Testing Clustered Equal Predictive Ability with Unknown Clusters

Abstract: We develop new tests of clustered equal predictive ability (C-EPA) in panels where the clusters are unknown and estimated by a Panel Kmeans algorithm. This algorithm differs from the standard Kmeans algorithm by employing the time series variation of the panel rather than relying merely on time averages of observations. To address the challenge of testing hypotheses that depend on data-driven cluster estimates, we adopt a selective conditional inference framework. Specifically, we derive a Wald-type test statistic for pairwise equality and show that the limiting distribution of its square root conditional on the estimated cluster structure is that of a truncated $\chi$ random variable. We characterize the associated truncation set as a polyhedron in the data space. As a test of the C-EPA hypothesis, we propose a $p$-value combination method which aggregates the evidence against the pairwise equality and overall EPA null hypotheses. In addition, we prove that using an information criterion to select the unknown number of clusters under the alternative hypothesis prior to testing does not require further conditioning to obtain a valid test. Monte Carlo simulations confirm the excellent finite sample performance of the proposed tests. An empirical application to forecasting exchange rates using traditional time series models as well as machine learning methods illustrates the practical importance of our procedure.