Elasticity theory of structuring∗

Andrei N. Soklakov†
arXiv:1304.7535v6 [q-fin.GN] 12 Dec 2016

Financial derivatives have often been criticized as casino-style betting instruments.

It turns out that many naive ways of making them are indeed equivalent to gambling.
Fortunately, this inadvertent effect can be understood and prevented. We present a
theory of product design which achieves that.

1 Introduction
In [1] we reviewed the shortcomings of the pre-crisis approach to product design and
advocated the need for a better, more quantitative, approach. We started to build this
approach in [2] and provided further illustrations in [3].
So far, most of our examples have been centered around the important special case of
the growth-optimizing investor. This investor defines growth in terms of the compounded
rate of return (logarithmic rate of return) and seeks to maximize the expectation of this
rate regardless of the risks. First introduced by Bernoulli in 1738, this is probably the
oldest and one of the best researched benchmarks of investment behavior [4, 5].
With the growth-optimizing (very risky) investor on the one hand and the risk-free (bond)
investor on the other hand, we begun to understand the range of products which most
sensible investors would want to use [2]. In this paper we refine our framework allowing
investors to control their risk appetite in a more precise way.
Combining Bayesian information processing with rational optimization gives us a simple
new tool for product design – the payoff elasticity equation. This is the main technical
result of the paper. After some demonstrations of the equation, including detection and
prevention of inadvertent gambling, we summarize our approach to product design as a
self-contained manufacturing process. In terms of infrastructure, this is fully compatible
with existing business environments.

∗
Original version 23 April 2013. Journal-ref: Risk, December (2016), pp. 81-86.
†
Head of Strategic Development, Asia-Pacific Equities, Deutsche Bank.
The views expressed herein should not be considered as investment advice or promotion. They represent
personal research of the author and do not necessarily reflect the view of his employers, or their associates
or affiliates. Andrei.Soklakov@(db.com, gmail.com).

1
2 Deriving derivatives
Let x be a random variable with some financial significance – this can be anything from a
stock price or an average commodity value to temperature readings in Texas. The variable
x can be a vector, containing several such examples as components. Vector-valued x can
also be used to hold time series of a market variable (see [3] for a concrete example).
Financial product is called a derivative of x if its payoff F is defined as a function of x.
Behavioral Finance teaches us to be very careful when using intuition in finance. In terms
of payoff functions, this means that not every ad-hoc proposal for F is good; not even
when it is clear and well-defined. Superstitions and compulsive disorders often manifest
themselves as transparent, well-defined procedures which lead people to disastrous results.
The best constructive way of supporting human intuition is to use the scientific approach.
In the case of financial derivatives it begins with making sure that investment structuring
is guided by the logical laws of information processing.

3 Investor equivalence principle

In this presentation we consider the simplest case of a real-valued underlying, x ∈ R.
This assumption is not crucial in any way and we use it purely for notational clarity.
Following [2], we partition the range of possible values of x into non-overlapping intervals
using a discrete mesh (. . . , xi , xi+1 , . . . ). Imagine now a set of securities {si }, where each
si pays 1 when x fixes between xi and xi+1 and zero otherwise. We can get a quote for
purchasing the securities, pi = quoted price(si ). It is tempting to call {si } the Arrow-
Debreu securities and connect (undiscounted) {pi } to the risk-neutral probabilities for x.
Let us resist this temptation as it would force us to make a lot of unnecessary assumptions.
At this point we do not require a two-way liquid arbitrage-free market of securities for
each distinct state of the economy. All we need is someone who would be willing to sell
{si } – a set which may contain just a couple of securities.
In [2] we introduced {si } as binary spreads and argued that for a very large class of
investors the problem of optimal investment is equivalent to optimal splitting of capital
across the {si }.
Let
P {βi } be the proportions in which the investor decides to partition their capital
( i βi = 1). Only one security among {si } can mature in the money, so the payoff
F from the investment is easy to compute

Fk = βk rk , (1)

where k is the index of the security that matures in the money and rk = 1/pk is the
quoted return on that
Psecurity. We define the market-implied distribution by normalizing
the prices mk = pk / i pi and rewrite the above equation in the “Bayesian” form

βk = Fk mk , (2)

where we decided not to burden our notation with trivial normalization constants (this
can be arranged formally by re-defining Fk , or by leaving mk not normalized).

2
Equation (2) covers all investment scenarios which can be broken down into mutually
exclusive events {si }. Given that the choice of the underlying variable remains completely
up to the investor this is a very general setting. Indeed, not only can we use x to mean
anything we want (from exotic strategies to weather readings), the choice of βk in (1) is
also very general – so far we didn’t even assume the investor to be rational in any way.
Equation (2) is just basic accounting. In the following we use Eq. (2) as a property which
defines any investor .
Let {bk } be investor-believed probabilities that the corresponding securities {sk } mature
in the money. In [2] we considered a special case of a growth-optimizing investor. In this
special case the fractions {βk } coincide with {bk } [6]. Equation (2) becomes

bk = fk mk , (3)

where fk denotes the growth-optimal payoff. We revisit this case in more detail later.
Although growth-optimizing investors form a very specific group, we see that focusing on
this group does not at all reduce the range of possible investment decisions. Indeed, both
{βi } and {bi } must add up to one, but with no further constraints, Eqs. (2) and (3) are
solved by the same set of payoffs, {f } = {F }. By pairing up these payoffs we can now
make an even more detailed observation, which, in view of its generality, we formulate
here as a principle:

Any investor can be viewed as growth-optimizing.

More precisely, general investor chooses the same product, F = f , i.e. behaves in the same
way, as some growth-optimizing investor whose beliefs {bk } happen to coincide with {βk }.
The principle focuses our attention on the actions of the investor – their net economic
behavior. The logical integrity of these actions can be checked in many ways. The
principle suggests doing this from the point of view of a growth-optimizing investor – a
very convenient choice.

4 Logical investors
In [2] we made a connection between Eq. (3) and Bayes’ theorem. We recognized the
market-implied {mk } and the investor-believed {bk } as the prior and the posterior dis-
tributions and highlighted the fundamental connection between the optimal payoff struc-
tures and the likelihood functions. This gave us some understanding of how the growth-
optimizing investors learn and then express their knowledge through trading.
In the previous section we also discovered that, in terms of their net economic behavior, the
growth-optimizing investors are surprisingly versatile. It looks like they can be educated
to act as if they had sophisticated risk preferences. The aim of this section is to describe
and understand this education process mathematically. To this end we narrow the scope
of our investigations and focus on the case of rational investors.
It can be shown that rational investors behave as if they were maximizing the expected
value of utility [7]. The expected utility approach which follows from this observation is
well known in economic theory. We will use it here as well. It is important, however, to

3
emphasize that by adopting the expected utility approach we do not reduce the generality
of our arguments in the same way as it happens in economic theories. For us, rationality
is not an assumption – it is part of the goal. We understand that being rational is not
easy, so we build tools which facilitate rational behavior.
Consider a rational investor with utility u() and a view on x given by the probabilities
{bi }, where each bi measures the degree of investor’s belief for x to end up between xi and
xi+1 . Because the logarithm is a monotonically increasing function, we can, without loss
of generality, write the utility as a function of the logarithmic rate of return. We also allow
the utility to depend on x explicitly. The optimal investment is found by maximizing the
expectation X  
bi u ln(βi ri ), xi (4)
i
P
over all possible proportions {βi } subject to the budget constraint i βi = 1. The La-
grangian for this optimization reads
X   X
L({βi }, λ) = bi u ln(βi ri ), xi + λ( βi − 1) . (5)
i i

By setting ∂L/∂βk = 0 we compute

 
bk u′ ln(βk rk ), xk = −λβk , (6)

where prime denotes derivative with respect to the first argument. This implies
X   X
bi u′ ln(βi ri ), xi = −λ βi = −λ . (7)
i i

Substituting this back into Eq. (6) we obtain

 
u′ ln(βk rk ), xk bk
βk = P   . (8)
i u ′ ln(β r ), x
i i i bi

The positivity of βk is guaranteed by the monotonicity of the utility function with respect
to return u′ > 0 (the limiting case of zero βk is trivial and can be removed from the
optimization problem (5)). Solving this equation for βk gives us the optimal investment
strategy. This includes the growth-optimizing investor as a special case. Indeed, in this
case u′ = 1, Eq. (8) reduces to the Kelly equation, βk = bk , which in view of (2), leads to
our basic equation (3).
Research in economics and finance is often done in the comfort of fully continuous settings
where the tools of calculus are more readily available. Although our approach does not
need that, it would be easier for us to see the emergence of classical concepts if we present
our main arguments in a similar fashion. So, substituting Eq. (1) into (8) and writing the
result in a fully continuous notation we get1
 
u′ ln F (x), x
β(x) = R   b(x) . (9)
u′ ln F (x̃), x̃ b(x̃) dx̃
1
Discrete treatment is reviewed in Appendix A.

4
 1 Learning process and investment structuring

m(x) b(x) β(x)

b(x) β(x)
f (x) = F (x) =
m(x) m(x)
Growth-optimizing General investor

We are now in a position to understand the logic of how our investor learns and how this
learning process translates into investment decisions.
The investor needs to learn two kinds of information. On the one hand, they need to
learn about the market. On the other hand, they need to learn about themselves (their
preferences and goals). It is a fundamental property of Bayesian learning that, given a set
of data, it does not matter in which order the various parts of this data are included in the
calculation [8]. Indeed, conditional probabilities P (x|A, B) and P (x|B, A) are numerically
equal. So we have a lot of freedom in how to organize our calculations.
Great, but how do we start? This question is infamous in statistics as the problem of
choosing the prior. In general, choosing the prior can be very hard and the end solution is
often ad-hoc. We are lucky, however, to have a natural starting point. This is a growth-
optimizing investor who chooses m(x) as the prior information on the market. Why
growth-optimizing? Because of the investor equivalence principle. Why m(x)? Because,
so far, this is all that we have given them – the prices {pi } of the securities {si }. We
emphasize that this is just a starting point of a calculation.
The investor is now in a position to include the rest of their data. They discover related
markets, analyze historical records. At this point they may notice differences in drifts
implied by their research vs m(x). In some such cases, they may want to speak in
terms of “real” and “risk-neutral” worlds to articulate their findings. The reader has
almost certainly witnessed such conversations in practice. Eventually the investor stops
researching the markets and writes down their believed distribution b(x).
A true growth-optimizing investor would normally stop here and proceed directly to the
construction of the payoff via equation (3), as shown on Figure 1. In general, however,
the investor might have additional information about their personal preferences. The
investor needs to incorporate this information into the calculation given, of course, what
they already included about the market. The investor equivalence principle tells us how
we can describe this additional learning step mathematically.
Indeed, according to the investor equivalence principle, we can continue to view our in-
vestor as growth-optimizing as long as we can imagine persuading them to believe β(x).
It is important to emphasize that we do not simply imagine someone who already believes

5
β(x). We demand logical continuity: we started with a growth-optimizing investor, they
have already learned about the market, and now we want them to learn additional infor-
mation which would make them behave as if they had a different system of preferences.
In other words, we want to see if our investor can take their belief b(x) and update it once
again using Bayes’ theorem arriving at β(x). Equation (9) describes exactly this addi-
tional learning step and Figure 1 summarizes the whole process including the derivation
of the final optimal payoff F .
The above discussion leads us to two important observations. First, although not entirely
obvious from its structure, Eq. (9) is about incorporating information on risk aversion
into product design. In subsequent sections we further distill this idea and demonstrate
its power on practical applications. Second, we see that incorporating all information
that is relevant to investments (objective or personal) is clearly an important ability for
an investor. Those investors which can do that without contradicting the logic of Bayes’
theorem are especially important, so we give them a name. Let us call them logical
investors. In the Appendix B we provide additional context both to justify this name and
to give a better feel for the role such investors play in our theory.

5 Payoff elasticity equation

Dividing both sides of Eq. (9) by m(x) and using equations (3) and (2) we derive
 
u′ ln F (x), x
F (x) = R   f (x) . (10)
u′ ln F (x̃), x̃ b(x̃) dx̃

This is an integral equation for F (x). We want to convert it into a more practical differ-
ential form.
The concept of elasticity gives us a particular way of differentiating that proved to be
especially useful in economics and finance. Elasticity of a function, φ(x), with respect to
its argument, x, is defined as the derivative
d ln φ(x)
. (11)
d ln x
This measures the percentage change in the function’s value with respect to percentage
change in its argument.
The classical notion of utility does not depend on x explicitly [7]. In our notation this is
the case when the second argument of u happens to be redundant: u(ln F, x) = u(ln F ).
We investigate this case first and then discuss what happens in general. Taking the
logarithm on both sides of Eq. (10), forgetting temporarily about the explicit dependence
of u on x and differentiating we obtain
1
d ln F = u′′ (ln F ) d ln F + d ln f . (12)
u′ (ln F )
Note that x is the only variable quantity. The denominator on the rhs of Eq. (10) does
not depend on x and drops out from the above differential equation.

6
Rearranging the terms,
d ln F u′ (ln F )
= ′ . (13)
d ln f u (ln F ) − u′′ (ln F )
Thinking about practical applications, removing the explicit use of the highly theoretical
concept of utility from the above equation is our next challenge.
It is well known that affine transformations of the utility function do not have any effect
on investors’ preferences. Utility functions effectively ignore the two most fundamental
mathematical operations – addition and multiplication by a number. This fact limits
practical use of utility functions, especially when we want to talk about risk aversion.
To remedy the situation, a variety of measures for risk aversion were introduced. The
most popular of them are probably the Arrow-Pratt measures of absolute and relative
risk aversion:
U ′′ (F ) U ′′ (F )
A(F ) = − ′ , R(F ) = −F ′ , (14)
U (F ) U (F )
where U is the standard definition of utility which is connected to our slightly more general
notion, u, via the equation U(F (x)) = u(ln F (x)). By direct calculation we have

u′ (ln F ) u′′ (ln F ) − u′ (ln F )

U ′ (F ) = , U ′′ (F ) = . (15)
F F2
Using the definition of relative risk aversion (14), we can now rewrite Eq. (13) as

d ln F 1
= . (16)
d ln f R
This simple equation is the central technical result of this paper. It gives us a fundamental
link between payoff elasticity and risk aversion. The more risk aversion we have the less
elastic is the payoff.
On the practical side, this equation allows us to compute the optimal payoff F from the
growth-optimal f and the risk aversion profile R of the client. Conversely, we are now
also able to compute risk aversion profiles directly from clients’ positions.
For completeness, we mention some alternative forms of Eq. (16) which can be used
depending on the application. Considering the payoff elasticity equation (16) for two
different general investors we derive

d ln F (1) R(2)
= . (17)
d ln F (2) R(1)
This equation shows that other payoff profiles (not necessarily growth-optimal) can serve
us as building blocks. The concept of elasticity can be replaced by the ordinary differ-
entiation if we decide to work in terms of a less fundamental measure of absolute risk
aversion:
dF 1 dF (1) A(2)
= , = (1) . (18)
df fA dF (2) A
How would the above derivation change if we allowed for explicit dependence of u on x?
The payoff elasticity equation would stay the same but the expressions for both R and A
would become more complicated. In particular, they would acquire explicit dependence

7
on the state x (see Appendix A), but would of course reduce to the original Arrow-Pratt
definitions in the special case of state-independent preferences. On the practical level, the
important thing to mention is that essentially all of these equations have been solved and
present no further technical challenge. Indeed, the Picard-Lindelöf theorem provides the
necessary theory (for the one-dimensional case) and even offers an explicit construction
of the general solution (Picard iteration method).

6 Illustrations
In this section we illustrate practical usage of the payoff elasticity equation. We warm up
on a simple analytically solvable example and show how to examine structuring ideas that
do not necessarily come from our theory. We then turn to the main power of the payoff
elasticity equation – the ability to understand and to adjust clients’ risk aversion. We
show that naive attempts to express risk aversion can be dangerous and that the payoff
elasticity equation provides us with a more sound technology.

6.1 One-parameter investor families

By a one-parameter family we mean a set of investors whose degree of risk aversion is
controlled by a single number, e.g. constant absolute or constant relative risk aversion.
The usefulness of one-parameter families comes from the fact that the position of an
investor within the family can be determined by asking the investor a single question
regarding, for instance, their maximum acceptable loss. To illustrate this point, let us
look at Eq. (18) and consider the case when it takes a particularly simple form: we set
a
A= , (19)
f
where a is a constant which controls the strength of risk aversion. We immediately derive
1
F = (f − 1) + 1 . (20)
a
We see that a effectively scales the payoff around the bond line F = 1. This is exactly
how Figure 2 in [2] was constructed. The value of a can be easily found by matching
investors maximum acceptable loss.

6.2 Product validation

Can our theory accommodate every imaginable investor? Of course not. One can imagine
investors that demand greater flexibility than our equations would allow. As explained
in Appendix B, we should be very cautious in offering assistance to such investors as not
every desire is necessarily wise. At the moment it may be more prudent to focus on the
many unexplored possibilities that are already offered by our theory.
Having said that, it is very important to acknowledge that investment ideas are born in
all sorts of ways and good ideas may not necessarily come through an explicit use of our

8
theory. To recognize such cases we want the ability to verify that a given investment
strategy is both rational and logical.
In order to see how this can be done, let us examine the family of investors considered
in [9]. The investors are looking for a payoff structure, h(x), which solves a certain
optimization problem. Using our notation, the optimization problem reads:
hZ Ra
Z i
max h(x)b(x) dx − h2 (x)m(x) dx (21)
h 2
| {z } | {z }
mean variance
Z
subject to h(x)m(x) dx = 0 , (22)

where Ra is a parameter that controls the degree of risk aversion. In Ref. [9] this
problem was introduced by analogy with the mean-variance optimization approach of
Markowitz [10]. It is important to note, however, that the mean and the variance, given
by the first and the second term of (21) respectively, are computed using two very different
distributions. In particular, the mean is computed using the investor-believed b(x), while
the definition of variance is using the market-implied m(x). As a result, the optimization
setup (21-22) does not fit well into utility maximization paradigm, so, on its own, the
setup (21-22) is not enough to recommend the investment as rational. The idea of penal-
izing market-based variance does, however, make significant intuitive sense, so let us not
rush into dismissing it on formal grounds. Let us independently investigate in what sense
the above idea may point towards a sensible product.
In what follows we put aside any attempts to justify the optimization (21-22) and pro-
ceed by directly examining its solution, h(x). By doing so we put ourselves in a rather
typical situation when an investment product is proposed which looks good yet with some
questions regarding its real quality. Before we can recommend the product, we want to
verify the assumptions under which h(x) may be viewed as a rational strategy pursued
by a logical investor. The expression for h(x) is given by the first equation in [9], namely
b(x) − m(x) 1
h(x) = . (23)
m(x) Ra
By rearranging the terms we can rewrite this equation as
 
b(x) = Ra h(x) + 1 m(x) . (24)

This allows us to understand the investment through the eyes of a growth-optimizing

investor. Indeed, using equation (3) we derive

Ra h(x) + 1 = f (x) , (25)

where f (x) is the growth-optimal payoff structure. Finally, we see that

1  
h(x) = f (x) − 1 . (26)
Ra
We immediately recognize this equation as a particular case of the one-parameter investor
family which we considered above. Indeed, the only difference between this equation and

9
equation (20) is the absence of an additive constant – the upfront investment cost of 1 –
which drops out from (26) on the account of the constraint (22). This constraint assumes
the existence of a perfectly liquid two-way market in options on x which effectively allows
the investor to borrow from the market and set up the investment at zero upfront cost.
Such assumptions are very important to note and check. Other than that we managed to
verify that the solution of the optimization (21-22) can be viewed as a rational investment
pursued by a logical investor with an understandable risk aversion strategy.

6.3 Implying and adjusting risk aversion

Empirical computations of risk aversion have always been a very challenging task. We
now have additional tools. We can imply risk aversion directly from clients’ positions or
ideas, adjust it if necessary and use it to structure the optimal product.
To emphasize the immediate relevance of our techniques in the context of current practices,
let us review one of the most popular ideas in finance – expressing higher risk aversion by
implementing a more conservative view. This very simple idea is not confined to finance. It
almost certainly came into our field from other more basic human behaviors. It obviously
has powerful evolutionary reasons yet it is not without faults. Failing strategies, later
described as “half-measures” or “indecisions”, often begin as conservative interpretations
of data. So let us examine how robust this idea really is in the mathematical context of
financial structuring.
Imagine a growth-optimizing investor which believes that the skew should be half of the
value observed on the market. This view and the corresponding optimal payout are
depicted in red on Figures 2.A and 2.B. The investor decides to be more cautious and to
avoid expressing their view in the wings. To this end they manually adjusts their original
view by gently forcing it to revert back to the market in the wings (blue line on Figure
2.A). The corresponding payoff structure is given by the blue line on Figure 2.B.
We can now use the payoff elasticity equation and examine the risk aversion profile which
corresponds to investor’s attempt to soften their view. Before we do that, let us think
what we expect to see. Risk aversion profile for a growth-optimizing investor is flat R = 1.
So, surely, what we should see is a profile that lies above the line R = 1 increasing further
in the wings.
Looking at Figure 2.C we see that this intuition could not be further from the truth. Not
only the profile does not form a convex line above growth-optimizing, it widely oscillates
crossing R = 1 and even goes negative. Was the investor too crude in modifying their
view? Were they too harsh or too lenient? Either way, how could they become less risk
averse and even crossed the line into risk-loving territory of gambling? We encourage the
reader to try their own ways of modifying views. The more you try the more evidence you
obtain supporting the inevitable conclusion – it is very difficult, practically impossible, to
achieve consistent sensible risk aversion by ad-hoc modification of views.

10
 2 Implying and adjusting risk aversion

d ln f
A. Implied vols (%) B. Payoffs C. Implied risk aversion
d ln f˜

55 1.5 100

m (mkt) f f˜ 10
45
b (belief) sanity ↑
1
35 b̃ (adjusted)
1 0

25 −1
gambling
−10
15
x x
0.5 −100
0.4 0.7 1 1.3 1.6 0.4 0.7 1 1.3 1.6 0.4 0.7 1 1.3 1.6
Strike/Fwd S6M /Fwd S6M /Fwd

D. Correcting risk aversion E. Payoffs from risk aversion F. Detecting state dependence

104 1.5 1.3

R̃ f F̃
3
10
R
102 F̃ F̃ (f )
1 1
10 F

x x f
0.1 0.5 0.7
0.4 0.7 1 1.3 1.6 0.4 0.7 1 1.3 1.6 0.7 1 1.3

Note: We use real market data for STOXX50E as of 28/Jan/2011. By S6M we denote the value of the index in
6 months, and by Fwd – the corresponding observed forward value. Figures A, B and C examine the quality of
ad-hoc attempts to express risk aversion via conservative modification of views. The vertical scale on Figure C
is linear between -1 and 1 and logarithmic outside this interval. Figures D and E show how better control can
be achieved by working directly with risk aversion profiles. Figure F illustrates detection of state-dependence
in investor’s preferences. For detailed explanations of all graphs see the main text.

Using the payoff elasticity equation gives us a much easier way of including risk aversion.
To demonstrate let us create a simple textbook example. We take the same market and
the same growth-optimizing investor as we considered above. This time, however, we
accommodate the investor’s desire to suppress their view in the wings by stating a very
high level of risk aversion – see Figure 2.D. Integrating the payoff elasticity equation
immediately gives us the payoff structure – Figure 2.E.
The two Figures 2.D and 2.E also illustrate the corrective effect of adjusting risk aversion
profiles. By comparison to growth-optimizing, the investor with the risk aversion R is
more risk-averse in the wings (R > 1) and less risk-averse near ATM (R < 1). Looking at
the payoff (blue line F on Figure 2.E vs the red line f ) we see that the investor effectively
reallocates their exposure from the wings to the ATM region. This reallocation effect can
be perfectly removed (if desired) by replacing R with R̃, i.e. by making sure that we are
at least as risk averse as the growth-optimizing investor. Having obtained the payoffs we
can easily illustrate them in terms of market views. This would give us pictures which
resemble Figure 2.A in the general shape with the difference that this time we don’t just
hope (and almost surely fail) to imply some reasonable risk aversion – we know exactly

11
what we are building (this is how Investor 2 on Figure 2.A in Ref [3] was constructed).
Direct modifications of payoffs is yet another example of popular structuring ad-hocery for
imitating risk aversion. The most benign examples of that are introducing ad-hoc floors
and caps on top of otherwise problematic payoffs. To see that direct payoff modifications
can be dangerous, all we have to do is to examine modifications of f presented on Figures
2.B and 2.E. At first glance all of them look reasonable.
As the above examples illustrate, modifying payoff structures by changing risk aversion
profiles is easy and intuitive. With almost no practice you will see how to produce
structures that appear to have caps, floors and other recognizable features that can be
very useful in describing the resulting product. As before, we don’t just hope to produce
sensible risk aversion as a byproduct of ad-hoc modifications – we know exactly what risk
aversion profile we are using.

6.4 Rules of thumb

Handling risk aversion is a complicated and delicate enough task to require a specialist
tool. Payoff elasticity equation gives us such a tool. It would be even better, however,
if we could say something practical about products even when knowing next to nothing
about risk preferences. Amazingly enough, this is possible.
Note, for instance, that most investors would want to express positive risk aversion.2
We immediately see that positive risk aversion implies that F and f must agree on the
location and the type of their extremum points. Indeed, as we vary x both F (x) and f (x)
should be going up or down together for the elasticity d ln F/d ln f to remain positive.
Violation of this simple rule is what produced the wildly oscillating often negative profile
of Figure 2.C.
We can talk about state-agnostic and state-dependent investors. For the former kind the
utility of their investment depends only on its performance – the classical Arrow-Pratt
investor. The latter kind also cares about the underlying market at maturity. Clients
would normally know which kind they are and we can easily test if they get the right kind
of product. Indeed, for a state-agnostic investor (imagine a US investor who is exposed
to UK market only via their investment), R depends on x only via F . In this case, by the
payoff elasticity equation, F becomes a function of f . This we can test by simply plotting
F against f (while varying x as a parameter) and then examining the picture. Figure 2.F,
for instance, shows that F̃ cannot be computed from f alone – a typical signature of a
state-dependent investment.
We learned that the locations of the maxima and minima of F are essentially determined
by that of f . Given that F is just a function, the locations and the types of its extremal
points already give us a lot of information. A lot more can be said about F which belongs
to a state-agnostic investor. The fact that F depends on x only via f imposes severe
constraints. We can see, for example, that F and f intersect on a bond line, i.e. if
F (x1 ) = f (x1 ) and F (x2 ) = f (x2 ) then F (x1 ) = F (x2 ) = c, where c is a constant (not
necessarily 1).
2
This observation comes from understanding the difference between risk-averse investors and risk-
loving gamblers.

12
 3 Structuring as a manufacturing process

Capture Capture Check

Select mkt Capture Solve for Package
client’s client’s constraints,
variable mkt view payout for sale
view goals goals, etc.

U (x) b = fm
x m(x) b(x) d ln F 1
R(x) d ln f = R

Because structuring is often constrained by practical considerations (such as the number

of traded strikes which we could use for replicating the final product) such rules of thumb
provide us with a lot of information. Couple that with additional data like maximum
tolerable loss and you might be able to sketch the final optimal structure without detailed
knowledge of your client’s risk aversion.

7 Summary
At the heart of structuring we always have optimization which reflects the goals of the
client. This is preceded by preparatory steps defining the problem and followed by pack-
aging of the solution into a tradeable product. Together these steps form the backbone
of a manufacturing process which we summarized on Figure 3.
The preparatory steps include deciding on the market variable, capturing the relevant
market and client views and identifying the goals of the client. This leads us to the key
solution stage where our equations come in: the growth-optimal b = f m and the payoff
elasticity equation.
After the solution stage we check the quality of the derived product. At this point we
might be interested in assessing the expected performance of the product. The relevant
quantitative techniques are explained and illustrated in Refs. [11] and [12]. We might even
decide to go back and redefine the underlying variable. In [3] we considered a detailed
example when this happens. Having satisfied ourselves with the solution we proceed to
the final stage of packaging it into a tradeable product.

8 Appendix A: Discrete elasticity equation

In the derivation of the payoff elasticity equation we made two assumptions: first, we
considered continuous x and, second, we considered state-independent utility functions,
i.e. utilities which depend on x only via the value of the payoff F (x). We did it simply to
highlight the connection with the classic Arrow-Pratt measures of risk aversion which were
built on these assumptions. In the actual fact, none of these assumptions is important to
us except, of course, that the continuous case benefits from the rich toolbox of calculus.

13
A few lines of algebra give us a discrete fully general payoff elasticity equation. Indeed,
dividing both sides of Eq. (8) by mk and using Eqs. (1), (2) and (3) we obtain
 
u ln Fk , xk fk
′

Fk = P   . (27)
i u ′ ln Fi , xi bi

We then compute  
Fk+1 u′ ln Fk+1 , xk+1 fk+1
=   . (28)
Fk u ln Fk , xk fk
′

Taking logarithm on both sides and rearranging terms,

u′ (ln F ,xk+1 ) !
Fk+1 ln u′ (lnk+1
Fk ,xk ) fk+1
ln · 1− Fk+1
= ln . (29)
Fk ln fk
Fk

Finally we get a discrete payoff elasticity equation

ln FFk+1 1
k
fk+1
= , (30)
ln Rk,k+1
fk

where the expression for Rk,k+1 in terms of utility can be easily obtained from the previous
equation. We see that, in contrast to the original Arrow-Pratt notion, Rk,k+1 depends
not only on F but also on x. Detailed investigations of state-dependent utilities and
the emergence of Rk,k+1 as a generalization of the Arrow-Pratt risk aversion are very
interesting topics. These topics, however, would take us far beyond the scope of this paper.

9 Appendix B: Rationality and Logic

Together with rationality, as captured by the expected utility approach, there is another
technical aspect of decision making which is extremely important but which is often taken
for granted. This aspect is the mathematical foundation of logic itself.
The simplest and perhaps the most widely known example of logical framework is given
by the Boolean algebra which uses True=1 and False=0 to describe logical statements.
Boolean algebra has proven itself as an extremely powerful tool underpinning, among other
things, all computer-based calculations. Amazingly enough, despite all of its successes,
the Boolean approach leaves substantial room for improvement.
Indeed, what should we do with statements for which we cannot say if they are true
or false? Imagine, for example, a statement for which we only have a measure of how
often or how likely it is to be true. It turns out that the Boolean approach can be very
easily generalized to handle such cases [8]. All we have to do is to replace True and
False by numbers between zero and one and, of course, we have to generalize the rules
for manipulating these numbers. The resulting logical system turns out to coincide with
the standard probability theory. In particular, the product rule which we use to compute
joint probability, i.e.
p(A, B) = p(A)p(B|A) = p(B)p(A|B) , (31)

14
arises as a generalization of the Boolean product which we use to determine the truth
value of the joint statement (A and B). The only conceptual difference is that statements
can now have varying degree of information about one another which is handled by the
concept of conditioning (as in conditional probabilities). The last equality in (31), i.e.
Bayes’ theorem, emerges as a logical consistency requirement which comes from swapping
the order in which we include A and B into the calculation.
This intimate connection of Bayes’ theorem with logic underlies the emphasis we make
on logical investors. Indeed, violations of Bayes’ theorem do not just go against basic
probability theory. Such violations may also be a symptom of much deeper logical incon-
sistencies. It is our duty to help our investors to remain logical. Of course, the utility
theory itself is already based on logic. We can therefore expect that, at least partially,
logical behavior is already included in rationality. This, I think, is what allowed us to use
such simple derivations.
For a comprehensive literature review of the Bayesian methods in portfolio selection and
market risk management we point to [13] and references therein. For example applications
outside market risk (e.g. credit risk and stress testing) see Refs. [14] and [15].

References
[1] Soklakov A., “Why quantitative structuring?”, July (2015). arXiv:1507.07219.

[2] Soklakov A., “Bayesian lessons for payout structuring”, Risk, September (2011), 115-
119. arXiv:1106.2882.

[3] Soklakov A., “Deriving Derivatives”, Risk, July (2016), 78-83. arXiv:1304.7533.

[4] Bernoulli D., “Specimen Theoriae Nova de Mensura Sortis” (1738), reprinted in Econo-
metrica Bernoulli (1954).

[5] Christensen M. M., “On the history of the growth-optimal portfolio”, University of
Southern Denmark (2005).

[6] Kelly J. L. Jr, “A New Interpretation of Information Rate”, Bell System Technical
Journal, July (1956) 917-926.

[7] von Neumann J. and O. Morgenstern, “Theory of Games and Economic Behaviour”,
Princeton (1944, 2nd ed. 1947, 3rd ed. 1953).

[8] Jaynes E. T., “Probability Theory, The Logic of Science” (2003).

[9] Shimko D., “A tail of two distributions”, Risk, September (1994), 123-130.

[10] Markowitz H. M., “Portfolio Selection”, The Journal of Finance 7 (1), March (1952),
77-91.

[11] Soklakov A., “Model Risk Analysis via Investment Structuring”, July (2015).
arXiv:1507.07216.

15
[12] Soklakov A., “One trade at a time – unraveling the equity premium puzzle”, July
(2015). arXiv:1507.07214.

[13] Rachev S. T., Hsu J. S. J., Bagasheva B. S. and F. J. Fabozzi, “Bayesian methods
in Finance”. Wiley (2008).

[14] Jacobs M. Jr. and N. M. Kiefer, “The Bayesian approach to default risk: a guide” in
Rethinking Risk Measurement and Reporting, Boecker, K. (ed), Risk Books, London
(2010), 319-343.

[15] Jacobs M. Jr., Karagozoglu A. K. and F. J. Sensenbrenner, “Stress testing and model
validation: application of the Bayesian approach to a credit risk portfolio”, Journal of
Risk Model Validation 9 (3), 41-70 (2015).

16