Forecasting Mixed-Frequency Time Series with Statistical Learning

Mixed-frequency time series, defined as data recorded at different time intervals such as quarterly, monthly or daily, are prevalent across various fields, particularly macroeconomics. For example, the gross domestic product (GDP) growth is a key economic indicator, but it is published quarterly and with a long delay, while higher-frequency data that impact GDP growth e.g., unemployment and inflation are released monthly and reflect real-time economic conditions. Unlike standard econometric models, mixed-frequency models can model and forecast variables recorded at different frequencies. However, when dealing with large datasets, the dimensionality of these models quickly increases, requiring alternative estimation techniques. This thesis develops statistical learning methods for forecasting high-dimensional mixed-frequency time series. It introduces new approaches and extends existing methods to enhance the estimation and predictive accuracy of mixed-frequency models, providing valuable tools for policymakers and practitioners to make timely and informed predictions. The methods are empirically evaluated on classical macroeconomic datasets as well as on more modern platform streaming datasets.