Klassifizierung von Hedge-Fonds durch das k-means Clustering von Self-Organizing Maps: eine renditebasierte Analyse zur Selbsteinstufungsgüte und Stiländerungsproblematik[Classifying Hedge Funds using k-means Clustering of Self-Organizing Maps: a return-based analysis of misclassification and the problem of style creep]

Through an implementation of the 2-level-approach due to Vesanto & Alhoniemi (2000), this paper addresses a number of problems typically seen when visualized interpretation of Self Organizing Maps (SOM) are applied to derive a systematic classification system in the hedge fund literature. Normally, a trained SOM does not result in an exact depiction of the detected structures of the input data, and is therefore challenging for visual interpretations. The 2-level-approach overcomes this problem and assures a consistent clustering of neighboring output units, and therefore an objective classification scheme. Through an empirical application, such an objective classification is derived. Building on this, further analyses concerning the misclassification and style creep problems are conducted. Within the ten-year sample period (31.01.1999 to 31.12.2008), which comprises 2789 hedge funds, organized in eleven strategies, six classes can be identified. This six-class taxonomy is fairly robust to different sub-sample periods, topologies and data-samples. According to the classification system applied here, it is shown that most of the analyzed hedge funds are inconsistent in their self-declared strategies. Furthermore, evidence of undisclosed trading style changes over time is identified – specifically, it is shown that misclassified hedge funds are more likely to change their trading style.