Model Risk Analysis for Entropic Hedging Strategies

arXiv:2610.06168 · q-fin.MF · Submitted 2026-10-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Model Risk Analysis for Entropic Hedging Strategies".

Warren: Here's the gist of the paper. Entropic risk optimisation is a general framework for pricing and hedging financial derivatives in incomplete markets that can be used to decompose P&L into market and model risk contributions.

Ruth: The single most informative sentence about the paper's subject and finding is Entropic risk optimisation is a general framework for pricing and hedging financial derivatives in incomplete markets that can be used to decompose P&L into market and model risk contributions.

Warren: Quadratic Gaussian Framework The paper introduces the quadratic Gaussian theorem, which allows for the decomposition of a quadratic Gaussian variable into a weighted sum of independent standard normal and noncentral chi-squared variables.

Ruth: This decomposition is crucial for understanding how trading P&L can be decomposed into convexity, dimension and funding contributions that are attributed in the explanation of realised P&L.

Paper summary: Ruth: So, this paper, "Model Risk Analysis for Entropic Hedging Strategies," it basically sets up a whole framework for pricing and hedging financial derivatives in markets that aren't complete. The main idea is to take the profit and loss, P andL, and break it down into what’s just market risk versus what’s coming from the model itself.

Warren: Right, so they use this entropic risk optimization framework to do that decomposition. It claims you can get a clear attribution of that P andL into convexity, dimension, and funding contributions. That's the core claim for understanding realized P andL >

Iris: From an econometric standpoint, what catches my eye is how they use the quadratic Gaussian theorem here. They decompose a quadratic Gaussian variable into a weighted sum of standard normal and noncentral chi-squared variables. This decomposition is what lets them attribute those contributions to convexity, dimension, and funding >

Ruth: Exactly. They write this variable X as a linear combination of independent variables, including the standard normal variable W and noncentral chi-squared variables CI > So they’re using this theorem to understand how trading P andL can be broken down into these three specific buckets: convexity, dimension, and funding contributions >

Warren: And then they use this decomposition to show how you can maximize the entropy-adjusted mean for a self-funded portfolio. They look at two main scenarios: just the underlying or when you’re holding both the derivative and the underlying >

Iris: That's interesting. Theorem II.one identifies a unique self-funded underlying portfolio that maximizes that entropy-adjusted mean for just the underlying market > But then in Theorem II <ref:2610.06168#pg1>.two they give you an expression for how much incremental entropy-adjusted mean you get per unit of derivative notional when you consider both the derivative and the underlying simultaneously > <ref:2610.06168#pg2>

Ruth: And that leads into their P andL decomposition. They assess P andL at the initial time using risk penalties, which they subtract from that entropy-adjusted mean. Then at the final time, they explain what all that P andL is made of by summing up terms related to convexity, dimension, funding, and strategy >

Warren: The risk penalties themselves are non-negative numbers and they’re defined as the incremental entropy-adjusted means generated by those attributions > And then there's this concept of indifference pricing which adds a constraint that aligns the maximum entropy-adjusted means between the underlying only market and the derivative and underlying market >

Iris: Theorem II.six gives you that indifference pricing condition, which is quadratic in the funding spread s star, and it shows those risk penalties from convexity, dimension, funding, and strategy are non-negative numbers that can't be eliminated by replication > So it’s saying these risk components are inherent limitations > <ref:2610.06168#pg1>

Ruth: And they extend this using an Ornstein-Uhlenbeck process to handle multiple hedging intervals. They generate standard normal variables on each interval based on the initial state xzero driven by exponential terms and standard normal variables > This lets them derive the incremental entropy-adjusted mean per unit of derivative notional for each interval, showing how those risk penalties are calculated based on local parameters >

Warren: It’s a detailed look at how these local parameters affect the risk attribution across different time steps. It really moves beyond a single point in time analysis > So where does this lead us when we look at the actual strategy and hedging decisions?

Conclusion: Ruth: Looking at "Model Risk Analysis for Entropic Hedging Strategies," the paper’s main contribution is showing how to use an entropic risk optimization framework to price and hedge derivatives in incomplete markets > It provides a way to decompose P andL into market and model risk contributions, specifically isolating convexity, dimension, and funding risks >

Warren: The authors are using the quadratic Gaussian theorem to solve this. They show that when prices follow Gaussian processes like fractional Ornstein-Uhlenbeck processes, you can get closed-form solutions for both the price and the hedge ratios > That’s useful because it enables a comprehensive analysis of trading P andL >

Iris: From my view, what they've done is rigorously defined these risk penalties from convexity, dimension, funding, and strategy as non-negative numbers that you can’t just replicate away > That sets a firm boundary on what replication can actually achieve in these situations >

Ruth: And the title itself points to this—it’s about analyzing model risk within hedging strategies. The implications are that if you're dealing with incomplete markets, you need a way to explicitly account for these model risks when trying to manage your hedge >

Warren: For someone who just listens to the show, what this means is that when you’re trading something complex, it’s not just about the price movement; it’s about understanding how much of your profit or loss comes from the structure of the model you used versus just market noise >

Iris: And they also showed how indifference pricing works by making a constraint on those maximum entropy-adjusted means between different market scenarios > It gives you a condition, quadratic in the funding spread s star, that fixes both the funding spread and the derivative price at the start of each interval >

Ruth: So it’s a tool for assessing model risk contribution and figuring out how to structure hedges when replication isn't perfect because of things like convexity or dimension > The authors show you how these structural risks generate penalties that influence your pricing and hedging decisions across multiple time steps >

Warren: The paper offers parametric expressions for hedge ratios that can be used in regressions for deep hedging, which is a practical application > But they also flag something important: the method itself is based on the assumption of quadratic Gaussian variables, and if you move away from that, you’re not using this specific framework >

McCloud Research

q-fin.MF

Submitted: 2026-10-05

Updated: 2026-11-02

Comments: Presented at QuantMinds 2026

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 71/100

The gist: Entropic risk optimisation is a general framework for pricing and hedging financial derivatives in incomplete markets that can be used to decompose P&L into market and model risk contributions.

Key concepts

Entropic Risk Optimisation
A general framework for pricing derivatives in incomplete markets that decomposes profit and loss (P&L) into market risk and model risk contributions. It uses an entropic measure to guide hedging decisions, aiming to maximize the expected P&L adjusted for this risk metric.
Quadratic Gaussian Framework
A mathematical tool used to decompose a quadratic Gaussian variable into a weighted sum of independent standard normal and noncentral chi-squared variables. This decomposition is essential for attributing realized P&L into specific components like convexity, dimension, and funding contributions.
Risk Penalties
Non-negative values subtracted from the entropy-adjusted mean to assess risk contributions from convexity, dimension, funding, and strategy. These penalties are derived from the attributions of the total P&L at different time points.
Indifference Pricing
A constraint that aligns the maximum entropy-adjusted means between different market scenarios (e.g., underlying only versus derivative and underlying). This condition is quadratic in the funding spread and helps determine initial prices for hedging.

Terminology

Summary

Entropic risk optimisation is a general framework for pricing and hedging financial derivatives in incomplete markets that can be used to decompose P&L into market and model risk contributions. p. 0

Where the paper says it

Model Risk Analysis for Entropic Hedging Strategies Paul McCloud McCloud Research (Dated: 5 October 2026) Entropic risk optimisation is a general framework for pricing and hedging financial derivatives in incomplete markets that can be used to decompose P&L into market and model risk contributions. When the prices are quadratic Gaussian, the coupled equations for price and hedge ratios are solved in closed form.

The single most informative sentence about the paper's subject and finding is Entropic risk optimisation is a general framework for pricing and hedging financial derivatives in incomplete markets that can be used to decompose P&L into market and model risk contributions.

Quadratic Gaussian Framework

The paper introduces the quadratic Gaussian theorem, which allows for the decomposition of a quadratic Gaussian variable into a weighted sum of independent standard normal and noncentral chi-squared variables. This decomposition is crucial for understanding how trading P&L can be decomposed into convexity, dimension and funding contributions that are attributed in the explanation of realised P&L. p. 0

Where the paper says it

When the prices are quadratic Gaussian, the coupled equations for price and hedge ratios are solved in closed form. This enables comprehensive analysis of trading P&L, with a decomposition of the model value-at-risk into convexity, dimension and funding contributions that are attributed in the explanation of realised P&L.

The variable X is written as a linear combination of independent variables, including the standard normal variable W, and noncentral chi-squared variables CI. p. 0
Where the paper says it

Since the zero eigenvalue needs to be handled separately, this is made explicit in the spectral decomposition: 1 = 10 + X I̸=0 1I (3) γ = X I̸=0 γI 1I in terms of the orthogonal projections 1I onto the eigenspaces for the distinct eigenvalues γI , where the case I = 0 corresponds to the zero eigenvalue γ0 = 0. The variable X is written as: X = η − X I̸=0 ∥1I ϵ∥ 2 2γI + ∥10ϵ∥W + 1 2 X I̸=0 γICI (4) This expression decomposes the variable into a linear combination of independent variables, including the standard normal variable: W := 10ϵ · 10Z ∥10ϵ∥ (5) when ∥10ϵ∥ ̸= 0, and the noncentral chi-squared variables: CI := ∥1I (Z + γ −1 I ϵ)∥ 2 (6) with noncentrality parameter ∥1I ϵ/γI ∥ 2 and degrees of freedom rank[1I ] for each I ̸= 0.

Pricing and Hedging Strategies

The paper details how to maximize the entropy-adjusted mean for a self-funded portfolio by considering two main scenarios: underlying only and derivative and underlying. For the underlying only portfolio, Theorem II. p. 0

Where the paper says it

CONTENTS I. Quadratic Gaussian theorem 1 II.

p. 2
Where the paper says it

QUADRATIC GAUSSIAN PRICING This section introduces the following variables: q (Nm, 1) Underlying initial prices p (1, 1) Derivative initial price Q (Nm, 1) Underlying final prices P (1, 1) Derivative final price ω, δ (Nm, 1) Underlying notionals λ (1, 1) Derivative notional µ, ¯ µ˘ (Nm, 1) Mean of underlying prices σ, ¯ σ˘ (Nm, Nf ) Volatility of underlying prices ν, ¯ ν˘ (Nm, Nm) Covariance of underlying prices η, ¯ η˘ (1, 1) Mean of derivative price ϵ,¯ ϵ˘ (Nf , 1) Volatility of derivative price γ, ¯ γ˘ (Nf , Nf ) Convexity of derivative price which are listed with their dimensions, where Nm is the number of underlying securities. The underlying and derivative final prices are modelled as quadratic Gaussian variables: Q := ¯µ + ¯σZ (14) P := ¯η + ¯ϵ · Z + 1 2 Z · γZ¯ For a portfolio comprising notional λ of the derivative security and notionals ω of the underlying securities, the portfolio final price is the quadratic Gaussian variable: λP +ω·Q = (λη¯+ω·µ¯)+ (λϵ¯+ ¯σ †ω)·Z + 1 2 Z ·λγZ¯ (15) Using the quadratic Gaussian theorem, the entropyadjusted mean of the final price is: E[α][λP + ω · Q] = (16) 1 2α log det[1 + αλγ¯] + (λη¯ + ω · µ¯) − 1 2 α(λϵ¯+ ¯σ †ω) · (1 + αλγ¯) −1 (λϵ¯+ ¯σ †ω) which is valid for αλ within an open interval containing zero.

1 identifies a unique self-funded underlying portfolio that maximizes the entropy-adjusted mean. p. 2
Where the paper says it

Let ν¯ := ¯σσ¯ † be the covariance of the underlying final prices and assume that it is invertible. This enables the definitions: σ¯ −1 := ¯σ † ν¯ −1 σ¯ −† := ¯ν −1σ¯ (19) 1ν¯ := ¯σ † ν¯ −1σ¯ of the right-inverse σ¯ −1 , left-inverse σ¯ −† and projection 1ν¯, satisfying (¯σ −1 ) † = ¯σ −† and: σ¯σ¯ −1 = 1 = ¯σ −†σ¯ † (20) σ¯ −1σ¯ = 1ν¯ = ¯σ †σ¯ −† The unique self-funded underlying portfolio ω ∗ generating the maximum entropy-adjusted mean is obtained as the solution of the self-funding and stationary conditions: ϕ · q = 0 (21) µ¯ − (1 + r)q − νϕ¯ = 0 where r is the funding rate used to implement the selffunding condition and ϕ := αω∗ .

When considering the derivative and underlying portfolio, Theorem II. p. 2
Where the paper says it

QUADRATIC GAUSSIAN PRICING This section introduces the following variables: q (Nm, 1) Underlying initial prices p (1, 1) Derivative initial price Q (Nm, 1) Underlying final prices P (1, 1) Derivative final price ω, δ (Nm, 1) Underlying notionals λ (1, 1) Derivative notional µ, ¯ µ˘ (Nm, 1) Mean of underlying prices σ, ¯ σ˘ (Nm, Nf ) Volatility of underlying prices ν, ¯ ν˘ (Nm, Nm) Covariance of underlying prices η, ¯ η˘ (1, 1) Mean of derivative price ϵ,¯ ϵ˘ (Nf , 1) Volatility of derivative price γ, ¯ γ˘ (Nf , Nf ) Convexity of derivative price which are listed with their dimensions, where Nm is the number of underlying securities. The underlying and derivative final prices are modelled as quadratic Gaussian variables: Q := ¯µ + ¯σZ (14) P := ¯η + ¯ϵ · Z + 1 2 Z · γZ¯ For a portfolio comprising notional λ of the derivative security and notionals ω of the underlying securities, the portfolio final price is the quadratic Gaussian variable: λP +ω·Q = (λη¯+ω·µ¯)+ (λϵ¯+ ¯σ †ω)·Z + 1 2 Z ·λγZ¯ (15) Using the quadratic Gaussian theorem, the entropyadjusted mean of the final price is: E[α][λP + ω · Q] = (16) 1 2α log det[1 + αλγ¯] + (λη¯ + ω · µ¯) − 1 2 α(λϵ¯+ ¯σ †ω) · (1 + αλγ¯) −1 (λϵ¯+ ¯σ †ω) which is valid for αλ within an open interval containing zero.

2 provides an expression for the incremental entropy-adjusted mean per unit of derivative notional. p. 3
Where the paper says it

2 (Measure and convexity). The incremental entropy-adjusted mean per unit of derivative notional is: 1 λ (E[α][λP + (ω ∗ − λδ) · Q] − E[α][ω ∗ · Q]) (26) = (˘η − (1 + r)p) − 1 2 αλ ∥ϵ˘− σ˘ † δ∥ 2 + tr αλγ¯ − log[1 + αλγ¯] (αλ) 2 where the parameters are adjusted for measure and convexity: σ˘ := ¯σ(1 + αλγ¯) −1/2 (27) η˘ := ¯η − ϵ¯· σ¯ †ϕ + 1 2 σ¯ †ϕ · γ¯σ¯ †ϕ + 1 2 tr[¯γ] ϵ˘ := (1 + αλγ¯) −1/2 (¯ϵ − γ¯σ¯ †ϕ) Proof.

Profit and Loss Decomposition

The paper establishes a method for assessing the profit and loss P − δ · Q by first assessing it at initial time by risk penalties, which are the risk penalties they subtract from the entropy-adjusted mean. The P&L is then explained at final time by its attribution, where the total P&L is given by a sum of terms related to convexity, dimension, funding and strategy. p. 5

Where the paper says it

The P&L of the funded derivative is P − δ · Q. The P&L is assessed at initial time by the risk penalties: χ γ := tr αλγ¯ − log[1 + αλγ¯] (αλ) 2 (51) χ σ := ∥(1 − 1ν˘)˘ϵ∥ 2 χ s := s 2 δ s · q χ δ := ∥σ˘ † (δ − δ ∗ )∥ 2 for convexity, dimension, funding and strategy, and the entropy-adjusted mean of the P&L is: 1 λ (E[α][ω ∗ · Q + λ(P − δ · Q)] − E[α][ω ∗ · Q]) = (52) (˘η − (1 + r)p) − 1 2 αλ(χ γ + χ σ + χ s + χ δ ) The P&L is explained at final time by the attribution: Xγ := 1 2 (Y · γY¯ − tr[¯γ]) (53) Xσ := (¯ϵ − γ¯σ¯ †ϕ) · Y − δ σ · (Q − (1 + r)q) Xs := −sδs · (Q − (1 + r)q) Xδ := −(δ − δ ∗ ) · Q for convexity, dimension, funding and strategy, where Y := Z + ¯σ †ϕ are the measure-adjusted factors, and the total P&L is: P − δ · Q = (54) (˘η − (1 + r)p) + (Xγ + Xσ + Xs + Xδ ) The risk penalties are non-negative and are the incremental entropy-adjusted means generated by the attributions: − 1 2 αλχγ = 1 λ (E[α][ω ∗ · Q + λXγ ] (55) − E[α][ω ∗ · Q]) − 1 2 αλχσ = 1 λ (E[α][ω ∗ · Q + λ(Xγ + Xσ )] − E[α][ω ∗ · Q + λXγ ]) − 1 2 αλχs = 1 λ (E[α][ω ∗ · Q + λ(Xγ + Xσ + Xs )] − E[α][ω ∗ · Q + λ(Xγ + Xσ )]) − 1 2 αλχδ = 1 λ (E[α][ω ∗ · Q + λ(Xγ + Xσ + Xs + Xδ )] − E[α][ω ∗ · Q + λ(Xγ + Xσ + Xs )]) Proof.

The risk penalties are non-negative and are the incremental entropy-adjusted means generated by the attributions. p. 5
Where the paper says it

The P&L of the funded derivative is P − δ · Q. The P&L is assessed at initial time by the risk penalties: χ γ := tr αλγ¯ − log[1 + αλγ¯] (αλ) 2 (51) χ σ := ∥(1 − 1ν˘)˘ϵ∥ 2 χ s := s 2 δ s · q χ δ := ∥σ˘ † (δ − δ ∗ )∥ 2 for convexity, dimension, funding and strategy, and the entropy-adjusted mean of the P&L is: 1 λ (E[α][ω ∗ · Q + λ(P − δ · Q)] − E[α][ω ∗ · Q]) = (52) (˘η − (1 + r)p) − 1 2 αλ(χ γ + χ σ + χ s + χ δ ) The P&L is explained at final time by the attribution: Xγ := 1 2 (Y · γY¯ − tr[¯γ]) (53) Xσ := (¯ϵ − γ¯σ¯ †ϕ) · Y − δ σ · (Q − (1 + r)q) Xs := −sδs · (Q − (1 + r)q) Xδ := −(δ − δ ∗ ) · Q for convexity, dimension, funding and strategy, where Y := Z + ¯σ †ϕ are the measure-adjusted factors, and the total P&L is: P − δ · Q = (54) (˘η − (1 + r)p) + (Xγ + Xσ + Xs + Xδ ) The risk penalties are non-negative and are the incremental entropy-adjusted means generated by the attributions: − 1 2 αλχγ = 1 λ (E[α][ω ∗ · Q + λXγ ] (55) − E[α][ω ∗ · Q]) − 1 2 αλχσ = 1 λ (E[α][ω ∗ · Q + λ(Xγ + Xσ )] − E[α][ω ∗ · Q + λXγ ]) − 1 2 αλχs = 1 λ (E[α][ω ∗ · Q + λ(Xγ + Xσ + Xs )] − E[α][ω ∗ · Q + λ(Xγ + Xσ )]) − 1 2 αλχδ = 1 λ (E[α][ω ∗ · Q + λ(Xγ + Xσ + Xs + Xδ )] − E[α][ω ∗ · Q + λ(Xγ + Xσ + Xs )]) Proof.

Indifference Pricing and Risk Penalties

Indifference pricing imposes an additional constraint that aligns the maximum entropy-adjusted means of the underlying only market and the underlying and derivative market. Theorem II.6 gives the indifference pricing condition, which is quadratic in the funding spread s∗. p. 7

Where the paper says it

(66) The indifferent derivative initial price p ∗ is the discounted and penalised adjusted mean: p ∗ := 1 1 + r η˘ − 1 2 αλ(χ γ + χ σ + χ s ∗ ) (67) where the funding penalty is evaluated at the indifferent funding spread. The indifference pricing condition is quadratic in the funding spread, which enters only through the total funding, and rearranges to: (1 + r + αλs∗ ) 2 = (1 + r) 2 1 + 2αλζ (1 + r) 2δ s · q (68) Taking the square root, and using 1+r > 0, gives the two branches for the funding spread.

The risk penalties arising from convexity, dimension, funding and strategy are shown to be non-negative and cannot be eliminated by replication. p. 6
Where the paper says it

The penalties arise from three sources of incompleteness in the underlying market and the suboptimality of the funding strategy. Convexity: The convexity between the derivative and underlying final prices cannot be eliminated by linear replication, which leaves residual risks in the portfolio when ¯γ ̸= 0.

Ornstein-Uhlenbeck Process Application

The paper extends the framework to multiple hedging intervals using the Ornstein-Uhlenbeck process to generate standard normal variables on each interval. The drivers for each interval are constructed iteratively from x0 via a formula involving exponential terms and standard normal variables. p. 1

Where the paper says it

The decomposition of X into a standard normal and noncentral chi-squared variables follows from the spectral decomposition of γ as provided. Use this decomposition together with the orthogonality 1I 1J = δIJ 1I and symmetry 1† I = 1I of the projections to generate the decomposition of X: X = η + 10ϵ · 10Z (10) + X I̸=0 (1I ϵ · 1IZ + 1 2 γI 1IZ · 1IZ) = η − X I̸=0 ∥1I ϵ∥ 2 2γI + ∥10ϵ∥ 10ϵ · 10Z ∥10ϵ∥ + 1 2 X I̸=0 γI ∥1I (Z + γ −1 I ϵ)∥ 2 The expectation is given by the ratio of Gaussian integrals: E[exp[−X]f[Z]] = (11) R z exp[−η − ϵ · z − 1 2 z · (1 + γ)z]f[z] dz R z exp[− 1 2 z · z] dz Complete the square in the exponent via the change of variables: y = (1 + γ) 1/2 z + (1 + γ) −1/2 ϵ (12) z = (1 + γ) −1/2 y − (1 + γ) −1 ϵ Use this transformation to generate the expressions: dz = 1 p det[1 + γ] dy (13) −ϵ · z − 1 2 z · (1 + γ)z = 1 2 ϵ · (1 + γ) −1 ϵ − 1 2 y · y These expressions are inserted into the integrals to complete the proof.

The incremental entropy-adjusted mean per unit of derivative notional is then derived for each interval, showing how the risk penalties are calculated based on the local parameters. p. 3
Where the paper says it

2 (Measure and convexity). The incremental entropy-adjusted mean per unit of derivative notional is: 1 λ (E[α][λP + (ω ∗ − λδ) · Q] − E[α][ω ∗ · Q]) (26) = (˘η − (1 + r)p) − 1 2 αλ ∥ϵ˘− σ˘ † δ∥ 2 + tr αλγ¯ − log[1 + αλγ¯] (αλ) 2 where the parameters are adjusted for measure and convexity: σ˘ := ¯σ(1 + αλγ¯) −1/2 (27) η˘ := ¯η − ϵ¯· σ¯ †ϕ + 1 2 σ¯ †ϕ · γ¯σ¯ †ϕ + 1 2 tr[¯γ] ϵ˘ := (1 + αλγ¯) −1/2 (¯ϵ − γ¯σ¯ †ϕ) Proof.

Literature Context

The entropic pricing principle applied in this article is developed in [1, 2], where price is derived from a log-martingale condition in the measure with minimum entropy relative to economic expectations, and the risk aversion of the investor is the parameter of the entropic risk metric. The algebra of quadratic Gaussian variables, with its expectation and measure groups, is developed in [14] for the pricing of interest rate and cross asset derivatives in term structure models. p. 0

Where the paper says it

Model Risk Analysis for Entropic Hedging Strategies Paul McCloud McCloud Research (Dated: 5 October 2026) Entropic risk optimisation is a general framework for pricing and hedging financial derivatives in incomplete markets that can be used to decompose P&L into market and model risk contributions. When the prices are quadratic Gaussian, the coupled equations for price and hedge ratios are solved in closed form.

The Ornstein-Uhlenbeck process that generates the factors on each hedging interval originates in the theory of Brownian motion of Uhlenbeck and Ornstein [15], and underlies the term structure model of Vasicek [16]. p. 0
Where the paper says it

The equations of the quadratic Gaussian model are directly applicable when the underlying prices follow Gaussian processes, such as fractional Ornstein-Uhlenbeck processes. The model also provides simple parametric expressions for hedge ratios that can be used for regressions in deep hedging.

The paper also references literature on deep hedging [17, 18] for which the quadratic Gaussian model provides parametric expressions for the hedge ratios. p. 0
Where the paper says it

The equations of the quadratic Gaussian model are directly applicable when the underlying prices follow Gaussian processes, such as fractional Ornstein-Uhlenbeck processes. The model also provides simple parametric expressions for hedge ratios that can be used for regressions in deep hedging.

Case Study Illustration

A case study is provided where an underlying market comprises a bond and a swap, demonstrating how market incompleteness via convexity, dimension and funding contributions reduces the effectiveness of the replication strategy. The optimal hedging in the mid model (αλ = 0) is identified by varying the hedge ratio across the range −1 ≤ δ ≤ 3, with the optimal choice at δ∗ = 1 highlighted. p. 8

Where the paper says it

P&L is a combination of normal and noncentral chi-squared variables. Optimal hedging in the mid model (αλ = 0) is identified by varying the hedge ratio across the range −1 ≤ δ ≤ 3, with the optimal choice at δ ∗ = 1 highlighted.

Introducing market incompleteness via convexity, dimension and funding contributions reduces the effectiveness of the replication strategy. p. 8
Where the paper says it

Optimal hedging in the mid model (αλ = 0) is identified by varying the hedge ratio across the range −1 ≤ δ ≤ 3, with the optimal choice at δ ∗ = 1 highlighted. Introducing market incompleteness via convexity, dimension and funding contributions reduces the effectiveness of the replication strategy.

The quadratic coefficient drives the convexity penalty, and dialling these parameters away from the baseline generates the three penalties and their contributions to pricing and hedging. p. 9
Where the paper says it

The bottom-right graph demonstrates that the quadratic coefficient drives the convexity penalty. (no funding penalty), and dialling these parameters away from the baseline generates the three penalties and their contributions to pricing and hedging.

Distribution Analysis

The impact of the quadratic coefficient γ on the indifference price and hedge ratio is illustrated in Figure 4, showing results for notionals in the range −1 ≤ αλ ≤ 1, highlighting the mid case αλ = 0. The bottom-left graph shows the P&L distribution as γ is varied in the range −10 ≤ γ ≤ 10, with the γ = 0 case highlighted. p. 9

Where the paper says it

In the top-right graph, long positions correspond to the decreasing curves and short positions correspond to the increasing curves. The bottom-left graph shows the P&L distribution as γ is varied in the range −10 ≤ γ ≤ 10, with the γ = 0 case highlighted.

The bottom-right graph demonstrates that the quadratic coefficient drives the convexity penalty (no funding penalty), and dialling these parameters away from the baseline generates the three penalties and their contributions to pricing and hedging. p. 9
Where the paper says it

The bottom-right graph demonstrates that the quadratic coefficient drives the convexity penalty. (no funding penalty), and dialling these parameters away from the baseline generates the three penalties and their contributions to pricing and hedging.

The noise coefficient ρ drives the dimension penalty. p. 9
Where the paper says it

The bottom-left graph shows the P&L distribution as γ is varied in the range −10 ≤ γ ≤ 10, with the γ = 0 case highlighted. The bottom-right graph demonstrates that the quadratic coefficient drives the convexity penalty.

The bond volatility drives the funding penalty. p. 9
Where the paper says it

The bottom-right graph demonstrates that the quadratic coefficient drives the convexity penalty. (no funding penalty), and dialling these parameters away from the baseline generates the three penalties and their contributions to pricing and hedging.

Final Price Structure

The final price is thus a linear combination of two independent variables, one noncentral chi-squared with one degree of freedom, and one standard normal. The final price has entropy-adjusted mean: η + 1/2 γ(1 + ρ 2)σ 2q − p − 1/2 αλ(χγ + χσ + χs). p. 2

Where the paper says it

The underlying and derivative final prices are modelled as quadratic Gaussian variables: Q := ¯µ + ¯σZ (14) P := ¯η + ¯ϵ · Z + 1 2 Z · γZ¯ For a portfolio comprising notional λ of the derivative security and notionals ω of the underlying securities, the portfolio final price is the quadratic Gaussian variable: λP +ω·Q = (λη¯+ω·µ¯)+ (λϵ¯+ ¯σ †ω)·Z + 1 2 Z ·λγZ¯ (15) Using the quadratic Gaussian theorem, the entropyadjusted mean of the final price is: E[α][λP + ω · Q] = (16) 1 2α log det[1 + αλγ¯] + (λη¯ + ω · µ¯) − 1 2 α(λϵ¯+ ¯σ †ω) · (1 + αλγ¯) −1 (λϵ¯+ ¯σ †ω) which is valid for αλ within an open interval containing zero. Entropic risk optimisation seeks to maximise the entropy-adjusted mean, at the risk aversion α > 0 of the investor, for the final price of a self-funded portfolio whose initial price is zero.

The risk penalties for convexity, dimension, funding and strategy are given by expressions involving the parameters. p. 5
Where the paper says it

The P&L of the funded derivative is P − δ · Q. The P&L is assessed at initial time by the risk penalties: χ γ := tr αλγ¯ − log[1 + αλγ¯] (αλ) 2 (51) χ σ := ∥(1 − 1ν˘)˘ϵ∥ 2 χ s := s 2 δ s · q χ δ := ∥σ˘ † (δ − δ ∗ )∥ 2 for convexity, dimension, funding and strategy, and the entropy-adjusted mean of the P&L is: 1 λ (E[α][ω ∗ · Q + λ(P − δ · Q)] − E[α][ω ∗ · Q]) = (52) (˘η − (1 + r)p) − 1 2 αλ(χ γ + χ σ + χ s + χ δ ) The P&L is explained at final time by the attribution: Xγ := 1 2 (Y · γY¯ − tr[¯γ]) (53) Xσ := (¯ϵ − γ¯σ¯ †ϕ) · Y − δ σ · (Q − (1 + r)q) Xs := −sδs · (Q − (1 + r)q) Xδ := −(δ − δ ∗ ) · Q for convexity, dimension, funding and strategy, where Y := Z + ¯σ †ϕ are the measure-adjusted factors, and the total P&L is: P − δ · Q = (54) (˘η − (1 + r)p) + (Xγ + Xσ + Xs + Xδ ) The risk penalties are non-negative and are the incremental entropy-adjusted means generated by the attributions: − 1 2 αλχγ = 1 λ (E[α][ω ∗ · Q + λXγ ] (55) − E[α][ω ∗ · Q]) − 1 2 αλχσ = 1 λ (E[α][ω ∗ · Q + λ(Xγ + Xσ )] − E[α][ω ∗ · Q + λXγ ]) − 1 2 αλχs = 1 λ (E[α][ω ∗ · Q + λ(Xγ + Xσ + Xs )] − E[α][ω ∗ · Q + λ(Xγ + Xσ )]) − 1 2 αλχδ = 1 λ (E[α][ω ∗ · Q + λ(Xγ + Xσ + Xs + Xδ )] − E[α][ω ∗ · Q + λ(Xγ + Xσ + Xs )]) Proof.

The leading term in the entropy-adjusted mean is the adjusted mean of the derivative final price minus its initial price compounded at the funding rate. p. 6
Where the paper says it

Since the risk penalties are scaled by λ 2 , they are penalties for both long and short positions in the derivative security. The leading term in the entropy-adjusted mean is the adjusted mean of the derivative final price minus its initial price compounded at the funding rate.

Indifference Price Determination

The indifference pricing condition is quadratic in the funding spread, which resolves both the funding spread and the derivative price at the start of the interval. The derivative initial price is obtained from this condition written in terms of risk penalties. p. 7

Where the paper says it

The positive branch is finite as αλ → 0, whereas the funding spread diverges on the negative branch. The derivative initial price is obtained from the indifference pricing condition written in terms of the risk penalties, utilising the funding spread as determined.

The total funding 1 + r + αλs combines the funding rate r of the underlying market with the funding spread s charged for the hedge funding gap, and indifference pricing fixes its square. p. 7
Where the paper says it

The derivative initial price is obtained from the indifference pricing condition written in terms of the risk penalties, utilising the funding spread as determined. The total funding 1 + r + αλs combines the funding rate r of the underlying market with the funding spread s charged for the hedge funding gap, and indifference pricing fixes its square.

The solution is real precisely when the argument of the square root is non-negative. p. 7
Where the paper says it

The total funding 1 + r + αλs combines the funding rate r of the underlying market with the funding spread s charged for the hedge funding gap, and indifference pricing fixes its square. The solution is real precisely when the argument of the square root is non-negative; at that boundary the two branches merge where the total funding vanishes, and beyond it the derivative security at this notional has no indifference price.

Multiple Intervals

The problem is framed in a format where results from the previous section can be used to solve for the hedge portfolio and derivative price at the start of each interval. The maximum entropy-adjusted mean attained by this portfolio is given by an expression involving the adjusted parameters for each interval. p. 2

Where the paper says it

This enables the definitions: σ¯ −1 := ¯σ † ν¯ −1 σ¯ −† := ¯ν −1σ¯ (19) 1ν¯ := ¯σ † ν¯ −1σ¯ of the right-inverse σ¯ −1 , left-inverse σ¯ −† and projection 1ν¯, satisfying (¯σ −1 ) † = ¯σ −† and: σ¯σ¯ −1 = 1 = ¯σ −†σ¯ † (20) σ¯ −1σ¯ = 1ν¯ = ¯σ †σ¯ −† The unique self-funded underlying portfolio ω ∗ generating the maximum entropy-adjusted mean is obtained as the solution of the self-funding and stationary conditions: ϕ · q = 0 (21) µ¯ − (1 + r)q − νϕ¯ = 0 where r is the funding rate used to implement the selffunding condition and ϕ := αω∗ . These conditions have solution: r := q · ν¯ −1 (¯µ − q) q · ν¯−1q (22) ϕ := ¯ν −1 (¯µ − (1 + r)q) The maximum entropy-adjusted mean attained by this portfolio is: E[α][ω ∗ · Q] = 1 2α ∥σ¯ †ϕ∥ 2 (23) which is non-negative, and vanishes if and only if ϕ = 0, in which case µ¯ = (1 + r)q.

The increment in the entropy-adjusted mean is set to zero in indifference pricing, which resolves both the funding spread and the derivative price at the start of each interval. p. 7
Where the paper says it

Indifference pricing imposes the additional constraint that these two entropy-adjusted means align. The derivative initial price p and the funding spread s, related by the expression p = δ σ · q + sδs · q, are both resolved by this condition.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Precision PnL Decomposition for Trading Systems This system can decompose trading Profit and Loss into market risk, model risk contributions (convexity, dimension, funding), and strategy contributions based on realized PnL attribution (Theorem II.5).

  2. Optimal Hedging Strategy Determination This AI can determine a unique hedge portfolio that maximises the entropy-adjusted mean for various portfolios by solving self-funding and stationary conditions (Theorem II.1, II.2, II.3).

  3. Indifference Pricing for Robust Valuation This system can calculate an indifference derivative initial price by aligning the maximum entropy-adjusted means of underlying and derivative markets (Theorem II.6), providing a more robust valuation than standard no-arbitrage principles alone when covariance is singular or positive final prices exist.

  4. Dynamic Hedging in Multiple Time Intervals This system can generate iterative hedging schedules using the Ornstein-Uhlenbeck process to model price movements over multiple hedging intervals, allowing for dynamic adjustment based on evolving market conditions (Section III).

  5. Risk Penalty Quantification and Control This system can quantify residual risks as non-negative risk penalties (convexity, dimension, funding) which are the incremental entropy-adjusted means generated by the attributions, enabling investors to control or attribute these specific sources of model incompleteness.

Abstract

Market incompleteness causes residual profit and loss after hedging that is accounted for with valuation adjustments and bid-offer spreads acting as model risk buffers for the derivative price. Entropic risk optimisation is a general framework for modelling these adjustments. When the prices are quadratic Gaussian, the coupled equations for price and hedge ratios are solved exactly, with a decomposition of the entropic risk into convexity, dimensionality, funding and strategy contributions that are attributed in the explanation of realised P&L. The results are applicable when the underlying prices follow Gaussian processes, such as fractional Ornstein-Uhlenbeck processes. The article concludes with analysis of the continuous hedging limit for these processes, and a discussion of their application in data-driven regressions.

Sources

Related papers