Non-asymptotic calibration and resolution

Abstract

We analyze a new algorithm for probability forecasting of binary labels, without making any assumptions about the way the data is generated. The algorithm is shown to be well calibrated and to have high resolution for big enough data sets and for a suitable choice of its parameter, a kernel on the Cartesian product of the forecast space [0,1] and the object space. Our results are non-asymptotic: we establish explicit inequalities for the performance of the algorithm. © Springer-Verlag Berlin Heidelberg 2005.