On the use of multiple models in macro-economic forecasting and decision-making

Chernis, Edward

Thesis

On the use of multiple models in macro-economic forecasting and decision-making

下载PDF文件

Creator

Chernis, Edward

Rights statement

Strathclyde Thesis Copyright

Awarding institution

University of Strathclyde

Date of award

2024

Thesis identifier

T17173

Person Identifier (Local)

202171131

Qualification Level

Doctoral (Postgraduate)

Qualification Name

Doctor of Philosophy (PhD)

Department, School or Faculty

Department of Economics

Abstract

This thesis studies prediction and decision-making with multiple models in a Bayesian context. It is common practice to use multiple models in a forecasting and decision-making environment. This is because the true model is unknown, if it even exists, and even the ‘best’ model can be hard to identify. Consequently, there is often large amounts of uncertainty around the choice of model and decision-makers resort to multiple models. Using multiple models in practice is an open area of research where this thesis will contribute over the course of three essays. The three problems I seek to address are how to combine large numbers of forecasts; how to explain why combination techniques place higher, or lower, weight on certain models; and finally, how to use multiple models in a monetary policy decision-making context. The first essay (Chapter 2) investigates model combination with large numbers of models and predictions. To this end I use Bayesian Predictive Synthesis (BPS) which is a flexible method of combining density predictions. The flexibility comes from the ability to choose an arbitrary synthesis function to combine predictions. Through careful choice of this synthesis function I show how to combine large numbers of predictions - which is a common occurrence in macroeconomics. Specifically, I consider shrinkage priors and factor modelling techniques which are common choices for high-dimensional problems in macroeconomics. Additionally, these techniques provide an interesting contrast between the sparse weights implied by shrinkage priors and dense weights of factor modeling techniques. I find that the sparse weights of shrinkage priors perform well across exercises. The second essay (Chapter 3) addresses a common issue which is that it can be difficult to understand the reason why models are chosen in a combination. This is of particular importance in decision-making contexts. As in Chapter 2 we develop a synthesis function to address this problem. Typically, synthesis functions are specified parametrically as a dynamic linear regression. Instead, we develop a nonparametric treatment of the synthesis function using regression trees. We are able to explain the combination weights since we introduce observable variables in the regressions trees’ splitting rule. We can then examine the tree splits to see which variables are most important for explaining the weights. We show the advantages of our tree-based approach in two macroeconomic forecasting applications. The first uses density forecasts for GDP growth from the euro area’s Survey of Professional Forecasters. The second combines density forecasts of US inflation produced by many regression models involving different predictors. Both applications demonstrate the benefits – in terms of improved forecast accuracy and interpretability – of modeling the synthesis function nonparametrically. The third essay (Chapter 4) goes beyond the previous chapters going from prediction to decision-making. We show to make optimal monetary policy decisions with multiple models. We use Bayesian predictive decision synthesis (BPDS) as a formal Bayesian decision theory-based approach to monetary policy decision-making. BPDS draws on recent developments in model combination and statistical decision theory that make it possible to combine models in a manner that incorporates decision goals, expectations, and outcomes. We develop a BPDS procedure for a case study of monetary policy decision-making with an inflation-targeting central bank. Our procedure searches for an optimal monetary policy decision through maximizing the decision-maker’s utility function and weighting models conditionally on that decision. The model weights are determined by their empirical fit, past and expected decision-making performance, and the model-based plausibility of the policy decision. We find that BPDS produces quite different decisions and weights from standard approaches, such as Bayesian Model Averaging, that only consider forecasting performance.

Advisor / supervisor