Empirical studies using survey data on expectations have frequently observed that forecasts are biased and have concluded that agents are not rational. We establish that existing rationality tests are not robust to even small deviations from symmetric loss and hence have little ability to tell whether the forecaster is irrational or the loss function is asymmetric. We quantify the exact trade-off between forecast inefficiency and asymmetric loss leading to identical outcomes of standard rationality tests and explore new and more general methods for testing forecast rationality jointly with flexible families of loss functions that embed quadratic loss as a special case. An empirical application to survey data on forecasts of nominal output growth demonstrates the empirical significance of our results and finds that rejections of rationality may largely have been driven by the assumption of symmetric loss.

This paper proposes a new regression model — a smooth transition mixed data sampling (STMIDAS) approach — that captures recurrent changes in the ability of a high frequency variable in predicting a variable only available at lower frequency. The model is applied to the use of …nancial variables, such as the slope of the yield curve, the short-rate and stock returns, to forecast US output growth both in- and out-of-sample. I …nd evidence that the use of the predictor sampled weekly improves output growth forecasts, which may also be improved when changes in …nancial variables’predictive power are considered.

In this paper we first develop two statistical tests of the null hypothesis that early release data are rational. The tests are consistent against generic nonlinear alternatives, and are conditional moment type tests, in the spirit of Bierens (1982,1990), Chao, Corradi and Swanson (2001) and Corradi and Swanson (2002). We then use this test, in conjunction with standard regression analysis in order to individually and jointly analyze a real-time dataset for money, output, prices and interest rates. All of our empirical analysisiscarriedoutusingvariousvariable/vintage combinations, allowing us to comment not only on rationality, but also on a number of other related issues. For example, we discuss and illustrate the importance of the choice between using first, later, or mixed vintages of data in prediction. Interestingly, it turns out that early release data are generally best predicted using first releases. The standard practice of using “mixed vintages ” of data appears to always yield poorer predictions, regardless of what we term “definitional change problems ” associated with using only first releases for prediction. Furthermore, we note that our tests of first release rationality based on ex ante prediction find no evidence that the data rationality null hypothesis is rejected for a variety of variables (i.e. we find strong evidence in favor of the “news ” hypothesis). Thus, it appears that there is little benefit to using later releases of data for prediction and policy analysis, for example. Additionally, we argue that the notion of finaldata is misleading, and that definitional and other methodological

In the past years the academical and industrial interest in portfolio risk forecasting has grown rapidly. The importance of this issue is to due to the fact that the regulation of banks is based on a certain risk measure, namely Value-at-Risk (VaR). After many financial market crises occurred all over the world in the last two decades, the

Acknowledgements: We thank Christiane Baumeister for providing access to the world and OECD industrial production data and Ryan Kellogg for providing the Michigan survey data on gasoline price expectations. We thank Domenico Giannone for providing the code generating the Bayesian VAR forecasts. We have benefited from discussions with Christiane Baumeister, Mike

Abstract. When comparing two competing approximate models, the one having smallest ‘expected true error ’ is closest to the data generating process (according to the specified loss function) and is therefore to be preferred. In this paper we consider a data-driven method of testing whether two competing approximate models, for instance a parametric and a nonparametric model, are equivalent in terms of their expected true error (i.e., their expected performance on unseen data drawn from the same data generating process). The proposed test is quite flexible with regards to the types of models and data types that can be compared (i.e., time-series, cross section, panel etc.). Moreover, by applying our method to time-series models we can overcome two of the drawbacks associated with existing approaches, namely, the reliance on only one split of the data and the need to have a sufficiently large hold-out sample in order for the test to have power. Some useful graphical summaries are also presented. Finite-sample performance and several illustrative applications are considered. 1.