Schools rankings based upon their added value, i.e. the learning progress registered by their students, are subject to large estimation errors. The paper shows how those errors may be reduced, by comparing strategies based purely upon the enlargement of the information/data set about the students' starting point with those using machine learning techniques. A real life application to Italian middle schools is presented.
Machine learning techniques are an inexpensive way to boost the precision of estimates, providing a low-cost and more reliable alternative to widening the control variables set. The ranking precision improves, especially for the two extremes of the distribution, where the interest of the policy maker is concentrated in order to identify either very strong or very weak schools (which are to be the object of specific attention and supervisory efforts, using tools which may differ according to the institutional and political environment).
