This paper investigates an auto-regressive method for time series forecasting and applied to the challenge of predicting groundwater levels. An auto-regressive method estimates a next value of a time series by regression with the historical values of the series. Several regression methods can be used. In this paper, experiments are presented to identify the best setting to accurately predict the groundwater levels. Different classifiers, different modes of learning (by series or by group of series), and different usages of exogenous data are compared. Intensive experiments have been conducted and allow us to conclude on the best method we will use to answer the challenge.