This thesis explores the application of neural networks to option pricing and model calibration, focusing on the Black-Scholes and Variance Gamma models. Traditional numerical methods, such as the Carr-Madan approach and iterative calibration algorithms, can be computationally intensive, limiting their scalability in real-world financial applications. To address this, we analyze the performance of three deep learning architectures—Multilayer Perceptrons (MLP), Residual Networks (ResNet), and Highway Networks—in approximating option prices and calibrating model parameters. Our study evaluates the ability of these architectures to efficiently learn the pricing function and inverse calibration mapping, significantly reducing computational costs while maintaining accuracy. Using a large dataset generated through Latin Hypercube Sampling, we train and test the models, assessing their performance with multiple error metrics. The results indicate that advanced architectures, such as ResNet and Highway Networks, improve prediction accuracy and generalization compared to traditional MLPs, offering a viable alternative to classical numerical methods. This work contributes to the growing field of machine learning in quantitative finance, demonstrating the potential of neural networks in accelerating financial computations.
Questa tesi esplora l'applicazione delle reti neurali al pricing delle opzioni e alla calibrazione dei modelli, con un focus sui modelli di Black-Scholes e Variance Gamma. I metodi numerici tradizionali, come l’approccio di Carr-Madan e gli algoritmi iterativi di calibrazione, possono essere computazionalmente onerosi, limitandone la scalabilità nelle applicazioni finanziarie reali. Per affrontare questo problema, analizziamo le prestazioni di tre diverse architetture di deep learning—Multilayer Perceptron (MLP), Residual Networks (ResNet) e Highway Networks—nell'approssimare i prezzi delle opzioni e nel calibrare i parametri del modello. Il nostro studio valuta la capacità di queste architetture di apprendere in modo efficiente la funzione di pricing e la mappatura inversa della calibrazione, riducendo significativamente i costi computazionali senza compromettere l'accuratezza. Utilizzando un ampio dataset generato tramite Latin Hypercube Sampling, addestriamo e testiamo i modelli, valutandone le prestazioni con diverse metriche di errore. I risultati indicano che architetture avanzate, come ResNet e Highway Networks, migliorano la precisione e la capacità di generalizzazione rispetto agli MLP tradizionali, offrendo un'alternativa valida ai metodi numerici classici. Questo lavoro contribuisce al crescente campo del machine learning nella finanza quantitativa, dimostrando il potenziale delle reti neurali nell'accelerare i calcoli finanziari.
Neural networks for option pricing and calibration: a comparative study of architectures
Raeli, Fabrizio
2023/2024
Abstract
This thesis explores the application of neural networks to option pricing and model calibration, focusing on the Black-Scholes and Variance Gamma models. Traditional numerical methods, such as the Carr-Madan approach and iterative calibration algorithms, can be computationally intensive, limiting their scalability in real-world financial applications. To address this, we analyze the performance of three deep learning architectures—Multilayer Perceptrons (MLP), Residual Networks (ResNet), and Highway Networks—in approximating option prices and calibrating model parameters. Our study evaluates the ability of these architectures to efficiently learn the pricing function and inverse calibration mapping, significantly reducing computational costs while maintaining accuracy. Using a large dataset generated through Latin Hypercube Sampling, we train and test the models, assessing their performance with multiple error metrics. The results indicate that advanced architectures, such as ResNet and Highway Networks, improve prediction accuracy and generalization compared to traditional MLPs, offering a viable alternative to classical numerical methods. This work contributes to the growing field of machine learning in quantitative finance, demonstrating the potential of neural networks in accelerating financial computations.File | Dimensione | Formato | |
---|---|---|---|
2025_04_Raeli_Tesi.pdf
accessibile in internet per tutti
Descrizione: file tesi
Dimensione
1.41 MB
Formato
Adobe PDF
|
1.41 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/236089