Robust Som for Realistic Data Completion

Self-Organizing Maps aims ideally to group homogeneous individuals, highlighting a neighbourhood structure between classes in a chosen network. Recent approaches propose to exploit the homogeneity of the underlying classes for data completion purposes (see [2]). The aim of this paper is two-fold. First, we present and slightly modified two complementary approaches in completing the stochastic method proposed by Rousset and Maillet [11] based on bootstrap process for increasing the reliability of the induced neighbourhood structure and, second, we use the induced Robust Map of the last approach for data completion, generalising the results by Merlin and Maillet [9] with robust statistics of the moments of the series. An empirical illustration of this new completion scheme is finally provided based on a sample of Hedge Fund Net Asset Values.