Mathematics of Differential Machine Learning in Derivative Pricing and Hedging

Pedro Duarte Gomes

2024-05-02

Abstract:This article introduces the groundbreaking concept of the financial differential machine learning algorithm through a rigorous mathematical framework. Diverging from existing literature on financial machine learning, the work highlights the profound implications of theoretical assumptions within financial models on the construction of machine learning algorithms.

Mathematical Finance,Machine Learning,Computational Finance

What problem does this paper attempt to address?

This paper attempts to solve the key problems in the pricing and hedging of financial derivatives, especially how to reliably estimate the prices and risks of these financial instruments through machine - learning methods. Specifically, the main objectives of the paper include:

1. **Introducing the Differential Machine Learning algorithm**: The paper proposes a new Differential Machine Learning algorithm and introduces it through a strict mathematical framework. Different from the existing financial machine - learning literature, this paper emphasizes the profound influence of theoretical assumptions in financial models on the construction of machine - learning algorithms.

2. **Solving the unbiased estimation problem**: An important challenge is how to obtain unbiased differential label estimates from data sources. The paper solves this problem and proposes a method for unbiased estimation under the assumption of local integrability, instead of relying on the assumptions of almost everywhere differentiability and almost everywhere Lipschitz continuity. This enables the method to be applied to a wider range of financial products.

3. **Establishing a unified theoretical basis**: The paper provides a unified theoretical framework, which is not only comprehensively compared theoretically but also verified by experimental results. This theoretical basis provides solid support for the experimental results and proves the optimality of the Differential Machine Learning method in the current context.

4. **Bridging abstract concepts and practical applications**: Through strict mathematical derivations, the paper bridges the gap between abstract financial concepts and actual algorithm implementations. This enables researchers and practitioners to better understand and apply these complex financial models.

### Main content overview

- **Theoretical framework**: The paper first regards the pricing and hedging functions as elements in Hilbert space and extends them to time series containing cumulative information through the Hahn - Banach extension theorem.

- **Derivation of loss functions**: The paper elaborately derives the loss functions of two risk - neutral methods, namely the least - squares Monte Carlo method and the Differential Machine Learning method. In particular, the loss function of the Differential Machine Learning method is derived using the theory of generalized functions, relaxing the assumptions on the payoff function.

- **Numerical experiments**: To verify the effectiveness of the proposed Differential Machine Learning method, the paper conducts numerical experiments to simulate the pricing and hedging processes of European call options. The experimental results show the profit - and - loss distributions and relative hedging errors of the method on different paths.

### Formula summary

- **Expected loss function**:
\[
J(g):=\int_{\Omega}L(g(z),x)p(z,x)\,dz\,dx
\]

- **Minimizing empirical loss**:
\[
\min_{\theta\in\Theta}J(\theta)=\min_{\theta\in\Theta}\frac{1}{n}\sum_{i = 1}^nL(g_{\theta}(x_i),y_i)
\]

- **Risk - neutral valuation formula**:
\[
g_{0,t}(Z)=E_Q[X|F_t]
\]

- **Loss function of Differential Machine Learning**:
\[
\int_{\Omega}|X - g_{0,t}|^2\,dQ+\lambda\int_{\Omega}|\partial X-\partial g_{0,t}|^2\,dQ
\]

Through these theories and experiments, the paper provides a new and effective solution for the pricing and hedging of financial derivatives.