Rough volatility models, in which the instantaneous variance is driven by a fractional Brownian motion with Hurst index H < 1 2 , have become a central paradigm in quanti- tative finance owing to their ability to reproduce the steep short-maturity behaviour of the at-the-money implied-volatility skew observed for major equity indices [16]. Among these, the rough Bergomi model of Bayer et al. [4] is parsimonious, fits equity smiles across maturities with few parameters, and admits a tractable exponential-Gaussian vari- ance structure. A significant practical obstacle is that the model lacks a closed-form characteristic function, so option prices must be obtained by Monte Carlo simulation. This makes calibration—which requires many pricing evaluations inside an optimisation loop—prohibitively slow, and shifts the bottleneck to the generation of training data for neural-network surrogates [2, 18]. Thisthesisaddressesthecalibrationproblemalongthreecomplementarydirections. First, we derive a small-time approximation of the transition density of the rescaled log-price under a rough Bergomi-type model. The derivation proceeds by computing the con- ditional Gaussian density of the driftless rescaled process given the Gaussian Volterra driver, and then applying Varadhan’s theorem to identify the rate function. The resulting rate function is shown to coincide with the one obtained by Forde and Zhang [14] via the contraction principle, providing an independent verification through a more direct probabilistic route. Second, we exploit this density formula to price European options for all maturities simultaneously from a single set of Monte Carlo paths of the Gaussian driver, dramatically reducing the cost of training-set generation. We further document, through careful numerical experiments, that a single global Monte Carlo simulation intro- duces systematic short-maturity distortions in the implied-volatility surface—particularly in the at-the-money skew—and demonstrate that the short-time approximation corrects these distortions without requiring additional time discretisation. Third, building on the efficient training set, we train a feedforward neural-network surrogate following the random-grid pointwise framework of Baschetti et al. [2], and evaluate three experimen- tal variants that differ in label quality and maturity-based capacity specialisation. The results show that explicitly treating the short-maturity regime—through improved labels and, when beneficial, dual-surrogate gating—yields a substantially more balanced cali- bration across maturities.
I modelli di volatilità rough, in cui la varianza istantanea è guidata da un moto browniano frazionario con indice di Hurst H < 1 2 , sono diventati un paradigma centrale nella finanza quantitativa grazie alla loro capacità di riprodurre il comportamento ripido dello skew della volatilità implicita at-the-money a breve scadenza osservato per i principali indici azionari [16]. Tra questi, il modello rough Bergomi di Bayer et al. [4] è parsimonioso, calibra i sorrisi azionari su tutte le scadenze con pochi parametri, e ammette una strut- tura varianza esponenziale-gaussiana trattabile sia teoricamente che tramite simulazione Monte Carlo. Un ostacolo pratico significativo è che il modello non dispone di una fun- zione caratteristica in forma chiusa, rendendo i prezzi delle opzioni accessibili solo tramite simulazione Monte Carlo. Ciò rende la calibrazione tradizionale—che richiede molte valu- tazioni del pricer all’interno di un ciclo di ottimizzazione—proibitivamente lenta, e sposta il collo di bottiglia alla generazione dei dati di addestramento per surrogati basati su reti neurali [2, 18]. Questa tesi affronta il problema della calibrazione lungo tre direzioni complementari. Il primo contributo è teorico: deriviamo un’approssimazione a tempo breve della densità di transizione del log-prezzo riscalato sotto una dinamica di tipo rough Bergomi, recu- perando la funzione di tasso di Forde and Zhang [14] tramite un percorso probabilistico indipendente basato sul teorema di Varadhan. Il secondo contributo è numerico: mostri- amo come questa formula della densità consenta di prezzare opzioni europee su tutte le scadenze simultaneamente a partire da un unico insieme di traiettorie Monte Carlo del driver gaussiano, e documentiamo con esperimenti numerici le distorsioni sistematiche a breve scadenza introdotte da una simulazione globale nella costruzione della superficie di volatilità implicita, dimostrando che l’approssimazione a tempo breve corregge tali distor- sioni senza richiedere ulteriore discretizzazione temporale. Il terzo contributo è applica- tivo: implementiamo e validiamo una pipeline di calibrazione neurale su griglie casuali seguendo il framework di Baschetti et al. [2], e mostriamo che trattare esplicitamente il regime a breve scadenza—tramite etichette di addestramento migliorate e specializzazione del modello per regime di maturità—produce una calibrazione sostanzialmente più uni- forme su tutte le scadenze
Short-time asymptotics and deep calibration of the rough Bergomi volatility model
Pellegrinotti Mari, Paolo;PIZZO, LUCA
2024/2025
Abstract
Rough volatility models, in which the instantaneous variance is driven by a fractional Brownian motion with Hurst index H < 1 2 , have become a central paradigm in quanti- tative finance owing to their ability to reproduce the steep short-maturity behaviour of the at-the-money implied-volatility skew observed for major equity indices [16]. Among these, the rough Bergomi model of Bayer et al. [4] is parsimonious, fits equity smiles across maturities with few parameters, and admits a tractable exponential-Gaussian vari- ance structure. A significant practical obstacle is that the model lacks a closed-form characteristic function, so option prices must be obtained by Monte Carlo simulation. This makes calibration—which requires many pricing evaluations inside an optimisation loop—prohibitively slow, and shifts the bottleneck to the generation of training data for neural-network surrogates [2, 18]. Thisthesisaddressesthecalibrationproblemalongthreecomplementarydirections. First, we derive a small-time approximation of the transition density of the rescaled log-price under a rough Bergomi-type model. The derivation proceeds by computing the con- ditional Gaussian density of the driftless rescaled process given the Gaussian Volterra driver, and then applying Varadhan’s theorem to identify the rate function. The resulting rate function is shown to coincide with the one obtained by Forde and Zhang [14] via the contraction principle, providing an independent verification through a more direct probabilistic route. Second, we exploit this density formula to price European options for all maturities simultaneously from a single set of Monte Carlo paths of the Gaussian driver, dramatically reducing the cost of training-set generation. We further document, through careful numerical experiments, that a single global Monte Carlo simulation intro- duces systematic short-maturity distortions in the implied-volatility surface—particularly in the at-the-money skew—and demonstrate that the short-time approximation corrects these distortions without requiring additional time discretisation. Third, building on the efficient training set, we train a feedforward neural-network surrogate following the random-grid pointwise framework of Baschetti et al. [2], and evaluate three experimen- tal variants that differ in label quality and maturity-based capacity specialisation. The results show that explicitly treating the short-maturity regime—through improved labels and, when beneficial, dual-surrogate gating—yields a substantially more balanced cali- bration across maturities.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Pellegrinotti_Pizzo_Tesi.pdf
accessibile in internet per tutti
Descrizione: Thesis
Dimensione
2.54 MB
Formato
Adobe PDF
|
2.54 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_Pellegrinotti_Pizzo_Executive_Summary.pdf
accessibile in internet per tutti
Descrizione: Executive_Summary
Dimensione
1.24 MB
Formato
Adobe PDF
|
1.24 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/10589/253277