Essays on econometric forecasting

Abstract

This thesis contributes towards the improvement of model-based econometric forecast performance under realistic forecast environments, such as when information about in-sample structural breaks is unknown or when the forecast users' loss function is not based on squared-errors. The thesis consists of three essays within this theme. The first essay focuses on the scenario when there might be structural breaks in the provided sample period and break information is unknown to forecasters. Under structural break uncertainty the forecasting performance of a specific model depends on the choice of the estimation window. Aiming to improve the forecasting ability of econometric models, we propose two new weighting methods for averaging forecasts generated using different estimation windows. The first method builds on the intuition that the combination weights reflect the probability of each time point being the most-recent break point, and we use the reversed ordered Cusum test statistics to capture this intuition. The second weighting method imposes heavier weights on the forecasts that use more recent information. Both the Monte Carlo study and the empirical application that uses the NAIRU Phillips curve to forecast U.S. inflation show that our proposed weighting methods improve forecasts and often outperform other existing strategies that deal with structural break uncertainty. The second essay provides a better understanding of the role of decision making in forecast evaluation. We assume that an investor allocates her assets between a developed stock market, the domestic emerging stock market, and the domestic bill market, based on her excess stock return forecasts in two markets. Given a recursive forecasting strategy and a portfolio-switching trading strategy, we calculate the dollar-measured profits at the end of the investment period as the economic benefit of stock return forecasts. We compare forecast evaluation results implied by conventional statistics with the economic benefits arising from stock return forecasts and find inconsistency between them. This essay emphasizes the necessity of incorporating decision making into forecast evaluation. In the third essay, we examine how to generate forecasts when forecast results are evaluated using an asymmetric loss function that implies different forecasts costs resulting from over-prediction and under-prediction. In empirical applications that forecast the monthly changes of Australian unemployment rates and US industrial production, we incorporate the asymmetric quadratic loss function into model selection, model estimation and forecast combination. In particular, to select forecasting models based on the asymmetric loss function, we suggest the use of cross-validation. Our results show that the incorporation of the same loss function used in forecast evaluation when producing forecasts can improve forecast performance. However, model selection based on asymmetric loss has no obvious advantage over conventional model selection methods based on squared-errors. Show more