In this paper, we introduce and analyze Discriminative State-Space Models for forecasting non-stationary time series. We provide data-dependent generalization guarantees for learning these models based on the recently introduced notion of discrepancy. We provide an in-depth analysis of the complexity of such models. Finally, we also study the generalization guarantees for several structural risk minimization approaches to this problem and provide an efficient implementation for one of them which is based on a convex objective.

Based on these bounds, a structural risk minimization formulation is proposed to estimate forecasting models with learning guarantees both in the case where the state-space predictor is not neccesarily accurate and in the case where we assume that it is. The authors show that for many models of interest, this is a reasonable assumption. The convex objective function of the resulting SRM is then solved using a coordinate descent algorithm, with some encouraging empirical results presented in the appendix.

In this paper, we introduce and analyze Discriminative State-Space Models for forecasting non-stationary time series. We provide data-dependent generalization guarantees for learning these models based on the recently introduced notion of discrepancy. We provide an in-depth analysis of the complexity of such models. Finally, we also study the generalization guarantees for several structural risk minimization approaches to this problem and provide an efficient implementation for one of them which is based on a convex objective. Papers published at the Neural Information Processing Systems Conference.