Competing Forecasts · Comparing the Predictive Ability of Forecasting Models

Comparing the Predictive Ability of Forecasting Models

This project studies the problem of identification failure in a variety of econometric models. The main aim is to analyze the impact of this issue on estimation, inference and forecast comparison, using several models. The research shows that whenever the model is not strongly identified, the finite sample distributions of estimators and test statistics are affected, in the sense that these can be nonstandard. This might result then in misleading estimation and inference, which could eventually affect policy decisions when using these models at central banks, or other policy institutions. In order to overcome these issues, the projects proposes robust methods for inference. From a theoretical point of view, identification of econometric models is crucial for valid estimation, inference and testing. From a policy perspective, as the policy implications of observational equivalent parameter points can be distinct, the reliability of policy recommendations rests upon the primitive assumptions of identifiability. This could potentially lead to important problems, if the researcher does not acknowledge the issue, but assumes a standard normal distribution. Decisions about the significance of parameters, and confidence intervals could be misleading. Thus, this research promotes the importance of developing tools and methods that are robust to this issue. Results emerging from this project are of interest to a large international academic community interested in predictive ability evaluation, central banks and other governmental organizations that could take-up the new knowledge for policy decisions, as well as to forecasting institutions and businesses that produce predictions which could improve their evaluation methodologies. These governmental and industry institutions are direct potential users of the project results. The following scientific overall objectives have been achieved: - The project shows that whenever the model is not strongly identified, the finite sample properties of estimators are affected. The distribution of the estimators for different parameters will have nonstandard/unexpected shapes: bimodal, uniform distributions appear quite often. - Similarly, the distribution of tests such as the t-test or the likelihood ratio tests are also affected. The identification issues transfer from estimators to tests. Thus, decisions about the significance of parameters and confidence intervals could be misleading. - The problem can also appear when comparing the predictive ability of different models and thus the finite sample and asymptotic properties of predictive ability tests are studied. - In order to circumvent the problem three robust critical values are proposed. These rely on the actual finite sample distribution of tests and are shown to perform well in simulations and applications. - Two empirical application showcase the prevalence of this issue in financial and macroeconomic data and illustrate the use of the proposed critical values.

Текст от CORDIS, на английски · Данни: CORDIS, © Европейски съюз