On Optimality of Score Driven Models

The contribution of this thesis consists in proving that score driven models possess a novel, intuitive, high dimensional and global optimality criterion, called Conditional Expected Variation optimality that formalizes the following words from Creal et al. (2013) "The use of the score is intuitive. It defines a steepest ascent direction for improving the model's local fit in terms of the likelihood or density at time t given the current position of the parameter. This provides the natural direction for updating the parameter. " Indeed, the fact that the score defines a steepest ascent direction is crucial in deriving the results and for the proposed optimality criterion to hold. To prove the aforementioned property, a point of contact between the econometric literature and the time varying optimization literature will be established. As a matter of fact, the Conditional Expected Variation optimality can be naturally viewed as a generalization of the monotonicity property of the gradient descent scheme. A number of implications on the specification of score driven models are analyzed and discussed, even in the case of model misspecification.