ETD

• Artificial Neural Networks
• Deep Calibration
• Heston-Nandi GARCH
• Rough Bergomi Model
• Rough Volatility

The calibration of advanced derivative pricing models, such as rough volatility frameworks, represents a severe computational bottleneck that traditionally inhibits real-time application in financial markets. This thesis investigates the integration of Artificial Neural Networks (ANNs) as deterministic surrogate models to eliminate this latency. By adopting a Two-Step Pointwise architecture, the computationally intensive mathematical mechanics are isolated to an offline training phase, reducing online calibration to a highly efficient sequence of linear algebraic evaluations. The methodology is applied to the discrete-time Heston-Nandi GARCH(1,1) model, utilizing the Fourier-Cosine (COS) expansion method for synthetic data generation and demonstrating the critical necessity of global optimization algorithms, specifically Differential Evolution (DE), to overcome structural equifinality. Finally, the study addresses the non-Markovian rough Bergomi (rBergomi) model, leveraging the Turbocharged Monte Carlo method to generate training data. A comparative analysis of network architectures demonstrates that a 64-neuron model utilizing the Swish activation function outperformed the other tested configurations, establishing it as the optimal surrogate.
By replacing traditional numerical solvers with this optimized ANN-DE pipeline, the online calibration time for the rBergomi model is reduced from several hours to approximately 15 seconds. Ultimately, this thesis shows that the computational bottleneck of the calibration process can be entirely eliminated using Artificial Neural Networks.