Econometrics, Statistics and Empirical Economics

Robust estimation of time-varying moments, mutual information, and transfer entropy by means of quantile regression based density forecasts

Our goal is to simplify and accelerate the detection of non-linear structures and their characteristics in multivariate time series and cross-sectional data. To do so, we will examine the estimation of conditional and unconditional moments based on quantile regression. Therefore, we build on the ideas in Baur and Dimpfl (Journal of Financial Econometrics), to estimate the conditional variance of returns. Furthermore, we want to develop estimation and testing methods for relative entropy measures, such as mutual information or transfer entropy. The key to our proposed methodology is the decomposition of multivariate, joint density functions into conditional and unconditional densities which are needed to calculate entropy measures and moments. These can be modeled by a quantile regression. Due to its semi-parametric character and based on the available literature on the asymptotic theory of estimated quantiles, few assumptions are needed to model the densities, the computational effort is reduced, a smaller data size (compared to for example multivariate kernel density estimators) is sufficient and a consistent interpretation of the individual measures is to be expected. In particular, with the proposed method to calculate the entropy measures for continuous random variables, it is possible to dispense with the discretization of data which currently prevails in the literature. In addition, an asymptotic distribution theory for the entropy and moment estimators can be obtained by appropriate formulation of the quantile regression. This is the basis for developing for the first time flexible test procedures in the entropy context (as opposed to purely permutation-based tests), but also for novel tests for (conditional) moments of a random variable. Still, there are also new issues that need to be solved, like smoothing the estimated density functions, or the optimal support during the numerical integration. To investigate the advantages and disadvantages of the proposed methodology, in detail, is the focus of this project.

Robuste Schätzung von zeitvariierenden Momenten, Transinformation und Transferentropie mittels Dichteprognosen aus der Quantilsregression

Mit unserem Vorhaben wollen wir dazu beitragen, das Aufdecken von nicht-linearen Strukturen und deren Charakteristika in multivariaten Zeitreihen- und Querschnittdaten zu vereinfachen und zu beschleunigen. Dazu werden wir mittels Quantilsregressionen die Schätzung von bedingten und unbedingten Momenten untersuchen. Dabei bauen wir auf den Ideen zur Schätzung der bedingten Varianz von Renditen in Baur und Dimpfl (Journal of Financial Econometrics) auf. Darüber hinaus wollen wir Schätz- und Testmethoden für relative Entropiemaße wie Transinformation oder Transferentropie entwickeln. Grundlegend für die Anwendung ist die Zerlegung von multivariaten, gemeinsamen Dichtefunktionen in bedingte und unbedingte Dichten, die für deren Berechnung benötigt werden. Diese können durch eine Quantilsregression modelliert werden. Durch deren semi-parametrischen Charakter und aufbauend auf der vorhandenen Literatur zur asymptotischen Theorie der geschätzten Quantile sind eine annahmenarme Modellierung, vereinfachter Rechenaufwand, geringerer Datenumfang (im Vergleich beispielsweise zu multivariaten Kerndichteschätzern) sowie eine konsistente Interpretation der einzelnen Maße zu erwarten. Insbesondere kann mit der vorgeschlagenen Methode bei der Berechnung der Entropiemaße für stetige Zufallsvariablen auf die bisher in der Literatur vorherrschende Diskretisierung verzichtet werden. Zudem erschließt sich durch geeignete Formulierung der Quantilsregression eine asymptotische Verteilungstheorie für die geschätzten Entropiemaße und Verteilungsmomente. Diese stellt die Grundlage dar, um erstmals flexible Testverfahren im Entropiekontext (im Gegensatz zu rein permutationsbasierten Tests), aber auch neuartige bezüglich bedingter Momente einer Zufallsvariablen zu entwickeln. Andererseits stellen sich neue Probleme bei der Glättung der geschätzten Dichtefunktionen oder die optimale Stützzahl bei einer numerischen Integration. Die Vor- und Nachteile der Methodik im Detail auszuarbeiten soll Gegenstand unserer Forschungsarbeit sein.