The classic ARIMA models use the information criteria for lag selection since 1990s. The information criteria are based on the summation of two expressions: a function of Residual Sum of Squares (RSS) and a penalty for decrease of degrees of freedom. However, the information criteria have some disadvantages since these two expressions do not have the same scale, so the information criteria are mainly based on the first expression (because of its bigger scale). In this paper, we propose a hybrid ARIMA model, which uses the Data Envelopment Analysis (DEA) model to select the best lags of AR and MA process called DEA-ARIMA. DEA is a linear programming technique, which computes a comparative ratio of multiple outputs to multiple inputs for each Decision Making Unit (DMU), which is reported as the relative efficiency score. We identify inputs as the number of AR and MA terms and outputs of the model are inverse of Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). In fact, in our proposed model, inputs consider as resources, so we are looking for some models with fewer resources and high efficiency. The DEA unlike the information criteria may have more than one solution and all of them are efficient so to compare this two models selection the mean of best DMUs is calculated. Experimental results demonstrate DEA-ARIMA will not trap in over fitting problem in contrast to classic ARIMA models because of considering a set of efficient ARIMA models.