Beyond computation, generalization is a cornerstone of stationary time series analysis. Stronger than good on-average performance, generalization guarantees that the expected performance of an estimator improves over every time step of a data stream. Consequently, low regret does not imply generalization. It is, however, possible to replace pure regret with a hybrid objective that becomes a batch (instantaneous) excess risk when the data is sufficiently regular. I propose to investigate whether a CFA would remain possible, or how a CFA in the pure regret setting can generalize, if not optimally.