Automatic identification of ARIMA models: The case of water consumption forecasting

Selecting an ARIMA model is addressed using the usual goodness of fit measures, and, also, with some proposed predictive criteria. Both methods are compared with the aim of obtaining optimum forecasts. Several statistics such as BIC and others, obtained within the sample range, are used with simulated data, and some smoothing predictive criteria are compared as a tool for automatic selection of a 'real' model. Classical methods identify the correct model barely in a third of the cases, as it is show with simulated series, and, very often a different model produces better forecasts. A case study with water consumption in urban areas is presented with the comparative results attained with both approaches.