In decision-making problems, it is essential to quantify the uncertainty of predictive models. Conformal Prediction provides a theoretically sound solution, but requires data to be exchangeable, excluding time series data. Recent research has addressed this limitation, but this work argues that Adaptive Conformal Inference (ACI) (ACI, Gibbs & Candès, 2021), developed for time series with distribution shifts, is an effective method also for data with more general dependencies. It's theoretically analyzed the effect of the learning rate on ACI’s efficiency in both exchangeable and autoregressive cases. It's proposed a new parameter-free method, AgACI, which extends ACI by leveraging adaptive aggregation based on online experts. For time series forecasting, we compare ACI with alternative methods through rigorous simulations. Finally, it's conducted a real-world case study on electricity price forecasting, demonstrating that the proposed algorithm provides efficient predictive intervals for day-ahead forecasting.
Nei problemi di decision-making è essenziale quantificare l’incertezza dei modelli predittivi. La Conformal Prediction rappresenta una soluzione teoricamente solida, ma richiede che i dati siano scambiabili, escludendo così le serie storiche temporali. Recenti ricerche hanno affrontato questo limite, ma in questo elaborato si sostiene che l’Adaptive Conformal Inference (ACI) (ACI, Gibbs & Candès, 2021), sviluppata per serie storiche temporali con variazioni di distribuzione, sia un metodo efficace anche per dati con dipendenze più generali. Si analizza teoricamente l’effetto del tasso di apprendimento sull’efficienza dell’ACI nei casi scambiabile e autoregressivo. Si propone un nuovo metodo senza parametri, AgACI, che estende l’ACI utilizzando un’aggregazione adattiva basata su ’online experts’. Per la previsione di serie temporali, confrontiamo l’ACI con metodi alternativi attraverso scrupolose simulazioni. Infine, si conduce uno studio di caso reale sulla previsione dei prezzi dell’energia elettrica, dimostrando che l’algoritmo in questione fornisce intervalli predittivi efficienti per la previsione del giorno dopo.
Conformal prediction per serie storiche: una rassegna su teoria, metodi e applicazioni.
MOLENA, SAMANTHA
2024/2025
Abstract
In decision-making problems, it is essential to quantify the uncertainty of predictive models. Conformal Prediction provides a theoretically sound solution, but requires data to be exchangeable, excluding time series data. Recent research has addressed this limitation, but this work argues that Adaptive Conformal Inference (ACI) (ACI, Gibbs & Candès, 2021), developed for time series with distribution shifts, is an effective method also for data with more general dependencies. It's theoretically analyzed the effect of the learning rate on ACI’s efficiency in both exchangeable and autoregressive cases. It's proposed a new parameter-free method, AgACI, which extends ACI by leveraging adaptive aggregation based on online experts. For time series forecasting, we compare ACI with alternative methods through rigorous simulations. Finally, it's conducted a real-world case study on electricity price forecasting, demonstrating that the proposed algorithm provides efficient predictive intervals for day-ahead forecasting.File | Dimensione | Formato | |
---|---|---|---|
Molena_Samantha.pdf
accesso aperto
Dimensione
2.32 MB
Formato
Adobe PDF
|
2.32 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/84139