Study Finds Generalized Highway Networks Excel in Option Pricing Machine Learning

Global: Study Finds Generalized Highway Networks Excel in Option Pricing Machine Learning

In July 2023, a research team released a preprint on arXiv that investigates how neural network architecture influences the accuracy and training efficiency of supervised learning models for option pricing and implied volatility estimation. The authors aim to determine whether newer architectures can outperform the conventional feed‑forward networks commonly used in financial modeling.

Problem Context

The paper addresses the longstanding challenge of approximating option prices and implied volatilities from model parameters such as those in the Black‑Scholes and Heston frameworks. Accurate and fast approximations are essential for traders and risk managers who require real‑time valuations.

Methodological Approach

To evaluate architectural impact, the researchers trained several neural network designs—including plain feed‑forward networks, generalized highway networks, and a simplified Deep Galerkin Method (DGM) variant—under comparable parameter budgets. Training performance was measured using mean squared error (MSE) and computational time.

Findings for Classic Pricing Models

According to the authors, the generalized highway network architecture achieved the lowest MSE while also reducing training time relative to the other models when applied to both the Black‑Scholes and Heston option pricing problems.

Results for Implied Volatility Estimation

For the transformed implied volatility task, the study reports that the simplified DGM variant produced the smallest error among the tested architectures, suggesting a particular suitability for this type of inverse problem.

Capacity‑Normalised Evaluation

The authors also performed a capacity‑normalised comparison, ensuring each architecture operated with an equal number of trainable parameters. This analysis confirmed the earlier observations, with the generalized highway network and DGM variant retaining their performance advantages.

Real‑World Market Data Experiments

Extending beyond synthetic datasets, the researchers incorporated real market data into the implied volatility experiments. The results mirrored the abstract‑based findings, reinforcing the practical relevance of the identified architectures.

Implications and Future Work

The authors suggest that adopting generalized highway networks or DGM‑style models could enhance the efficiency of machine‑learning‑driven pricing tools in finance. They recommend further exploration of these architectures across broader asset classes and market conditions.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.