Latent SDEs have recently emerged as a powerful probabilistic framework to model time series and stochastic systems in continuous time. However, training procedures require to estimate from the data both the parameters of the model and the latent state of the system. This thesis develops Light SDE Matching as a novel solution to this problem, bridging the classical theory of filtering and smoothing with the recent approach of SDE Matching based on variational inference. An alternative algorithm is obtained by replacing the original variational posterior with a structured smoother approximation. The approach is validated on a set of benchmark problems, including nonlinear and memory dependent systems. Finally, the thesis explores applications in finance, illustrating how latent SDEs can be used as a general framework for arbitrage free, data driven pricing models.
Le Latent SDEs sono recentemente emerse come un potente framework probabilistico per modellizzare serie storiche e sistemi stocastici in tempo continuo. Tuttavia, la fase di training richiede di stimare a partire dai dati sia i parametri del modello, sia lo stato latente del sistema. Questa tesi sviluppa Light SDE Matching come una nuova soluzione a questo problema, collegando la teoria classica del filtering e dello smoothing con il recente approccio di SDE Matching basato sul metodo dell’inferenza variazionale. Un algoritmo alternativo viene ottenuto sostituendo il posterior variazionale originale con l’approssimazione strutturata di uno smoother. L’approccio è validato su un insieme di problemi di riferimento, inclusi sistemi non lineari e sistemi con dipendenza dalla memoria. Infine, la tesi esplora alcune applicazioni in ambito finanziario, illustrando come le SDE latenti possano essere utilizzate come un framework generale per modelli di pricing privi di arbitraggio e data-driven.
Light SDE matching: a data-driven framework for learning stochastic differential equations with applications in finance
Meschieri, Andrea
2024/2025
Abstract
Latent SDEs have recently emerged as a powerful probabilistic framework to model time series and stochastic systems in continuous time. However, training procedures require to estimate from the data both the parameters of the model and the latent state of the system. This thesis develops Light SDE Matching as a novel solution to this problem, bridging the classical theory of filtering and smoothing with the recent approach of SDE Matching based on variational inference. An alternative algorithm is obtained by replacing the original variational posterior with a structured smoother approximation. The approach is validated on a set of benchmark problems, including nonlinear and memory dependent systems. Finally, the thesis explores applications in finance, illustrating how latent SDEs can be used as a general framework for arbitrage free, data driven pricing models.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi Andrea Meschieri Light SDE Matching.pdf
accessibile in internet per tutti
Descrizione: Testo principale tesi
Dimensione
5.48 MB
Formato
Adobe PDF
|
5.48 MB | Adobe PDF | Visualizza/Apri |
|
Executive Summary Light SDE Matching Andrea Meschieri.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
1.44 MB
Formato
Adobe PDF
|
1.44 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/10589/247587