Abstract
Forward-looking analysis is valuable for policymakers as they need effective strategies to mitigate imminent risks and potential challenges. Panel data sets contain time series information over a number of cross-sectional units and are known to have superior predictive abilities in comparison to time series only models. This PhD thesis develops novel panel data methods to contribute to the advancement of short-term forecasting and nowcasting of macroeconomic and environmental variables. The two most important highlights of this thesis are the use of cross-sectional dependence in panel data forecasting and to allow for timely predictions and ‘nowcasts’.Although panel data models have been found to provide better predictions in many empirical scenarios, forecasting applications so far have not included cross-sectional dependence. On the other hand, cross-sectional dependence is well-recognised in large panels and has been explicitly modelled in previous causal studies. A substantial portion of this thesis is devoted to developing cross-sectional dependence in panel models suited to diverse empirical scenarios. The second important aspect of this work is to integrate the asynchronous release schedules of data within and across panel units into the panel models. Most of the thesis emphasises the pseudo-real-time predictions with efforts to estimate the model on the data that has been released at the time of predictions, thus trying to replicate the realistic circumstances of delayed data releases.
Linear, quantile and non-linear panel models are developed to predict a range of targets both in terms of their meaning and method of measurement. Linear models include panel mixed-frequency vector-autoregression and bridge equation set-ups which predict GDP growth, inflation and CO2 emissions. Panel quantile regressions and latent variable discrete choice models predict growth-at-risk and extreme episodes of cross-border capital flows, respectively. The datasets include both international cross-country panels as well as regional subnational panels. Depending on the nature of the model and the prediction targets, different precision criteria evaluate the accuracy of the models in out-of-sample settings. The generated predictions beat respective standard benchmarks in a more timely fashion.
| Date of Award | 1 Dec 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Jack Fosten (Supervisor) & Martin Weale (Supervisor) |