Based on large simulations study, this paper investigates which strategy to adopt in order to choose the best forecasting model --- in terms of accuracy --- for Mixed causal-noncausal AutoRegressions (MAR) data generating processes: always differencing (D), never differencing (L) or unit root pretesting (P). Relying on recent econometric developments regarding forecasting and unit root testing in this MAR framework, the main results suggest that from a practitioner's point of view, the P strategy at the 1\%-level is a good compromise. In fact, it never departs too much from the best model in terms of forecast accuracy, unlike the L (respectively D) strategy when the DGP becomes very persistent (respectively with little persistence).