1: \begin{abstract}

2: Motivated by a variety of applications,

3: high-dimensional time series have become an active topic of research.

4: In particular, several methods and finite-sample theories for individual stable autoregressive processes with known lag have become available very recently. We, instead, consider multiple stable autoregressive processes that share an unknown lag. We use information across the different processes to simultaneously select the lag and estimate the parameters. We prove that the estimated process is stable, and we establish rates for the forecasting error that can outmatch the known rate in our setting.

5: Our insights on the lag selection and the stability are also of interest for the case of individual autoregressive processes.

6: % We design efficient stochastic algorithm for learning a family of $M$ stable sparse autoregressive processes --- of identical lag-order $L_0$ --- with iid gaussian innovations, in the setting where number of samples from at most $cM$ of the models are less than optimal for best available $M$-dimensional VAR-model learning algorithm to attain the expected (inverse linear) rate of convergence; we show that our algorithm provably produces parameter-vector that lies in the region of stability, while matching the best available bound on the in-sample, one-step-ahead mean squared forecasting error. In doing so, we successfully implement the techniques of transfer learning in the settings of time-series forecasting.

7: \end{abstract}

8: