In this work we address the calibration of stochastic and rough volatility models via neural networks. We first present the three considered models: Heston, rough Heston and rough Bergomi. The first allows for an analytical representation of the characteristic function, the second for a semi-analytical, and the third only relies on Monte Carlo simulations. Then we compare the most important neural networks approaches present in the literature to calibrate these models: the two-steps approach and the direct approach. In the two-steps approach we first train the network to learn the pricing functional and then employ it in the second stage: the calibration. Conversely, in the direct approach we train the networks directly for the calibration purpose. We start from the image based approach, where the network learns how to produce an entire volatility surface on a fixed grid of options parameters. Then we describe the consequent pointwise approach, where the network learns how to price a single option and even explore the new pointwise random grids approach, where the options parameters are generated randomly. After that, we adopt the direct approach, reproducing it with both Feed-Forward Neural Networks (FFNNs) and Convolutional Neural Networks (CNNs). We motivate the use of CNNs and extend the interpretability analysis of neural networks in this context by introducing saliency maps for CNN-based calibration models. Our original contribution also relies on the adaptation of the two-steps pointwise random grids approach to the direct calibration setting, implementing it with both FFNN and CNN. We systematically compare all the approaches calibration error and computational time, underlying the trade-off between speed and precision. Moreover, bring to light interpretability issues for rough volatility models and suggest an intermediate approach: the smile-FFNN, a compromise between the image based and the pointwise approach. Lastly we discuss the bottlenecks to lift in order to solve the calibration paradigm.
In questo lavoro affrontiamo la calibrazione di modelli di volatilità stocastica e ruvida tramite reti neurali. Presentiamo innanzitutto i tre modelli considerati: Heston, rough Heston e rough Bergomi. Il primo ammette una rappresentazione analitica della funzione caratteristica, il secondo una semi-analitica e il terzo si basa esclusivamente su simulazioni Monte Carlo. Successivamente confrontiamo i principali approcci basati su reti neurali presenti in letteratura per la calibrazione di tali modelli: l'approccio a due stadi e l’approccio diretto. Nell'approccio a due stadi la rete viene prima addestrata ad apprendere la funzione di pricing e successivamente utilizzata nella seconda fase: quella di calibrazione. Al contrario, nell’approccio diretto le reti vengono addestrate direttamente ad effettuare la calibrazione. Partiamo dall’approccio basato su immagini, in cui la rete impara a generare un’intera superficie di volatilità su una griglia fissa di parametri dell'opzione. Poi descriviamo il conseguente approccio pointwise, in cui la rete impara a prezzare una singola opzione ed esploriamo anche il nuovo approccio pointwise con griglie randomiche, in cui i parametri delle opzioni vengono generati in modo casuale. Adottiamo infine l’approccio diretto, riproducendolo sia con reti Feed-Forward (FFNN), sia con reti Convoluzionali (CNN). Motiviamo l’utilizzo delle CNN ed estendiamo l’analisi di interpretabilità delle reti neurali introducendo le saliency maps per i modelli di calibrazione basati su CNN. Il nostro contributo originale include anche l’adattamento dell’approccio a due stadi pointwise a griglie randomiche al contesto della calibrazione diretta, implementandolo sia con FFNN sia con CNN. Confrontiamo sistematicamente tutti gli approcci in termini di errori di calibrazione e tempo computazionale, mettendo in evidenza il compromesso tra velocità e precisione. Mettiamo in luce problematiche di interpretabilità nei modelli di volatilità ruvida e proponiamo un approccio intermedio, la smile-FFNN, come compromesso tra l’approccio basato su immagini e quello pointwise. Infine discutiamo i principali colli di bottiglia da sollevare per risolvere il paradigma della calibrazione.
An overview of Neural Networks calibration of stochastic and rough volatility models
Fioravanti, Mattia
2024/2025
Abstract
In this work we address the calibration of stochastic and rough volatility models via neural networks. We first present the three considered models: Heston, rough Heston and rough Bergomi. The first allows for an analytical representation of the characteristic function, the second for a semi-analytical, and the third only relies on Monte Carlo simulations. Then we compare the most important neural networks approaches present in the literature to calibrate these models: the two-steps approach and the direct approach. In the two-steps approach we first train the network to learn the pricing functional and then employ it in the second stage: the calibration. Conversely, in the direct approach we train the networks directly for the calibration purpose. We start from the image based approach, where the network learns how to produce an entire volatility surface on a fixed grid of options parameters. Then we describe the consequent pointwise approach, where the network learns how to price a single option and even explore the new pointwise random grids approach, where the options parameters are generated randomly. After that, we adopt the direct approach, reproducing it with both Feed-Forward Neural Networks (FFNNs) and Convolutional Neural Networks (CNNs). We motivate the use of CNNs and extend the interpretability analysis of neural networks in this context by introducing saliency maps for CNN-based calibration models. Our original contribution also relies on the adaptation of the two-steps pointwise random grids approach to the direct calibration setting, implementing it with both FFNN and CNN. We systematically compare all the approaches calibration error and computational time, underlying the trade-off between speed and precision. Moreover, bring to light interpretability issues for rough volatility models and suggest an intermediate approach: the smile-FFNN, a compromise between the image based and the pointwise approach. Lastly we discuss the bottlenecks to lift in order to solve the calibration paradigm.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Fioravanti.pdf
accessibile in internet per tutti
Descrizione: Tesi
Dimensione
2.49 MB
Formato
Adobe PDF
|
2.49 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Fioravanti_executive_summary.pdf
accessibile in internet per tutti
Descrizione: Executive summary
Dimensione
581.37 kB
Formato
Adobe PDF
|
581.37 kB | 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/10589/246221