Time series prediction model for multiple applications

• Main Author: Yu, Ying Cheng
• Other Authors: Vidya Sudarshan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/166682
• Institution: Nanyang Technological University

In recent times, the field of time series forecasting has gained much attention and popularity due to the effectiveness of statistical and Artificial Intelligence (AI) models in improving decision making. The focus of this paper will be on Neural Network models, and the plausibility of developing a model that can be applied to different kinds of time series datasets with minimal hyperparameter tuning. While the development of fronter technologies like AI and Machine Learning (ML) have been rapid over the years, hyperparameter is an essential and costly step in the process of building and optimizing ML models. These parameters can significantly affect the performance of models and tuning them is often a computationally expensive procedure. It is worthwhile to investigate how effective a hybrid model that is tuned on a particular type of time series dataset can be on other types of time series. This paper will explore the model development process for a hybrid model and its performance against 3 different baseline NN models (CNN, LSTM, GRU). The NN models are applied on 3 selected time series datasets that have been sourced online: COVID-19 pandemic data for countries, stock prices in NASDAQ, NYSE, AMEX stock exchanges, as well as temperature data for cities. The type of time series forecasting investigated in this project will be single-step, and various lookback period lengths will be explored for this supervised learning task. The goal of this research is to demonstrate and examine the effectiveness of a hybrid NN model that has been trained on 1 dataset, on other datasets.