AR-PolieDRO: DRO regression framework applied to time series

This work investigates the application of the PolieDRO framework in the estimation of autoregressive AR(p) models in time series. PolieDRO is a regression framework derived from robust optimization to distributions (DRO) formulations, which does not require choice of hyperparameters. To evaluate its feasibility, experiments are conducted with simulated time series of different orders and lengths, with and without outliers. The results show that the PolieDRO presents performance comparable to the estimated AR model via maximum likelihood. Although the benchmark is superior in some scenarios, especially with outliers, PolieDRO demonstrates potential as a robust alternative estimator without the need for manual adjustments.