Learning with little mixing

Neural Information Processing Systems 

We study square loss in a realizable time-series framework with martingale difference noise.