The Relative Entropy of Expectation and Price

Paul McCloud
(Dated: February 13, 2025)
As operators acting on the undetermined final settlement of a derivative security, expectation
is linear but price is non-linear. When the market of underlying securities is incomplete, non-
linearity emerges from the bid-offer around the mid price that accounts for the residual risks of the
optimal funding and hedging strategy. At the extremes, non-linearity also arises from the embedded
options on capital that are exercised upon default. In this essay, these convexities are quantified
in an entropic risk metric that evaluates the strategic risks, which is realised as a cost with the
introduction of bilateral margin. Price is then adjusted for market incompleteness and the risk of
default caused by the exhaustion of capital.
In the complete market theory, price is derived from a martingale condition. In the incomplete
market theory presented here, price is instead derived from a log-martingale condition:
arXiv:2502.08613v1 [q-fin.MF] 12 Feb 2025

1
pt = − log Et exp[−αpT ]
α
for the price p of a funded and hedged derivative security, where the price measure E has minimum
entropy relative to economic expectations, and the parameter α is determined from the risk appetite
of the investor. This price principle is easily applied to standard models for market evolution, with
applications considered here including model risk analysis, deep hedging and decentralised finance.

I. INTRODUCTION Incomplete markets: Market equilibrium identifies a

unique price model with minimum entropy relative to
A derivative transaction exchanges the initial price p economic expectations, calibrated to available funding
for a commitment to return the contractual final price and hedging. In the process, this resolves the ambigu-
p+dp at a later time. In determining its economic viabil- ity of pricing in incomplete markets.
ity, the counterparties assess whether the trade enhances
Market and model risk: Uniquely identifying price
or degrades the overall performance of their investment
from expectation, market and model risks are distin-
portfolios, balancing expected return against net risk.
guished respectively as sensitivities to observed prices
Asymmetry between the expectations and investment
and subjective economic assumptions. This enables a
objectives of participants then drives market evolution,
hierarchical approach that recognises market tiering.
following the dynamic that trades execute when parties
agree on price but disagree on value. Decentralised finance: The level of margin is con-
Capital is deployed to finance the transaction, entail- trolled by a single parameter, and the link with funding
ing economic consequences for the investor beyond the settlements is transparent and programmatic. Imple-
contractual terms of the derivative. In addition to the ex- mented in the scripting of smart contracts for deriva-
pected final settlements, valuation depends on the fund- tives, this supports the transition to decentralisation.
ing of initial and variation margin and the level of default
protection they confer. The impact of margin is typically Advocates of deep hedging replace risk neutralisation
developed as a valuation adjustment to a price model de- with risk optimisation as the guiding principle for pric-
rived from the hedging of contractual settlements. In ing, as data-driven models see the complex correlations
this essay, an entropic margin model is instead placed among economic variables as opportunity rather than
at the centre of the pricing methodology, providing the hindrance. Convex risk optimisation is similarly used
risk counterbalance to expected return in the investment here to derive the price measure from the expectation
strategy. There are numerous benefits to this approach. measure, though the objectives differ. Model risk is
quantified as the residual risk after funding and hedging,
Investment strategy: Traditional strategies balance which requires for its definition a precise understanding of
expected return against risk. Here, the risk metric is the relationship between expectation and price. Taking
encoded in the entropic margin model as an additional this idea to its logical conclusion, the convex risk metric
settlement. Maximising the return on capital net of all is implemented as margin held against model risk. Op-
settlements naturally embeds risk management. timisation of the strategic return net of all settlements
then naturally embeds risk optimisation.
Incentivising margin: In this approach, the invest- Suppose that two counterparties deploy initial capital
ment strategy accounts for both the funding costs and c1 > 0 and c2 > 0 to the transaction and deposit the
risk management benefits of margin, and quantifies this proceeds in liquid margin accounts with unit price b1 > 0
in the derivative valuation. Margin thus becomes a key and b2 > 0 respectively. The first counterparty buys the
consideration for pricing. derivative from the second counterparty, and the capital
 2

Initial Final

Trading Book
𝑝 𝑝 + 𝑑𝑝
Trading Book

Figure 1. The contractual settlements of the trade.

Initial Final
𝑐1 𝑥1 ≔ (𝑐1 − 𝑝)(1 + 𝑑𝑏1 /𝑏1 ) + (𝑝 + 𝑑𝑝)

Deposit Book
𝑐1 − 𝑝 (𝑐1 − 𝑝)(1 + 𝑑𝑏1 /𝑏1 )
Trading Book
𝑝 𝑝 + 𝑑𝑝
Trading Book
𝑐2 + 𝑝 (𝑐2 + 𝑝)(1 + 𝑑𝑏2 /𝑏2 )
Deposit Book

𝑐2 𝑥2 ≔ (𝑐2 + 𝑝)(1 + 𝑑𝑏2 /𝑏2 ) − (𝑝 + 𝑑𝑝)

Figure 2. The contractual and funding settlements of the trade.

Initial Final
+ +
𝑐1 𝑐1 + 𝑑𝑐1 = 𝑥1 − −𝑥2

Deposit Book
𝑐1 − 𝑝 (𝑐1 − 𝑝)(1 + 𝑑𝑏1 /𝑏1 )
Trading Book
𝑝 𝑝 + 𝑑𝑝 + −𝑥1 + − −𝑥2 +

Trading Book
𝑐2 + 𝑝 (𝑐2 + 𝑝)(1 + 𝑑𝑏2 /𝑏2 )
Deposit Book

𝑐2 𝑐2 + 𝑑𝑐2 = 𝑥2 + − −𝑥1 +

Figure 3. The contractual, funding and default settlements of the trade.

 3

deployed must be sufficient to enter the transaction: model based on the entropic risk metric – a convex risk
metric that is chosen for its convenient decomposition
−c2 ≤ p ≤ c1 (1) in tiered derivative markets. It is an idealisation of real
trading activity, and in particular does not account for
With the proceeds and unused capital deposited in the the default protections provided by additional capital
respective margin accounts, the final margin available to covering the net risks of the institution. Extensions to
settle the derivative is (c1 − p)(1 + db1 /b1 ) for the first multilateral margin, which can be incorporated into the
counterparty and (c2 +p)(1+db2 /b2 ) for the second coun- framework, are not considered here.
terparty. This is sufficient to settle the contractual com-
mitments of the derivative when:
II. ENTROPIC INVESTMENT STRATEGIES
−(c1 − p)(1 + db1 /b1 ) ≤ p + dp ≤ (c2 + p)(1 + db2 /b2 ) (2)

Define the default trigger levels: The objective of the investor is to optimise the return
dc on the investment strategy, a task that depends on
x1 := (c1 − p)(1 + db1 /b1 ) + (p + dp) (3) an initial assessment of the final economy that is both
subjective and indeterminate. This assessment is encap-
 
c1 p
= (b1 + db1 ) +d sulated in the expectation measure E of the investor. De-
b1 b1
fine the corresponding probability, variance and covari-
x2 := (c2 + p)(1 + db2 /b2 ) − (p + dp) ance operators:
 
c2 p
= (b2 + db2 ) −d
b2 b2 P[A] := E[1A ] (7)
t t
V[X] := E[XX ] − E[X]E[X]
The first counterparty defaults when x1 < 0 and the
second counterparty defaults when x2 < 0, and these V[X, Y ] := E[XY t ] − E[X]E[Y ]t
two conditions are mutually exclusive. The actual final
settlement of the derivative, net of default, is then: for the event A and the random vectors X and Y . In-
vestment performance is estimated using these operators,
p + dp + (−x1 )+ − (−x2 )+ (4) and the strategy is optimised accordingly.

and the final capital of the counterparties is:

A. Entropic risk metric
c1 + dc1 = (x1 )+ − (−x2 )+ (5)
c2 + dc2 = (x2 )+ − (−x1 )+ Maximising the mean E[dc] is an obvious strategic ob-
jective, though taken on its own this incentivises reckless
respectively. These expressions for the change in capital
investment and fails to reward effective risk management.
are rearranged to:
Mean-variance optimisation extends the strategy by max-
c1 imising the variance-adjusted mean:
d = (6)
b1
+ + 1
E[dc] − αV[dc] (8)
 
c1 p b2 + db2 p c2 p
− −d − d − +d 2
b1 b1 b1 + db1 b2 b2 b1
c2 In this expression, the parameter α controls the risk ad-
d = justment, and is interpreted as the Lagrange multiplier
b2
 +  + for the variance constraint in a strategy that maximises
p c2 b1 + db1 c1 p p
d − − − −d −d the mean for a fixed variance. Incorporating higher
b2 b2 b2 + db2 b1 b1 b2 moments, entropic-risk optimisation further extends the
strategy by maximising the entropy-adjusted mean:
demonstrating that the discounted return on capital
matches the discounted return on the derivative adjusted 1
by a put option on d(p/b1 ) with strike −c1 /b1 , exercised E[α][dc] := − log E exp[−α dc] (9)
α
when the first counterparty defaults, and a call option ∞
X (−α)n−1
on d(p/b2 ) with strike c2 /b2 , exercised when the second 1
= E[dc] − αV[dc] + Mn [dc]
counterparty defaults. When assessing the viability of 2 n=3
n!
the trade, it is these net returns that must be considered,
accounting for funding costs and the possibility that the where Mn is the nth cumulant operator of the investor,
settling counterparty defaults on its commitments. so that M0 = 0, M1 = E and M2 = V. The change
This simple framework for bilateral margin is used in of measure implicit in this definition reweights according
the following to derive equilibrium pricing expressions to the sign and magnitude of the return, with properties
that account for funding and default, using a margin that make it a useful metric of value-at-risk.
 4

Normal Poisson Uniform

Figure 4. Entropy-adjusted mean as a function of risk adjustment for three random variables. In each case, the variable is
standardised by linear transformation to have zero mean and unit standard deviation. The entropy-adjusted mean is monotonic
decreasing; it is unbounded for the normal variable, bounded below for the Poisson variable with ten expected jumps, and
bounded above and below for the uniform variable.

Theorem II.1 (Entropic risk metric). Let dc be a ran- Proof. These are standard results from information the-
dom variable whose cumulant generating function in the ory, and only the key elements of their proofs are pre-
probability measure E is defined. The entropy-adjusted sented here. Monotonicity is a consequence of Hölder’s
mean: inequality:
1
E[α][dc] := − log E exp[−α dc] (10) E[exp[−α1 dc]]1/α1 ≤ E[exp[−α2 dc]]1/α2 (16)
α
is monotonic decreasing as a function of the parameter for 0 < α1 < α2 , so that E[α1 ][dc] ≥ E[α2 ][dc]. The first
α > 0 and has limits: limit follows from the expansion of the cumulant generat-
ing function. For the second limit, Markov’s inequality:
E[0][dc] = E[dc] (11)
E[exp[−α dc]]
E[∞][dc] = ess inf[dc] P[dc ≤ γ] ≤ (17)
exp[−αγ]
Let Ē be an equivalent probability measure with strictly
positive Radon-Nikodym weight dĒ/dE. Define the en- for the parameters α > 0 and γ > ess inf[dc] (so that
tropy of Ē relative to E by: P[dc ≤ γ] > 0) is rearranged to:

dĒ
 
dĒ 1
S[Ē|E] := E log (12) E[α][dc] ≤ γ − log[P[dc ≤ γ]] (18)
dE dE α

Relative entropy is a metric of distance between equiva- From this, taking the limit α → ∞ followed by taking
lent probability measures: it is positive, S[Ē|E] ≥ 0, and the limit γ → ess inf[dc] derives E[∞][dc] ≤ ess inf[dc].
is zero when Ē = E. Let Eα be the equivalent probability The other side of the target inequality follows from the
measure with Radon-Nikodym weight dEα /dE given by: almost certainty of exp[−α dc] ≤ exp[−α ess inf[dc]], so
that E[α][dc] ≥ ess inf[dc].
dEα exp[−α dc] Gibb’s theorem states that S[Ē|E] ≥ 0, which follows
:= (13)
dE E[exp[−α dc]] from the observation that x log[x] ≥ x − 1 for x > 0.
The remaining result is a presentation of Donsker and
The entropy-adjusted mean E[α][dc] equals the mean Varadhan’s variational formula. Let m̄ := dĒ/dE and
Ē[dc] adjusted by relative entropy: mα := dEα /dE be the Radon-Nikodym weights. Then:
1
E[α][dc] = Ē[dc] + (S[Ē|E] − S[Ē|Eα ]) (14) S[Ē|E] − S[Ē|Eα ] (19)
α α
= E[m̄ log[m̄]] − E [(m̄/mα ) log[m̄/mα ]]
Instances of this equality for Ē = E and Ē = Eα generate = Ē[log[m̄]] − (Ē[log[m̄]] − Ē[log[mα ]])
bounds for the entropy-adjusted mean:
= Ē[−α dc − log E exp[−α dc]]
Eα [dc] ≤ (15)
= −α(Ē[dc] − E[α][dc])
1 1
S[Eα |E] + Eα [dc] = E[α][dc] = E[dc] − S[E|Eα ]
α α relating the entropy-adjusted mean to the relative en-
≤ E[dc] tropies of the measures.
 5

The lever α > 0 thus dials the entropy-adjusted mean Risk adjustment is controlled by the parameter α in the
E[α][dc] across the span of possible values for dc between range 0 ≤ α ≤ ∞, and entropic risk is monotonic in this
ess inf[dc] and E[dc] and no further. Entropic risk is, in range:
this sense, more targeted than variance adjustment as a
moments-based metric of value-at-risk. The equivalent ess inf[ω · dq] = (22)
measure Eα plays a central role in this result, and the E[∞][ω · dq] ≤ E[α][ω · dq] ≤ E[0][ω · dq]
gap between the mean E[dc] and the entropy-adjusted = ω · E[dq]
mean E[α][dc] is proportional to the distance between
the measure E and the measure Eα , as defined by their Large values of α thus imply that the investor takes
relative entropy. more account of negative outcomes in the strategic ob-
jective, corresponding to a lower risk appetite. The op-
timal strategy is the stationary point of entropic risk
B. Optimising entropic risk ω = (1/α)ϕ where the unit optimal portfolio ϕ satisfies
the calibration condition:
The investor has a choice of market strategies whose
returns over the investment period are the components 0 = Eϕ [dq] (23)
of the return vector dq. The return on the investment
portfolio whose weights are the components of the weight As a practical observation, the Newton-Raphson scheme:
vector ω is then dc := ω · dq, the weighted sum of the
returns on the market strategies. ϕ(0) = 0 (24)
Mean-variance and entropic-risk optimisation are the (n) (n)
ϕ(n+1) = ϕ(n) + Vϕ [dq]−1 Eϕ [dq]
basis for investment strategies, leading in each case to
equilibrium models of pricing and hedging. There are calibrates the unit optimal portfolio to the calibration
many similarities between these strategies, but also some condition when the variance is invertible. This portfolio
critical differences. Mean-variance optimisation does not is a property of the market return, generating the price
account for non-Gaussian features in the return, and the measure Eϕ equivalent to the expectation measure E.
strategy may need to be moderated to avoid arbitrage. In The optimal strategy is constructed on the assumption
contrast, entropic-risk optimisation incorporates all the that the investor has expertise in all the available mar-
cumulants of the return, and naturally leads to arbitrage- kets. In practice, markets are tiered and investors spe-
free equilibrium pricing. cialise in specific market sectors, and the optimal strat-
Introduce the families of equivalent measures Eϕ and egy across all sectors emerges as the net strategy of multi-
ϕ,ζ
E with Radon-Nikodym weights: ple specialist investors. This arrangement is conveniently
dEϕ exp[−ϕ · dq] captured in the entropic risk metric.
:= (20) Suppose that the market return vector decomposes as
dE E[exp[−ϕ · dq]
a direct sum dq = dq1 ⊕ · · · ⊕ dqn where dqi is the return
dEϕ,ζ exp[−ϕ · dq − 12 dq · ζ dq] vector for the ith market tier. The optimal portfolio is
:=
dE E[exp[−ϕ · dq − 12 dq · ζ dq]] similarly decomposed as a direct sum ϕ = ϕ1 ⊕ · · · ⊕ ϕn
where ϕi is the portfolio in the ith market tier, and is
parametrised by the vector ϕ and the positive-definite solved from the calibration conditions:
matrix ζ. Price measures that are equivalent to the ex-
pectation measure are identified within this collection 0 = Eϕ1 ⊕···⊕ϕn [dqi ] (25)
from the equilibrium of the entropic-risk optimisation
strategy. The parameter ϕ arises from the change of mea- for i = 1, . . . , n. This provides n conditions for the n
sure in the definition of the entropic risk metric, which de- components of the optimal portfolio. While these cali-
pends on the existence of the cumulant generating func- bration conditions need to be solved simultaneously, mar-
tion. If this does not exist, the quadratic-regularised ket tiering is exploited to incrementally build the optimal
measure with parameter ζ is used instead. Regularisa- portfolio from the independent contributions of sector in-
tion is removed in the limit ζ → 0; if the strategy does vestors, localising at each tier the risk management ac-
not survive this limit, then regularisation, which mod- tivity for the corresponding market expert. In this per-
erates the impact of large events, is retained as a pa- spective, there are n(n + 1)/2 calibration conditions for
rameter of the model. In the following, the cumulant n sub-markets:
generating function is assumed to exist, implying the use
1
of quadratic regularisation wherever necessary. 0 = Eϕ1 [dq1 ] (26)
In the entropic-risk optimisation strategy, the investor ϕ21 ⊕ϕ22 ϕ21 ⊕ϕ22
seeks the optimal portfolio ω of market strategies that 0=E [dq1 ] = E [dq2 ]
maximises the entropy-adjusted mean: ..
.
1 n n n n
E[α][ω · dq] := − log E exp[−αω · dq] (21) 0 = Eϕ1 ⊕···⊕ϕn [dq1 ] = · · · = Eϕ1 ⊕···⊕ϕn [dqn ]
α
 6

where the calibration conditions in the ith row maximise Since expanding the market can only increase the
the entropy-adjusted mean of the ith optimal portfolio opportunities for the investor, the maximum entropy-
ϕi = ϕi1 ⊕ · · · ⊕ ϕii constituted in the lowest i market tiers adjusted mean increases, µi ≥ µi−1 , as new tiers are
only. The ith investor then takes responsibility for de- introduced. Furthermore, the ith tier introduces no addi-
termining the primary component ψi of the ith optimal tional advantage for the strategy if the optimal portfolios
portfolio and its delta δ i = δ1i ⊕ · · · ⊕ δi−1
i
to the (i − 1)th ϕi and ϕi−1 satisfy the indifference condition µi = µi−1 .
optimal portfolio in the underlying sub-market. Stack- This is adopted as the equilibrium condition for pricing
ing these deltas for i = 1, . . . , n generates the optimal derivative strategies that have no intrinsic value beyond
portfolio of the full market. their contractual link with underlying strategies.
Theorem II.2 (Optimising entropic risk). Consider a
market of investment strategies whose return vector de- III. ENTROPIC PRICING AND HEDGING
composes as a direct sum dq = dq1 ⊕ · · · ⊕ dqn where dqi
is the return vector for the ith tier of the market. For The market of underlying securities accessible to the
each i = 1, . . . , n, define the ith sub-market with return investor performs the two key functions of funding and
vector dq i = dq1 ⊕ · · · ⊕ dqi and let ϕi = ϕi1 ⊕ · · · ⊕ ϕii be hedging for the derivative security. Fix a finite investment
a portfolio in this sub-market. Use the invertible linear horizon, and assume the market is tiered into a set of un-
change of variables: derlying securities with initial price vector q and final
ψi = ϕii price vector q + dq and a derivative security with initial
ϕij = ψj − (δjj+1 + · · · + δji ) (27) price p and final price p + dp. The investor quantifies a
δji = ϕi−1
j − ϕij priori the performance of the investment strategy using
an initial assessment of the final economic state, encap-
to represent the delta from the portfolio ϕi−1 to the port- sulated in the expectation measure E that maps variables
folio ϕi as the primary portfolio ψi in the leading market at final time t + dt to their expectations at initial time
tier minus the hedge portfolio δ1i ⊕ · · · ⊕ δi−1
i
in the un- t. Following the strategy developed in the previous sec-
derlying market tiers. tion, the optimal portfolio then maximises the entropy-
Let E be a probability measure. For each i = 0, . . . , n, adjusted mean return at a specified risk adjustment α
define the equivalent probability measure Ei with Radon- related to the risk appetite of the investor.
Nikodym weight: As summarised in Figure 5, there are two self-funded
investment strategies to consider: one that includes only
dEi exp[−(ϕi1 · dq1 + · · · + ϕii · dqi )] underlying securities; and one that also includes the
:= (28)
dE E[exp[−(ϕi1 · dq1 + · · · + ϕii · dqi )]] derivative security. In the first strategy, the portfolio ω
of underlying securities is optimal when it maximises the
and define the ith entropy-adjusted mean: entropy-adjusted mean return E[α][ω · dq]. The optimal
µi := − log E exp[−(ϕi1 · dq1 + · · · + ϕii · dqi )] (29) portfolio thus satisfies the calibration condition:
0 = E[(dq − qr dt) exp[−αω · dq]] (32)
At the ith tier, the portfolio ϕi maximises the entropy-
adjusted mean µi when the primary portfolio ψi and the where r dt is the Lagrange multiplier used to implement
hedge portfolio δ1i ⊕ · · · ⊕ δi−1
i
are calibrated to the hedge the self-funding condition ω · q = 0. Adjusting for the
conditions: opportunity and risk that adding the derivative security
presents, in the second strategy the optimal portfolio ω
0= (30) is modified by the hedge portfolio δ, and optimality is
E i−1
[dqj exp[−(ψi · dqi − (δ1i · dq1 + · · · + i
δi−1 · dqi−1 ))]] restored in the combined portfolio when it maximises the
entropy-adjusted mean return E[α][dp + (ω − δ) · dq]. The
for j = 1, . . . , i. Starting with the probability measure hedge portfolio thus satisfies the hedge condition:
E0 = E, these calibrations are iterated to obtain the unit 0 = E[(dq − q(r + αs) dt) exp[−α(dp + (ω − δ) · dq)]] (33)
optimal portfolio ϕn in the full market, with its associ-
ated equivalent probability measure En calibrated to the where (r + αs) dt is the Lagrange multiplier used to im-
calibration conditions for the market strategies. plement the self-funding condition δ · q = p.
Returns from correlated changes in the derivative and
Proof. This result follows from the measure-change rela- underlying prices are cancelled by hedging, with the
tionship: residual return on the combined portfolio maximising the
entropy-adjusted mean. When its price additionally sat-
log Ei−1 exp[−(ψi · dqi − (δ1i · dq1 + · · · + δi−1
i
· dqi−1 ))]
isfies the price condition:
= log E exp[−(ϕi1 · dq1 + · · · + ϕii · dqi )] (31)
E[α][ω · dq] = E[α][dp + (ω − δ) · dq] (34)
− log E exp[−(ϕi−1
1 · dq1 + · · · + ϕi−1
i−1 · dqi−1 )]
the derivative security neither advantages nor disadvan-
between the entropy-adjusted means. tages the investment strategy.
 7

Entropic Pricing and Hedging

Fix a finite investment horizon and let E be the expectation measure mapping variables at the final time t + dt to
their expectations at the initial time t. Assume the market is tiered into a set of underlying securities with initial price
vector q and final price vector q + dq and a derivative security with initial price p and final price p + dp. Consider two
self-funded strategies: one comprising underlying securities only; and one that also includes the derivative security.
Underlying only: In this strategy, the investment Underlying and derivative: In this strategy, the
comprises an underlying portfolio ω satisfying the initial investment comprises the derivative security and an un-
condition: derlying portfolio (ω − δ) satisfying the initial condition:

0=ω·q (35) 0 = p + (ω − δ) · q (38)

so that the investment is self-funded. The return on the so that the investment is self-funded. The return on the
investment has entropy-adjusted mean: investment has entropy-adjusted mean:

E[α][ω · dq] = (36) E[α][dp + (ω − δ) · dq] = (39)

1 1
− log E exp[−αω · dq] − log E exp[−α(dp + (ω − δ) · dq)]
α α
at the risk adjustment α > 0. The optimal portfolio ω at the risk adjustment α > 0. The hedge portfolio δ that
that maximises this satisfies the calibration condition: maximises this satisfies the hedge condition:

0 = E[(dq − qr dt) exp[−αω · dq]] (37) 0 = E[(dq − q(r + αs) dt) exp[−α(dp + (ω − δ) · dq)]] (40)

where r dt is the Lagrange multiplier used to implement where (r + αs) dt is the Lagrange multiplier used to im-
the self-funding condition. plement the self-funding condition.
These two strategies offer different entropy-adjusted mean returns to the investor. When the derivative security
has no intrinsic value beyond its contractual link with the underlying securities, the efficient market equilibrates to
the indifferent state wherein these entropy-adjusted means align, leading to the price condition:

E[α][ω · dq] = E[α][dp + (ω − δ) · dq] (41)

The calibration condition is solved to derive the optimal portfolio ω of underlying securities. The hedge and price
conditions are then solved simultaneously to derive the initial price p from the final price p + dp of the derivative
security and the hedge portfolio δ of underlying securities that re-establishes optimality for the combined portfolio.

Figure 5. In entropic pricing, economic principles are established that determine the optimal portfolio of underlying securities
and the hedge portfolio and price of the derivative security, based on the maximisation of the entropy-adjusted mean return.

Calibration generates the implied funding rate r and final derivative price. Using the price measure, the hedge
the price measure Eϕ equivalent to the expectation mea- and price conditions are re-expressed as:
sure E with Radon-Nikodym weight:
0 = Eϕ [(dq̄ − αqs dt) exp[−α(P − δ · dq̄)]] (44)
ϕ
dE exp[−ϕ · dq] 1
:= (42) p= Eϕ [α][P − δ · dq̄]
dE E[exp[−ϕ · dq]] 1 + r dt
where the unit optimal portfolio ϕ satisfies the self- These conditions express a non-linear relationship be-
funding condition ϕ · q = 0 and the calibration condition: tween the initial and final derivative prices. By definition,
the derivative security with notional λ has final price λP .
qr dt = Eϕ [dq] (43) The corresponding initial price pλ and hedge portfolio δλ
All investments are self-funded within the underlying per unit notional and the implied funding spread sλ are
market, and this model does not assume the existence determined from the hedge and price conditions:
of a risk-free funding rate for investments that are not 0 = Eϕ [(dq̄ − αqsλ dt) exp[−λα(P − δλ · dq̄)]] (45)
self-funded. The implied funding rate defined here orig-
1
inates in the optimisation as the Lagrange multiplier for pλ = Eϕ [λα][P − δλ · dq̄]
the self-funding condition. 1 + r dt
Let dq̄ := dq − qr dt be the underlying return cen- and the self-funding condition. Scaling the derivative no-
tralised in the price measure and let P := p + dp be the tional is thus equivalent to scaling the risk appetite.
 8

Delta Price
1 2

λ = -5 λ = -5
λ = -4 λ = -4
λ = -3 λ = -3
λ = -2 λ = -2
λ = -1 1 λ = -1
λ=0 λ=0
λ=1 λ=1
λ=2 λ=2
λ=3 λ=3
λ=4 λ=4
0 0
λ=5 λ=5
-2 -1 0 1 2 -2 -1 0 1 2
Strike Strike

Delta Price
1 2

λ = -5 λ = -5
λ = -4 λ = -4
λ = -3 λ = -3
λ = -2 λ = -2
λ = -1 1 λ = -1
λ=0 λ=0
λ=1 λ=1
λ=2 λ=2
λ=3 λ=3
λ=4 λ=4
0 0
λ=5 λ=5
-2 -1 0 1 2 -2 -1 0 1 2
Strike Strike

Delta Price
1 2

λ = -5 λ = -5
λ = -4 λ = -4
λ = -3 λ = -3
λ = -2 λ = -2
λ = -1 1 λ = -1
λ=0 λ=0
λ=1 λ=1
λ=2 λ=2
λ=3 λ=3
λ=4 λ=4
0 0
λ=5 λ=5
-2 -1 0 1 2 -2 -1 0 1 2
Strike Strike

Figure 6. Three examples of entropic hedging and pricing for a call option on an underlying security with return dq modelled
as a normal variable, a Poisson variable with ten expected jumps, and a uniform variable in the price measure. In each case,
the return is standardised by linear transformation to have zero mean and unit standard deviation, and the derivative payoff
for strike k is P := (dq − k)+ . Risk adjustment is fixed at one, α = 1, and the implied funding rate is set to zero, r = 0. The
graphs plot the delta and price per unit notional as a function of the strike for a range of notionals λ, highlighting the central
plot for the mid-pricing case λ = 0. Residual risk after hedging creates a bid-offer spread around the mid price and skews the
optimal hedge ratio for positions with large notional.

When α is small, the hedge and price conditions are This expansion in terms of α approximates the price
expanded to discover the solutions: when the risk appetite of the investor is large or when the
derivative security has small notional. In the marginal
Eϕ [P ]/(1 + r dt) − Vϕ [dq]−1 Vϕ [dq, P ] · q market, the model thus implies a limit order book for the
s= + O[α]
Vϕ [dq]−1 q dt · q derivative security with mid price and bid-offer spread:
δ = Vϕ [dq]−1 (Vϕ [dq, P ] + qs dt) + O[α] (46)
1

1
 pmid = Eϕ [P ]/(1 + r dt) (47)
p= Eϕ [P ] − αVϕ [P − δ · dq] + O[α2 ] ϕ
1 + r dt 2 pbid-offer = αV [P − δ · dq]/(1 + r dt)
 9

Delta Price
0.5 1

λ = -5 λ = -5
λ = -4 λ = -4
λ = -3 λ = -3
λ = -2 λ = -2
λ = -1 λ = -1
λ=0 λ=0
λ=1 λ=1
λ=2 λ=2
λ=3 λ=3
λ=4 λ=4
0 0
λ=5 λ=5
-2 -1 0 1 2 -2 -1 0 1 2
Strike Strike

Delta Price
0.5 1

λ = -5 λ = -5
λ = -4 λ = -4
λ = -3 λ = -3
λ = -2 λ = -2
λ = -1 λ = -1
λ=0 λ=0
λ=1 λ=1
λ=2 λ=2
λ=3 λ=3
λ=4 λ=4
0 0
λ=5 λ=5
-2 -1 0 1 2 -2 -1 0 1 2
Strike Strike

Delta Price
0.5 1

λ = -5 λ = -5
λ = -4 λ = -4
λ = -3 λ = -3
λ = -2 λ = -2
λ = -1 λ = -1
λ=0 λ=0
λ=1 λ=1
λ=2 λ=2
λ=3 λ=3
λ=4 λ=4
0 0
λ=5 λ=5
-2 -1 0 1 2 -2 -1 0 1 2
Strike Strike

Figure 7. Three examples of entropic hedging and pricing for a digital option on an underlying security with return dq modelled
as a normal variable, a Poisson variable with ten expected jumps, and a uniform variable in the price measure. In each case,
the return is standardised by linear transformation to have zero mean and unit standard deviation, and the derivative payoff
for strike k is P := 1dq>k . Risk adjustment is fixed at one, α = 1, and the implied funding rate is set to zero, r = 0. The
graphs plot the delta and price per unit notional as a function of the strike for a range of notionals λ, highlighting the central
plot for the mid-pricing case λ = 0. Residual risk after hedging creates a bid-offer spread around the mid price and skews the
optimal hedge ratio for positions with large notional.

The discounted mid price of the derivative security is a tinuous and discrete hedging, and accounting for the risk
martingale in the price measure and the bid-offer spread appetite of the investor. The case studies in Figures 6 and
prices incompleteness of the underlying market. Higher- 7 demonstrate entropic hedging and pricing for call and
order adjustments then accommodate the extended fea- digital options over a single investment horizon, and show
tures of the price distributions. the bid-offer spread around the mid model that prices the
This model is extensive in scope, consistently encom- residual risk after hedging. To further develop intuition
passing arbitrary economic assumptions, predictable and on the meaning of the hedge and price conditions and
unpredictable funding, complete and incomplete hedge their solutions, it is useful to consider boundary cases
markets, continuous and discontinuous evolution, con- where they simplify.
 10

A. Empty market C. Continuous diffusion

In this section, assume that there are no underlying In this section, assume that the underlying and deriva-
securities available to the investor for funding or hedging. tive prices diffuse continuously, and the investor is able
Assumption III.1 (Empty market). The market of un- to re-balance the hedge over short investment horizons.
derlying securities is empty. Assumption III.3 (Continuous diffusion). The diffu-
It is interesting to note that entropic pricing handles sion of the underlying and derivative prices is continuous
this trivial case in a sensible way. The calibration and almost certainly in the expectation measure E, so that
hedge conditions are empty, and the price condition im- their cumulant generating function satisfies:
plies that the derivative price is a log-martingale in the
expectation measure: log E exp[β dp + ψ · dq] = (53)
βE[dp] + ψ · E[dq]
p = E[α][P ] (48)
1 1
In the absence of any market information, price reduces + β 2 V[dp] + βψ · V[dq, dp] + ψ · V[dq]ψ + O[dt2 ]
2 2
to expectation, adjusted to match the risk appetite of the
investor. with V[dp] = O[dt], V[dq, dp] = O[dt] and V[dq] = O[dt].
Two assumptions are included here. The first assump-
B. Predictable funding
tion, that the market prices diffuse continuously, justifies
the short-horizon expression for their cumulant generat-
ing function. The second assumption is that the investor
In this section, assume that the underlying market re-balances frequently and at negligible cost, so that the
available to the investor includes a funding security whose O[dt] corrections to the investment strategy can be ig-
final price is predictable. nored, and the bid-offer spread around the mid price de-
Assumption III.2 (Predictable funding). The market rives exclusively from the post-hedging residual risks.
partitions into a vector of underlying securities with ini- Moving to the price measure Eϕ , the impact of the
tial price q and final price q + dq and a funding security measure change on the cumulant generating function de-
with strictly positive initial price b > 0 and predictable rives the drift adjustments in Girsanov’s theorem:
final price b + db that has zero variance V[db] = 0 in the
expectation measure E. log Eϕ exp[β dp + ψ · dq] = (54)
β(E[dp] − ϕ · V[dq, dp]) + ψ · (E[dq] − V[dq]ϕ)
With this assumption, the calibration and hedge condi-
tions are solved for the implied funding rate r and implied 1 1
+ β 2 V[dp] + βψ · V[dq, dp] + ψ · V[dq]ψ + O[dt2 ]
funding spread s: 2 2
E[db] Differentiating with respect to ψ then generates:
r= (49)
b dt
s=0 Eϕ [dq exp[β dp + ψ · dq]]
= (55)
Eϕ [exp[β dp + ψ · dq]]
When no additional underlying securities are available,
the price condition implies that the derivative price is a E[dq] + βV[dq, dp] + V[dq](ψ − ϕ) + O[dt2 ]
discounted log-martingale in the expectation measure:
These results are sufficient to solve the conditions. The
1 calibration condition is solved for the implied funding
p= E[α][P ] (50)
1 + r dt rate r and the unit optimal portfolio ϕ:
More generally, underlying securities are available for
V[dq]−1 E[dq] · q
hedging, and the calibration and hedge conditions for the r= + O[dt] (56)
unit optimal portfolio ϕ and hedge portfolio δ become: V[dq]−1 q dt · q
ϕ = V[dq]−1 (E[dq] − qr dt) + O[dt]
0 = Eϕ [dq̄] (51)
0 = Eϕ [dq̄ exp[−α(P − δ · dq̄)]] and the hedge and price conditions are solved for the
implied funding spread s, the hedge portfolio δ and the
The price condition then derives the discounted log- initial price p:
martingale expression in the price measure:
1 Eϕ [P ] − V[dq]−1 V[dq, P ] · q
p= Eϕ [α][P − δ · dq̄] (52) s= + O[dt] (57)
1 + r dt V[dq]−1 q dt · q
The funding security funds the investment and manifests δ = V[dq]−1 (V[dq, P ] + qs dt) + O[dt]
as discounting in the price expression, and the additional 1
underlying securities hedge the derivative risks. p = (1 − r dt)Eϕ [P ] − αV[P − δ · dq] + O[dt2 ]
2
 11

D. Markovian diffusion which is solved simultaneously with the self-funding con-

dition ϕ · q = 0 for the implied funding rate r and the
In this section, assume that the underlying and deriva- unit optimal portfolio ϕ. The hedge and price conditions
tive prices are deterministic functions of a state vector. become:
 
Assumption III.4 (Markovian diffusion). There is a ∂p ∂q ∂q ∂q
0= ·ν + qs − ·ν δ (63)
vector x of state variables whose cumulant generating ∂x ∂x ∂x ∂x
function in the expectation measure E satisfies: Z
1
+ (1 − exp[−α(∂j p − δ · ∂j q)]) ∂j q θϕ [dj] + O[dt]
1 j α
log E exp[η · dx] = (58)
dt ∂p ∂p ϕ 1 ∂ 2 p
1
Z 0= − pr + ·µ + 2
··ν
η · µ + η · νη + (exp[η · j] − 1) θ[dj] + O[dt] ∂t  ∂x 2 ∂x
2

j 1 ∂p ∂q ∂p ∂q
− α −δ· ·ν −δ·
where the vector µ ≡ µ[t, x] and positive-definite ma- 2 ∂x ∂x ∂x ∂x
Z
trix ν ≡ ν[t, x] are the mean and covariance rate for the 1
+ (1 − exp[−α(∂j p − δ · ∂j q)]) θϕ [dj] + O[dt]
continuous component of the diffusion, and the positive j α
measure θ[dj] ≡ θ[t, x][dj] is the frequency of jumps in
the discontinuous component of the diffusion. The un- which are solved simultaneously with the self-funding
derlying price q ≡ q[t, x] and derivative price p ≡ p[t, x] condition δ ·q = p for the implied funding spread s, hedge
are assumed to be functions of the time t and state x. portfolio δ and derivative price p.
Starting with its terminal contractual settlement, the
This form for the cumulant generating function of the numerical solution of these conditions for the derivative
state implies its distribution is divisible, which enables price repeats a Newton-Raphson scheme that determines
the transition to the limit dt → 0 of continuous hedging. the initial price p from the final price p + dp over each
Repeated differentiation with respect to η generates Itô’s interval in a discretised schedule. The high dimensional
lemma for the function f ≡ f [t, x] of the time t and state complexity of this procedure is a consequence of the non-
x in the expectation measure E: linearity of the hedge portfolio in the hedge condition.
1 ∂f ∂f 1 ∂2f There are two useful scenarios where this condition is
E[df ] = + ·µ+ ··ν (59) linear: the mid-pricing case α = 0; and the continuous
dt ∂t Z ∂x 2 ∂x2
diffusion case θ[dj] = 0. Pricing simplifies in these cases.
+ ∂j f θ[dj] + O[dt]
j Mid-pricing: When α = 0, the hedge and price condi-
tions simplify:
where ∂j f [t, x] := f [t, x+j]−f [t, x]. In the price measure
Eϕ , Itô’s lemma is modified by the drift adjustments of 
∂p ∂q
Z 
Girsanov’s theorem: p − V −1 ·ν + ∂j p ∂j q θϕ [dj] · q
∂x ∂x j
1 ϕ ∂f ∂f 1 ∂2f s= −1 + O[dt]
E [df ] = + · µϕ + ··ν (60) V q·q
dt ∂t Z ∂x 2 ∂x2 
∂p ∂q
Z 
δ = V −1 ·ν + ∂j p ∂j q θϕ [dj] + qs + O[dt]
+ ∂j f θϕ [dj] + O[dt] ∂x ∂x j
j
∂p ∂p ϕ 1 ∂ 2 p
where the mean rate and frequency of jumps are adjusted: 0= − pr + ·µ + ··ν (64)
∂t Z ∂x 2 ∂x2
 
ϕ ∂q + (∂j p − δ · ∂j q) θϕ [dj] + O[dt]
µ := µ − ν ϕ · (61)
∂x j

θϕ [dj] := exp[−ϕ · ∂j q] θ[dj] where:

With this assumption, the price model is established
Z
∂q ∂q
on the assumed divisible process for the state vector x V := ·ν + ∂j q ∂j q θϕ [dj] (65)
∂x ∂x j
and the underlying price vector q[t, x] given as a function
of time and state. Using the expression for Itô’s lemma in Embedding the expressions for the implied funding
the price measure, the calibration, hedge and price con- spread and hedge portfolio, the price condition becomes
ditions are transformed into partial differential-integral a linear parabolic differential-integral equation for the
form. The calibration condition becomes: derivative price. This equation balances the contribu-
∂q ∂q 1 ∂2q tions to the return from the continuous and discontinu-
0= − qr + · µϕ + ··ν (62) ous components of the diffusion. The underlying market
∂t Z ∂x 2 ∂x2
is incomplete and the hedge strategy does not eliminate
+ ∂j q θϕ [dj] + O[dt] market risk, but this choice for the hedge portfolio opti-
j mises the return after hedging.
 12

Continuous diffusion: When θ[dj] = 0, the hedge and the portfolio with underlying weight ϕ must have funding
price conditions simplify: weight −ϕq/b, and the cumulant generating function in
  the price measure Eϕ satisfies:
−1 ∂p ∂q
p−V ·ν ·q 1
∂x ∂x log Eϕ exp[β db + ψ dq + η dν] = (69)
s= + O[dt]
V −1 q · q dt
ϕ ϕ

∂p ∂q
 β br + ψ qµ + η (ν̄ − ν)κ
δ = V −1 ·ν + qs + O[dt] 1 1
∂x ∂x + ψ 2 q 2 ν + ψη qνρσ + η 2 νσ 2 + O[dt]
∂p ∂p ϕ 1 ∂ 2 p 2 2
0= − pr + ·µ + 2
··ν (66)
∂t  ∂x 2 ∂x  where the mean rate of the underlying price and the mean
1 ∂p ∂q ∂p ∂q reversion level of the variance are adjusted:
− α −δ· ·ν −δ· + O[dt]
2 ∂x ∂x ∂x ∂x
µϕ := µ − ϕqν (70)
where: ϕ ρσ
ν̄ := ν̄ − ϕqν
∂q ∂q κ
V := ·ν (67)
∂x ∂x The calibration condition then implies that the unit op-
Embedding the expressions for the implied funding timal portfolio has ϕ = (µ − r)/(qν), so that:
spread and hedge portfolio, the price condition becomes
a non-linear parabolic differential equation for the deriva- µϕ = r (71)
tive price. This equation modifies the expression for the ρσ
ν̄ ϕ = ν̄ − (µ − r)
price to adjust for residual risk. The underlying market κ
is incomplete and the hedge strategy does not eliminate
market risk, but this choice for the hedge portfolio opti- The state is also a Heston model in the price measure, al-
mises the return after hedging. beit with these two drift parameters adjusted to account
for the change of measure. In this approach, the Heston
parameters in the price measure can be estimated from
E. Heston model statistical analysis of the state in the expectation mea-
sure, applying the measure adjustment post-estimation
to calibrate to the underlying price.
In this section, the Heston model is proposed for the
The hedge and price conditions in the Heston model
expectation measure. This concrete example of a contin-
generate the expressions:
uous Markovian diffusion illustrates the key features of
the more general framework. ∂p ∂p ρσ
δ= + + O[dt] (72)
Assumption III.5 (Heston model). The state com- ∂q ∂ν q
prises a funding price b, an underlying price q and a ∂p ∂p ∂p ∂p
variance ν whose cumulant generating function in the 0= − pr + br + qr + ((ν̄ − ν)κ − (µ − r)ρσ)
∂t ∂b ∂q ∂ν
expectation measure E satisfies: 1 ∂2p 2 ∂2p 1 ∂2p 2
+ q ν + qνρσ + νσ
1 2 ∂q 2 ∂q ∂ν 2 ∂ν 2
log E exp[β db + ψ dq + η dν] = (68)  2
dt 1 ∂p
β br + ψ qµ + η (ν̄ − ν)κ − α νσ 2 (1 − ρ2 ) + O[dt]
2 ∂ν
1 1
+ ψ 2 q 2 ν + ψη qνρσ + η 2 νσ 2 + O[dt] for the hedge portfolio δ and the derivative price p. In
2 2
the first expression, the hedge ratio is adjusted when the
The funding price has deterministic increment with mean increment of the variance is correlated with the increment
br dt where r is the funding rate. The underlying price of the underlying price. The derivative price then satisfies
has increment with mean qµ dt and variance q 2 ν dt where a linear parabolic differential equation with a non-linear
µ is the mean rate. The variance has increment with correction term that prices the unhedgeable risks when
mean (ν̄ −ν)κ dt and variance νσ 2 dt where κ is the mean the increments are not perfectly correlated.
reversion rate, ν̄ is the mean reversion level and σ is the
volatility of variance. The final parameter ρ is the cor-
relation between the increments of the underlying price F. Data-driven model
and the variance.
Since the funding security has deterministic return, the In this section, assume that the model is supported
implied funding rate is r and the implied funding spread on a finite set of states with uniform probability in the
is s = 0. In order to satisfy the self-funding condition, expectation measure.
 13

Assumption III.6 (Data-driven model). The model is are solved to generate the implied funding spread s, the
supported on a finite set of states that are equally likely in hedge portfolio δ and the initial price p for the deriva-
the expectation measure. In the nth state, the underlying tive security. While the hedge neutralises the underlying
return is dqn and the final derivative price is Pn . component of the derivative risk, when the market is in-
For this data-driven approach, the expectation mea- complete there are unhedged risks that are sensitive to
sure is represented as a uniform distribution on a finite the choice of expectation measure. The entropic pricing
set of states, which could for example be sampled from a framework thus distinguishes between the market and
generative model trained on historical market data. The model risks in the derivative security.
calibration condition is solved for the unit optimal port- Consider an alternative choice Ē for the expectation
folio ϕ via the optimisation problem: measure, related to the original expectation measure E
nX o by the Radon-Nikodym weight:
ϕ = arg minψ exp[−ψ · dqn ] : ψ · q = 0 (73) dĒ exp[ϵW ]
n
= = 1 + ϵ(W − E[W ]) + O[ϵ2 ] (78)
The price measure is then supported on the same set of dE E[exp[ϵW ]]
states but with non-uniform probabilities: for a bounded weight W and a scale parameter ϵ. Using
exp[−ϕ · dqn ] this alternative expectation measure in the calibration
γn := X (74) conditions, the implied funding rate and unit optimal
exp[−ϕ · dqm ] portfolio are adjusted, r̄ = r + ϵ ∂r and ϕ̄ = ϕ + ϵ ∂ϕ,
m
with:
The hedge and price conditions are solved simultaneously
for the hedge portfolio δ and the initial derivative price Vϕ [dq]−1 Vϕ [dq, W ] · q
∂r = + O[ϵ] (79)
p via the optimisation problem: Vϕ [dq]−1 q dt · q
∂ϕ = Vϕ [dq]−1 (Vϕ [dq, W ] − qr̄ dt) + O[ϵ]
nX o
δ = arg minϵ γn exp[−α(Pn − ϵ · dqn )] : ϵ · q = p
n
1 X Applying these adjustments, the original and alternative
p = − log γn exp[−α(Pn − δ · dqn )] (75) price measures are then related by:
α n

These expressions are amenable to machine learning Ēϕ̄ [X] − Eϕ [X] = (80)
methods and are the starting point for a variety of appli- ϕ
ϵV [W − ∂ϕ · dq, X] + O[ϵ ] 2
cations in deep hedging.
This example demonstrates the reach of the entropic Ē [α][X] − Eϕ [α][X] =
ϕ̄

framework. With no theory to support development be-

 
ϕ 1 exp[−αX]
yond the information that the data expresses, data-driven ϵV W − ∂ϕ · dq, − + O[ϵ2 ]
α Eϕ [exp[−αX]]
models need a more general foundation that does not de-
pend on unrealistic assumptions such as complete mar- for the undetermined variable X.
kets, continuous hedging or risk-free funding. Advocates Similar adjustments apply to the implied funding
of deep hedging address this by replacing dynamic risk spread, hedge portfolio and initial price, derived from the
neutralisation with dynamic risk optimisation, using con- hedge and price conditions in the alternative expectation
vex risk metrics to assess strategic performance. While measure:
they suffer from lack of explainability, data-driven mod- p̄ = δ̄ · q (81)
els can uncover complex relationships between economic
ϕ̄
variables that may be hard to capture theoretically. 0 = Ē [(dq − q(r̄ + αs̄) dt) exp[−α(P − δ̄ · dq)]]
p̄ = Ēϕ̄ [α][P − δ̄ · dq]
IV. MARKET AND MODEL RISK Focusing on the price condition, the adjustment to the
initial price is p̄ = p + ϵ ∂p with:
Continuing from the previous section, with the expec- 1
tation measure E the calibration conditions: ∂p = × (82)
1 + (r + αs) dt
 
0=ϕ·q (76) 1 exp[−α(P − δ · dq)]
Vϕ W − ∂ϕ · dq, − + O[ϵ]
ϕ
0 = E [dq] − qr dt α Eϕ [exp[−α(P − δ · dq)]]

are solved to generate the implied funding rate r and the The hedge portfolio is the metric of market risk in the
unit optimal portfolio ϕ for the underlying securities, and entropic framework, and this adjustment is the metric
the hedge and price conditions: of model risk as it measures the sensitivity of the initial
price to the choice of expectation measure. If the un-
p=δ·q (77) derlying market is complete, the variance of the hedged
ϕ
0 = E [(dq − q(r + αs) dt) exp[−α(P − δ · dq)]] return is zero and the sensitivity to the expectation mea-
sure is eliminated. More generally, model risk increases
p = Eϕ [α][P − δ · dq] in proportion with the degree of market incompleteness.
 14

V. ENTROPIC XVA given by the condition:

 
c1
An unconstrained strategy that maximises the mean E1 [α1 ] d ≥0 (85)
b1
does not incentivise risk management and results in a  
portfolio with potentially disastrous downside risks. For c2
E2 [α2 ] d ≥0
this reason, a strategy that instead aims to maximise the b2
risk-adjusted mean creates a more balanced portfolio that
on the entropy-adjusted mean return on capital. To avoid
can be fine-tuned to the risk appetite of the investor.
the possibility of funding arbitrage, assume that the mar-
It would appear from this statement that risk manage-
gin accounts have the same price, b1 = b2 := b. With this
ment is an investment choice, with the risk appetite a
assumption, the return simplifies:
free parameter under the control of the investor. This
is not the case. Without risk management there is a c1 P̄ p
material risk that the investor will not meet their con- d = − (86)
b b + db b
tractual commitments. At the very least, this potentially c2 p P̄
tarnishes the reputation of the investor as a reliable coun- d = −
terparty, with consequent impact on the prices they will b b b + db
be offered. In regulated markets, risk management is also where:
a legal requirement with detailed and specific obligations
for margin and capital. P̄ := (87)
If the true costs of risk management are fully accounted max[−(c1 − p)(1 + db/b), min[(c2 + p)(1 + db/b), P ]]
in the return, reckless investments are naturally penalised
is the final price floored such that the first counterparty
by their associated margin costs. Returning to the bilat-
defaults when the floor is breached and capped such
eral model outlined in the introduction, suppose that two
that the second counterparty defaults when the cap is
counterparties deploy capital c1 and c2 to a transaction
breached.
with initial price p, where −c2 ≤ p ≤ c1 . Excess capital
The returns for the counterparties are equal and oppo-
is deposited in the margin account with price b1 > 0 and
site. This can only be viable if the counterparties have
b2 > 0 respectively. The transaction is settled from the
differing economic expectations and therefore attribute
margin account with no additional funds injected, creat-
different values to the transaction. The return is mono-
ing the risk that insufficient capital results in default by
tonic in the initial price, and the threshold price p1 and
one of the counterparties.
p2 solves the implicit relationship:
Net of all considerations, the returns are:
 
p1 P̄1
c1 = E1 [α1 ] (88)
d = (83) b b + db
b1  
 +  + p2 P̄2
c1 p b2 + db2 p c2 p = E2 [α2 ]
− −d − d − +d b b + db
b1 b1 b1 + db1 b2 b2 b1
c2 With these thresholds, the transaction is viable when the
d =
b2 initial price p is in the range p2 ≤ p ≤ p1 , and is there-

p c2
+
b1 + db1

c1 p
+
p fore not viable when p1 < p2 . Price is thus constrained
d − − − −d −d to the sector bounded by these viability thresholds, equi-
b2 b2 b2 + db2 b1 b1 b2 librating to a price within the range. Since the entropy-
adjusted mean is monotonic as a function of the notional,
To guarantee that default is avoided, the capital must
the range narrows as the notional increases and there is
exceed the maximum liability:
an equilibrium point at which no further transactions are

P
 viable. Equilibrium is thus discovered as the solution of
c1 ≥ p − ess inf (84) the relationships:
1 + db1 /b1
 
p P̄
 
P
c2 ≥ ess sup −p = E1 [λα1 ] (89)
1 + db2 /b2 b b + db
 
p P̄
where P := p + dp is the final price. The first counter- = E2 [λα2 ]
b b + db
party is at risk of default if the first condition is not met,
and the second counterparty is at risk of default if the for the initial price per unit notional p and the equi-
second condition is not met. These default risks need to librium notional λ. Terminating when the transaction
be accounted for in the valuation. exceeds the risk appetite of the investor, this procedure
Viability of the transaction is assessed by the counter- balances the differing expectations of the counterparties
party using their expectation measure E1 and E2 and risk to identify their equilibrium price, accounting for their
adjustment α1 and α2 , with the threshold for execution respective funding costs and risks of default.
 15

Entropy-Adjusted Mean Return On Capital Entropy-Adjusted Mean Return On Capital

1 1

First First
μ = -1 μ = -1
0 μ = -0.5 0 μ = -0.5
0 1 μ=0 0 1 μ=0
μ = 0.5 μ = 0.5
μ=1 μ=1

-1 -1
Option Price Per Unit Notional Option Price Per Unit Notional

λ = 0 and c = ∞ λ = 0 and c = 0.5

Entropy-Adjusted Mean Return On Capital Entropy-Adjusted Mean Return On Capital

1 1

First First
μ = -1 μ = -1
0 μ = -0.5 0 μ = -0.5
0 1 μ=0 0 1 μ=0
μ = 0.5 μ = 0.5
μ=1 μ=1

-1 -1
Option Price Per Unit Notional Option Price Per Unit Notional

λ = 1 and c = ∞ λ = 1 and c = 0.5

Entropy-Adjusted Mean Return On Capital Entropy-Adjusted Mean Return On Capital

1 1

First First
μ = -1 μ = -1
0 μ = -0.5 0 μ = -0.5
0 1 μ=0 0 1 μ=0
μ = 0.5 μ = 0.5
μ=1 μ=1

-1 -1
Option Price Per Unit Notional Option Price Per Unit Notional

λ = 2 and c = ∞ λ = 2 and c = 0.5

Figure 8. In this example, two counterparties have differing expectations for the future state. The first counterparty models the
state as a uniform variable with zero mean and unit standard deviation. The second counterparty models the state as a normal
variable with mean µ and unit standard deviation. The transaction is then a call option on the state with zero strike. In each
graph, the entropy-adjusted mean return on capital, computed with risk adjustment α = 1, is plotted for both counterparties as
a function of the option price per unit notional, including a range of values for the mean µ assumed by the second counterparty.
The first counterparty is fully capitalised against default, and each row includes cases c = ∞, 0.5 for the capital of the second
counterparty. Each column includes cases λ = 0, 1, 2 for the notional, where the case λ = 0 corresponds to mid-pricing.
The transaction is viable when the entropy-adjusted mean return on capital is positive for both counterparties. The first
counterparty buys the option when its price is low, and the second counterparty sells the option when its price is high. The
second counterparty is more likely to sell when the mean µ is lower, as this reduces the expected option payoff. Model risk
increases as the notional λ is increased, which pushes the threshold prices in opposite directions and makes the transaction less
viable. Finally, for the second counterparty, decreasing the capital c increases the risk of default. The CVA cost to the first
counterparty and DVA benefit to the second counterparty then pushes both threshold prices lower.
 16

VI. LITERATURE REVIEW timisation [14] and derivative pricing [15, 16] – see also
the review essay [17] which includes further references.
Understanding the relationship between the economic The article [18] by Frittelli proposes the minimal relative
and price measures has been the fundamental question entropy measure as a solution to the problem of pric-
of mathematical finance ever since Bachelier first applied ing in incomplete markets, and links this solution with
stochastic calculus in his pioneering thesis [1]. In this the maximisation of expected exponential utility. The
thesis, Bachelier develops a model for the logarithm of entropic solution is connected with risk optimisation in
the security price as Brownian motion around its equi- the introduction [19], which this essay extends with the
librium, deriving expressions for options that would not adoption of the entropic risk metric. As a measure of dis-
look unfamiliar today. Methods of portfolio optimisation order, entropy performs a role similar to variance but is
originated in the work of Markowitz [2] and Sharpe [3] better suited to the real distributions of market returns.
using Gaussian statistics for the returns, generalised by Proponents of the use of entropy justify the approach by
later authors to allow more sophisticated distributional appeal to the modelling of information flows in dynamical
assumptions and measures of utility. These approaches systems and the analogy with thermodynamics.
are embedded in the economic measure, explaining the While market evidence for discounting basis existed
origin of price as the equilibrium of market activity un- earlier, the basis widening that occurred as a result of
covering the expectations of participants. the Global Financial Crisis of 2007-2008 motivated re-
The impact of dynamic hedging on price was first search into funding and its impact on discounting. Work
recognised in the articles by Black and Scholes [4] and by Johannes and Sundaresan [20], Fujii and Takahashi
Merton [5]. Adopting the framework devised by Bache- [21], Piterbarg and Antonov [22–24], Henrard [25], Mc-
lier, these authors observe that the option return is repli- Cloud [26] and others established its theoretical justifi-
cated exactly by a strategy that continuously offsets the cation and practical application.
delta of the option price to the underlying price. The Most approaches to derivative pricing begin with the
theory matured with the work of Harrison and Pliska assumption of continuous settlement, and stochastic cal-
[6, 7], providing a precise statement of the conditions culus is an essential ingredient for these developments.
for market completeness and the martingale property of The representation of the stochastic differential equation
price in a measure equivalent to the economic measure. used in this article follows the discoveries of Lévy [27],
The discipline has since expanded in numerous directions, Khintchine [28] and Itô [29]. The technical requirements
with significant advances for term structure and default of the Lévy-Khintchine representation can be found in
modelling and numerical methods for complex derivative these articles and other standard texts in probability the-
structures. ory. The change from economic measure to price measure
Closer consideration of market incompleteness led to implied by the entropic principle extends the result from
the development of alternative pricing principles based Girsanov [30] to include the scaling adjustment of the
on the dynamic optimisation of convex risk metrics, as jump frequency in addition to the drift adjustment. The
presented in [8–10]. With foundations that extend well example model referenced here is the Heston model [31],
beyond the domain of applicability of risk-neutral pric- a popular model in quantitative finance.
ing, these principles enabled the development of data- It is remarkable that the principle of dynamic convex
driven methods including the pioneering work on deep risk optimisation enables the consistent and simultane-
hedging by Buehler and others [11, 12]. As a recent prac- ous application of models across electronic trading algo-
tical example, Oya [13] applies these methods to explain rithms, for pricing and hedging exotic derivatives, and to
the mark-to-market valuation adjustment for Bermudan determine economic and regulatory capital. In this con-
swaptions. text, the entropic risk metric is promoted for its simple
Entropy methods have found application across the implementation as an extension of the risk-neutral frame-
range of mathematical finance, including portfolio op- work, which supports integration into legacy platforms.

[1] L. Bachelier, Théorie de la spéculation, Annales scien- [5] R. C. Merton, Theory of Rational Option Pricing, The
tifiques de l’École Normale Supérieure 3e série, 17, 21 Bell Journal of Economics and Management Science 4,
(1900). 141 (1973).
[2] H. Markowitz, Portfolio Selection, The Journal of Fi- [6] J. M. Harrison and S. R. Pliska, Martingales and stochas-
nance 7, 77 (1952). tic integrals in the theory of continuous trading, Stochas-
[3] W. F. Sharpe, Capital asset prices: A theory of mar- tic Processes and their Applications 11, 215 (1981).
ket equilibrium under conditions of risk, The Journal of [7] J. M. Harrison and S. R. Pliska, A stochastic calcu-
Finance 19, 425 (1964). lus model of continuous trading: Complete markets,
[4] F. Black and M. Scholes, The Pricing of Options and Stochastic Processes and their Applications 15, 313
Corporate Liabilities, Journal of Political Economy 81, (1983).
637 (1973). [8] M. Xu, Risk measure pricing and hedging in incomplete
 17

markets, Annals of Finance 2, 51 (2006). [20] M. Johannes and S. Sundaresan, The impact of collater-
[9] S. Klöppel and M. Schweizer, Dynamic indifference val- alization on swap rates, The Journal of Finance 62, 383
uation via convex risk measures, Mathematical Finance (2007).
17, 599 (2007). [21] M. Fujii and A. Takahashi, Choice of collateral currency,
[10] A. Ilhan, M. Jonsson, and R. Sircar, Optimal static- Risk Magazine 24, 120 (2011).
dynamic hedges for exotic options under convex risk mea- [22] V. Piterbarg, Funding beyond discounting: collateral
sures, Stochastic Processes and their Applications 119, agreements and derivatives pricing, Risk Magazine 23,
3608 (2009). 97 (2010).
[11] H. Buehler, L. Gonon, J. Teichmann, and B. Wood, Deep [23] V. Piterbarg, Cooking with collateral, Risk Magazine 25,
hedging, Quantitative Finance 19, 1271 (2019). 46 (2012).
[12] H. Buehler, P. Murray, M. S. Pakkanen, and B. Wood, [24] A. Antonov and V. Piterbarg, Options for collateral op-
Deep hedging: learning to remove the drift under trading tions, Risk Magazine , 66 (2014).
frictions with minimal equivalent near-martingale mea- [25] M. Henrard, Interest rate modelling in the multi-curve
sures, arXiv (2021), 2111.07844. framework: Foundations, evolution and implementation
[13] K. Oya, Deep hedging bermudan swaptions, arXiv (Springer, 2014).
(2024), 2411.10079. [26] P. McCloud, Collateral convexity complexity, Risk Mag-
[14] G. C. Philippatos and C. J. Wilson, Entropy, market azine 26, 60 (2013).
risk, and the selection of efficient portfolios, Applied Eco- [27] P. Lévy, Sur les intégrales dont les éléments sont des vari-
nomics 4, 209 (1972). ables aléatoires indépendantes, Annali della Scuola Nor-
[15] P. W. Buchen and M. Kelly, The maximum entropy dis- male Superiore di Pisa-Classe di Scienze 3, 337 (1934).
tribution of an asset inferred from option prices, Journal [28] A. Khintchine, A new derivation of one formula by Levy
of Financial and Quantitative Analysis , 143 (1996). P., Bulletin of Moscow State University 1, 1 (1937).
[16] L. Gulko, The entropy theory of stock option pricing, [29] K. Itô, On stochastic processes (I), in Japanese journal
International Journal of Theoretical and Applied Finance of mathematics: transactions and abstracts, Vol. 18 (The
2, 331 (1999). Mathematical Society of Japan, 1941) pp. 261–301.
[17] R. Zhou, R. Cai, and G. Tong, Applications of entropy [30] I. V. Girsanov, On Transforming a Certain Class of
in finance: A review, Entropy 15, 4909 (2013). Stochastic Processes by Absolutely Continuous Substi-
[18] M. Frittelli, The minimal entropy martingale measure tution of Measures, Theory of Probability and its Appli-
and the valuation problem in incomplete markets, Math- cations 5, 285 (1960).
ematical Finance 10, 39 (2000). [31] S. L. Heston, A Closed-Form Solution for Options with
[19] P. McCloud, Expectation and Price in Incomplete Mar- Stochastic Volatility with Applications to Bond and Cur-
kets, arXiv (2020), 2006.16703. rency Options, The Review of Financial Studies 6, 327
(1993).