Title: Conjugate Processes: Theory and Application to Risk Forecasting

Abstract: Many dynamical phenomena display a cyclic behavior, in the sense that time can be partitioned into units in which distributional aspects of a process are homogeneous. In this paper, we propose a class of models which allow the marginal distributions of a cyclic, continuous-time process to evolve stochastically in time, in principle separated from the observable process itself. The connection between them is given by a compatibility condition on the conditional marginal distributions. The methodology relates to the concept of random measure and more generally to the theory of probability in function spaces. From the inferential point of view our method can be seen as Functional Data Analysis. We provide a constructive example which illustrates the method, and a statistical implementation of our model to risk forecasting in financial data is given. In summary, we generate forecasts of intraday asset returns variance and Value-at-Risk. The forecasts are attainable by reducing the dimension of the conditional distribution process into a latent scalar time series.