Component-Based Methods in Multivariate and Hierarchical Time Series Forecasting

With the rapid advancement of technology and data collection mechanisms, forecasting large collections of time series data has become a common objective in modern data analysis scenarios. This thesis develops three interconnected methods designed to reduce variations in forecast errors and improve accuracy by extracting and utilising shared information across multiple time series, supported by theoretical guarantees. Their effectiveness is demonstrated in applications of forecasting Australian tourism, Wikipedia pageviews, and macroeconomic indicators. This has immediate implications for forecasting in all disciplines including macroeconomics and finance.