This paper proposes and analyzes tests that can be used to compare the accuracy of alternative conditional density forecasts of a variable. The tests are also valid in the broader context of model selection based on out-of-sample predictive ability. We restrict attention to the case of density forecasts derived from non-nested parametric models, with known or estimated parameters. The evaluation makes use of scoring rules, which are loss functions defined over the density forecast and the realizations of the variable. In particular, we consider the logarithmic scoring rule, which leads to the development of asymptotic and bootstrap 'weighted likelihood ratio' tests. The name comes from the fact that the tests compare weighted averages of the scores over the available sample, as a way to focus attention on different regions of the distribution of the variable. For a uniform weight function, the asymptotic test can be interpreted as an extension of Vuong (1989)' s likelihood ratio test for non-nested hypotheses to time series data and to an out-of-sample testing framework. A Monte Carlo simulation explores the size and power properties of this last test in finite samples. An application using S&P500 daily returns shows how the tests can be used to compare the performance of density forecasts obtained from GARCH models with different distributional assumptions.

This paper proposes a method for comparing and combining conditional quantile forecasts based on the principle of 'encompassing'. Our test for conditional quantile forecast encompassing (CQFE) is a test of superior predictive ability, constructed as a Wald-type test on the coefficients of an optimal combination of alternative forecasts. The CQFE test is a 'model free' test that can be used to compare any given number of alternative forecasts, and is relatively easy to implement by GMM techniques appropriately modified to accommodate non-differentiable criterion functions. Further, our theoretical framework provides a basis for combining quantile forecasts, when neither forecast has superior predictive ability. A central feature of our method is the focus on conditional, rather than unconditional expected loss in the formulation of the encompassing test, which links our approach to Christoffersen's (1998) 'conditional coverage' test for evaluation of quantile forecasts. An empirical application to the problem of Value at Risk evaluation illustrates the usefulness of the proposed techniques.

We argue that the current framework for predictive ability testing (e.g., West, 1996) is not necessarily useful for real-time forecast selection, i.e., for assessing which of two competing forecasting methods will perform better in the future. We propose an alternative framework for out-of-sample comparison of predictive ability which delivers more practically relevant conclusions. Our approach is based on inference about conditional expectations of forecasts and forecast errors rather than the unconditional expectations that are the focus of the existing literature. We capture important determinants of forecast performance that are neglected in the existing literature by evaluating what we call the forecasting method (the model and the parameter estimation procedure), rather than just the forecasting model. Compared to previous approaches, our tests are valid under more general data assumptions (heterogeneity rather than stationarity) and estimation methods, and they can handle comparison of both nested and non-nested models, which is not currently possible. To illustrate the usefulness of the proposed tests, we compare the forecast performance of three leading parameter-reduction methods for macroeconomic forecasting using a large number of predictors: a sequential model selection approach, the "diffusion indexes" approach of Stock and Watson (2002), and the use of Bayesian shrinkage estimators.

In this paper we compare the relative efficiency of different methods of forecasting the aggregate of spatially correlated variables. Small sample simulations confirm the asymptotic result that improved forecasting performance can be obtained by imposing a priori constraints on the amount of spatial correlation in the system. One way to do so is to aggregate forecasts from a Space-Time Autoregressive model (Cliff et al., 1975), which offers a solution to the 'curse of dimensionality' that arises when forecasting with VARs. We also show that ignoring spatial correlation, even when it is weak, leads to highly inaccurate forecasts. Finally, if the system satisfies a 'poolability' condition, there is a benefit in forecasting the aggregate variable directly.

We propose a new family of density function that posses both flexibility and closed form expressions for moments and anti-derivatives, making them particularly appealing for applications. We illustrate its usefulness by applying our new family to obtain density forecasts of U.S. inflation. Our methods generate forecasts that improve on standard methods based on AR-ARCH models relying on normal or Student's t-distributional assumptions.