Essays on deep learning in option pricing

Heidarzadeh, A. (2026) Essays on deep learning in option pricing. PhD thesis, University of Reading. doi: 10.48683/1926.00129276

Abstract/Summary

Financial institutions must reprice large derivative portfolios quickly for risk management and regulatory X-Value Adjustments. Traditional numerical methods such as Monte Carlo and PDE solvers scale poorly in high dimensions, limiting their use in real-time settings. This thesis develops deep-learning models that provide efficient, accurate, and data-efficient pricing, with a focus on exotic options where computational challenges are most acute. First, I design and optimize a feed-forward deep neural network (DNN) framework using Black–Scholes vanilla options as a benchmark. Training over one hundred ReLU-based networks on simulated datasets shows that proper feature scaling and architectures aligned with the pricing formula’s structure accelerate convergence and improve accuracy. When extended to exotic options with closed-form Black–Scholes-based solutions, the optimized DNNs achieve interpolation-level accuracy while producing prices thousands of times faster than traditional methods, enabling near real-time risk workflows. Second, to address limited data availability for exotics, I develop a transfer-learning (TL) strategy that adapts pre-trained vanilla models. Across twelve TL configurations- including freezing, fine-tuning, and added layers- this approach achieves accuracy comparable to training from scratch while reducing training time by more than 50%. TL also allows deeper, more expressive networks to perform well on small datasets. Third, I propose a multi-task learning (MTL) framework that jointly learns vanilla and exotic pricing through a shared representation and task-specific outputs. Using a task indicator and option-specific inputs, MTL lowers exotic-option RMSE by up to 80% and halves training iterations in low-data regimes. Performance peaks when roughly 83% of training samples are vanilla, emphasizing the value of balanced auxiliary data. Together, optimized DNNs, TL, and MTL provide a coherent toolkit that reduces computational cost and data demands while preserving accuracy, extending the reach of machine-learning methods in option pricing.

Item Type	Thesis (PhD)
URI	https://centaur.reading.ac.uk/id/eprint/129276
Identification Number/DOI	10.48683/1926.00129276
Divisions	Henley Business School > Finance and Accounting
Date on Title Page	September 2025
Download/View statistics	View download statistics for this item