1: \begin{abstract}

2: In this work, we study the problem of aggregating a finite number of

3: predictors for nonstationary sub-linear processes. We provide oracle

4: inequalities relying essentially on three ingredients: (1)~a uniform

5: bound of

6: the $\ell^1$ norm of the time varying sub-linear coefficients, (2)~a Lipschitz

7: assumption on the predictors and (3)~moment conditions on the noise appearing

8: in the linear representation. Two kinds of aggregations are considered giving

9: rise to different moment conditions on the noise and more or less sharp

10: oracle inequalities. We apply

11: this approach for deriving an adaptive predictor for locally stationary time

12: varying autoregressive (TVAR) processes. It is obtained by aggregating a

13: finite number of well chosen predictors, each of them enjoying an optimal

14: minimax convergence rate under specific smoothness conditions on the TVAR

15: coefficients. We show that the obtained aggregated predictor achieves a

16: minimax rate while adapting to the unknown smoothness. To prove

17: this result, a lower bound is established for the minimax rate of the

18: prediction risk for the TVAR process. Numerical experiments complete this

19: study. An important feature of this approach is that the aggregated predictor

20: can be computed recursively and is thus applicable in an online prediction

21: context.

22: \end{abstract}