of complete and frictionless markets that do not admit

arbitrage opportunities, there is a risk neutral probability measure P under which the call option’s

price C is the expected value of its risklessly discounted payoff at expiration, i.e.

2
 h i
(5) C = EP max[x(T ) − K, 0]/r T

)
where the risk-neutral probabilities P satisfy the martingale constraint x(0) = EP [ x(T
rT
], rewritten

x(T ) T
(6) EP [ /r − 1] = 0
x(0)

Conventional risk-neutral pricing of options proceeds by specifying a parametric model for the

risk-neutral stochastic price process of the underlying stock. Parameter values are found that make

the model’s computed stock prices and/or option prices close (e.g. in the least squares sense)

to observed stock and/or option prices [3]. In the simplest case (the Black-Scholes model), this

procedure requires an estimate of the volatility parameter, found either from past stock returns

(i.e. historical volatility) or from market option prices (i.e. a best-fitting implied volatility).

But suppose one has doubts about the correct parametric model. The formalism of the previous
x(T ) T
section provides an alternative. Let the scalar function f (x, β) = x(0) /r − 1, where β = r. The

distribution µ is the forecast distribution of x(T ). To estimate this, one could just use a histogram

of past T -period stock returns as in [13] and [16], a conditional histogram [15], or a more complex

forecasting model [7]. The distribution is then substituted into (4) and solved to find the γ(β)

needed to estimate the density P (β) in (3), which is required to compute the option valuation (5).

It is possible to extend the approach to handle stochastic dividends and interest rates.

Another approach presumes a particular form for µ, and defines a vector f (x, β) with ith

component max[x(T ) − Ki, 0]/r T − Ci , where Ci is the observed market price for a call with exercise

price Ki, as in [10] and [4]. Then, (3)-(4) are used to study the nature of the measure P (β) implied

by those options’ market prices, while (5) could be used to value options other than those present

in f . See [2] for some theoretical and applied extensions of this approach.

3
 Further refinements were developed by Gray, et.al [9], showing that the associated dynamic

hedge (i.e. entropic hedge ratio) outperformed hedging benchmarks, and by Alcock and Carmichael

[1], extending the concept to enable valuation of American options.

The entropic approaches are not necessarily inconsistent with the conventional approach. In

fact, extant closed form option pricing models can be analytically derived by specification of the

process generating the stock price distributions, and systematically applying the Esscher Transform

(3) as in [8] and [6].

4
Notes

1 The limited length of this entry precludes writing a survey describing all uses of entropy in

finance and that cites all papers using it. Instead, focus is placed on its use in derivative security

valuation, arguably the most relevant application for readers of this book, and citations are limited

to early or illustrative papers.

2 Perhaps the first published use of this in finance was in Osborne [12], who rationalized as-

sumption of a lognormal stock price distribution as that maximizing Shannon entropy subject to

a vector of two constraints (1), constraining the mean and variance of the observed continuously

compounded returns to observed values.

5
References

[1] J. Alcock and T. Carmichael. Nonparametric American option pricing. Journal of Futures

Markets, 28:717–748, 2008.

[2] Marco Avellanada. Minimum relative-entropy calibration of asset-pricing models. International

Journal of Theoretical and Applied Finance, 1:447–472, 1998.

[3] David S. Bates. Testing option pricing models. In G.S. Maddala and C.R. Rao, editors,

Handbook of Statistics, Volume 15: Statistical Methods in Finance, pages 567–611. North-

Holland, Amsterdam, 1996.

[4] Peter W. Buchen and Michael Kelly. The maximum entropy distribution of an asset inferred

from option prices. Journal of Financial and Quantitative Analysis, 31:143–159, 1996.

[5] Imre Csiszar. I-divergence geometry of probability distributions and minimization problems.

Annals of Probability, 3(1):146–58, 1975.

[6] E. Eberlein, U. Keller, and K Prause. New insights into smile, mispricing, and value at risk:

The hyperbolic model. Journal of Business, 71:371–405, 1998.

[7] F. D. Foster and C.H. Whiteman. An application of Bayesian option pricing to the soybean

market. American Journal of Agricultural Economics, 81:222–272, 1999.

[8] Hans Gerber and Elias Shiu. Option pricing by Esscher Transforms. Transactions of the

Society of Actuaries, 46:99–140, 1994.

[9] P. Gray, S. Edwards, and E. Kalotay. Canonical pricing and hedging of index options. Journal

of Futures Markets, 27:771–790, 2007.

6
[10] R.J. Hawkins, Mark Rubinstein, and G.J. Daniell. Reconstruction of the probability density

function implicit in option prices from incomplete and noisy data. In K. Hanson and R. Silver,

editors, Maximum Entropy and Bayesian Methods. Kluwer, 1996.

[11] Yuichi Kitamura and Michael Stutzer. Connections between entropic and linear projections in

asset pricing estimation. Journal of Econometrics, 107:159–174, 2002.

[12] M.F.M. Osborne. Brownian motion in the stock market. In Paul Cootner, editor, The Random

Character of the Stock Market. MIT Press, 1970.

[13] Michael Stutzer. A simple nonparametric approach to derivative security valuation. Journal

of Finance, 51(4):1633–1652, 1996.

[14] Michael Stutzer. Portfolio choice with endogenous utility: A large deviations approach. Journal

of Econometrics, 116:365–386, 2003.

[15] Michael Stutzer and Muinul Chowdhury. A simple nonparametric approach to bond futures

option pricing. Journal of Fixed Income, 8:67–76, 1999.

[16] Joseph Zou and Emanuel Derman. Strike adjusted spread: A new metric for estimating the

value of equity options. Quantitative Strategies Research Note, Goldman, Sachs and Co., July

1999.