Abstract
dc:descriptionReal-life time series dynamically evolve through time, exhibiting chaotic patterns. They are governed both by non-linear dynamics and uncertainty. Capturing accurately their evolution through time based only on historical data is a very challenging task. This Ph.D. thesis deals with efficient dynamic online models applied for stock price, flight price, hotel rating, and COVID-19 cases prediction. In the former 3 cases, a time-evolving structure was constructed by employing dyads, i.e., pairs of latent state vectors, which model the time series evolution. Firstly, a dynamic dyadic particle filter was proposed for financial time series prediction (i.e., stock and flight prices), that is based on sequential importance resampling. This filter incorporates the tracking abilities of Kalman filtering. The variance of the observation noise has been treated as a random variable obeying a heavy-tailed distribution to model the uncertainty governing the price time series. However, financial time series exhibit erratic behavior since they are governed both by diffusion and jumps. To address this problem, a jump-diffusion particle filter was proposed for financial time series prediction. A particle coefficient was introduced which encodes these phenomena. The diffusion process has been assumed to be a geometric Brownian motion. The negative log-likelihood of the posterior distributions has been approximated by a Taylor expansion around the previously observed drift parameter. Novel approximation of the first and second-order derivatives of the negative log-likelihood of the posterior distribution with respect to the previously observed drift parameter have been derived. Sudden price jumps have been modeled using a reversible jump Markov chain Monte Carlo framework. Furthermore, an online dynamic weighted ordered probit model integrated with a collaborative Kalman filtering scheme was proposed for online hotel rating prediction. A dynamically evolving optimal standard deviation, modeling the hotel reputation evolution, has been learnt by the model through the application of a Newton-Raphson type update. An accept-reject sampling scheme has been introduced to compute the expected value of an auxiliary truncated normal random variable utilized to predict hotel ratings based on past ratings, which filled in a short sliding window. This process confronted efficiently both data sparsity and scalability problems. Last, a hybrid model comprised by an autoregressive filter, a graph convolutional neural network, and a long-short term memory neural network was proposed for accurate prediction of COVID-19 cases in U.S. Linearities and non-linearities governing the time series of cases have been capture effectively by the hybrid model. A graph convolutional neural network has exploited a novel adjacency matrix which relies on Granger causality tests applied to historical COVID-19 cases for each State in U.S., capturing efficiently latent information about the spread of the virus, while revealing which State actually affects the other ones. Consequently, the overall prediction performance of the hybrid model was strongly enhanced.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Ntemi, Myrsini
- Ντέμη, Μυρσίνη
Subjects
dc:subject × 14- Πρόβλεψη χρονοσειρών
- Συνεργατικά φίλτρα Kalman
- Σωματιδιακά Φίλτρα
- Συνελικτικά Νευρωνικά Δίκτυα Γραφημάτων
- Timeseries prediction
- Collaborative Kalman filters
- Particle filters
- Graph Convolutional Neural Networks
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/52030