This thesis addresses four main challenges within the field of energy finance and is structured into three parts. In Part I, we focus on electricity demand forecasting using deep learning techniques, concentrating on two key challenges: (i) developing a simple, interpretable, and high-performing model for probabilistic forecasts; (ii) generating reliable predictions that do not underestimate predictive uncertainty (a phenomenon known in the literature as overconfidence). To tackle these challenges, firstly, we propose a new class of Recurrent Neural Networks called RNN(p), studying their properties and focusing on how to optimise their training by leveraging their elementary architecture. Secondly, we introduce a new class of loss functions with a calibratable parameter, designed to counter the phenomenon of overconfidence. Part II is dedicated to the modelling of energy markets and to the pricing of derivatives through Lévy-driven Ornstein-Uhlenbeck processes. These stochastic processes are capable of reproducing some key stylised facts observed in energy markets; however, their simulation poses significant computational challenges. In this part, we develop a new technique for simulating these processes on discrete time grids, which is useful for pricing derivatives (both plain vanilla and path-dependent) having energy price as the underlying asset. The proposed technique is shown to be at least an order of magnitude faster than existing methods and to be more general, allowing for complete control of the numerical error. In Part III, we analyse the European carbon market, focusing on a well-known anomaly between spot prices and futures prices of emission allowances. This anomaly, already documented in the literature, manifests as a spread between the implicit spot/futures rate and the risk-free interest rate. By employing econometric techniques, we link this spread to a credit mechanism, showing the inefficiency of the carbon market and suggesting regulatory interventions to address this issue and to enhance the credibility of the European emissions reduction system.
Questa tesi affronta quattro sfide principali nel settore della finanza energetica ed è organizzata in tre parti. Nella Parte I, ci occupiamo del forecasting della domanda elettrica attraverso l’impiego di tecniche di deep learning, concentrandoci su due sfide chiave: (i) sviluppare un modello semplice, interpretabile e performante per le previsioni probabilistiche; (ii) generare previsioni affidabili che non sottostimino l’incertezza predittiva (un fenomeno noto in letteratura come overconfidenza). Per affrontare queste sfide, in primo luogo, proponiamo una nuova classe di Reti Neurali Ricorrenti, chiamata RNN(p), studiandone le proprietà e focalizzandoci su come ottimizzarne il training, sfruttando al meglio la loro architettura elementare. In secondo luogo, introduciamo una nuova classe di funzioni di costo dotate di un parametro calibrabile, progettate per contrastare il fenomeno dell’overconfidenza. La Parte II è dedicata alla modellistica dei mercati energetici e al pricing dei prodotti derivati mediante i processi Lévy-driven Ornstein-Uhlenbeck. Questi processi stocastici sono in grado di riprodurre alcuni importanti fatti stilizzati osservati nei mercati energetici, ma la loro simulazione presenta notevoli criticità in termini computazionali. In questa parte, sviluppiamo una nuova tecnica per simulare questi processi su griglie temporali discrete, utile per il pricing di derivati (sia plain vanilla che path-dependent) aventi il prezzo dell’energia come sottostante. La tecnica proposta si rivela più veloce di almeno un ordine di grandezza rispetto a quelle esistenti e offre un maggiore livello di generalità, consentendo inoltre un controllo completo sull’errore numerico. Nella Parte III, analizziamo il mercato europeo del carbonio, concentrandoci su un’anomalia tra i prezzi spot e futures delle quote di emissione. Questa anomalia, già documentata in letteratura, consiste in uno spread tra il tasso implicito spot/futures e il tasso di interesse privo di rischio. Attraverso l’uso di tecniche econometriche, colleghiamo tale spread ad un meccanismo di credito, mostrando l’inefficienza del mercato del carbonio e suggerendo un intervento regolatorio per affrontare tale problema ed accrescere la credibilità del sistema europeo di riduzione delle emissioni.
Four essays in energy finance
Manzoni, Pietro
2024/2025
Abstract
This thesis addresses four main challenges within the field of energy finance and is structured into three parts. In Part I, we focus on electricity demand forecasting using deep learning techniques, concentrating on two key challenges: (i) developing a simple, interpretable, and high-performing model for probabilistic forecasts; (ii) generating reliable predictions that do not underestimate predictive uncertainty (a phenomenon known in the literature as overconfidence). To tackle these challenges, firstly, we propose a new class of Recurrent Neural Networks called RNN(p), studying their properties and focusing on how to optimise their training by leveraging their elementary architecture. Secondly, we introduce a new class of loss functions with a calibratable parameter, designed to counter the phenomenon of overconfidence. Part II is dedicated to the modelling of energy markets and to the pricing of derivatives through Lévy-driven Ornstein-Uhlenbeck processes. These stochastic processes are capable of reproducing some key stylised facts observed in energy markets; however, their simulation poses significant computational challenges. In this part, we develop a new technique for simulating these processes on discrete time grids, which is useful for pricing derivatives (both plain vanilla and path-dependent) having energy price as the underlying asset. The proposed technique is shown to be at least an order of magnitude faster than existing methods and to be more general, allowing for complete control of the numerical error. In Part III, we analyse the European carbon market, focusing on a well-known anomaly between spot prices and futures prices of emission allowances. This anomaly, already documented in the literature, manifests as a spread between the implicit spot/futures rate and the risk-free interest rate. By employing econometric techniques, we link this spread to a credit mechanism, showing the inefficiency of the carbon market and suggesting regulatory interventions to address this issue and to enhance the credibility of the European emissions reduction system.File | Dimensione | Formato | |
---|---|---|---|
2025_02_ManzoniP_Thesis.pdf
accessibile in internet per tutti a partire dal 20/01/2028
Descrizione: Thesis
Dimensione
3.73 MB
Formato
Adobe PDF
|
3.73 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/10589/232693