Clustering Techniques And their Effect on Portfolio Formation and Risk Analysis

Abstract

This paper explores the application of three different portfolio formation rules using standard clustering techniques-K-means, K-mediods, and hierarchical-to a large financial data set (16 years of daily CRSP stock data) to determine how the choice of clustering technique may affect analysts' perceptions of the riskiness of different portfolios in the context of a prototype visual analytics system designed for financial stability monitoring. We use a two-phased experimental approach with visualizations to explore the effects of the different clustering techniques. The choice of clustering technique matters. There is significant variation among techniques, resulting in different "pictures" of the riskiness of the same underlying data when plotted to the visual analytics tool. This sensitivity to clustering methodolgy has the potential to mislead analysts about the riskiness of portfolios. We conclude that further research into the implications of portfolio formation rules is needed, and that visual analytics tools should not limit analysts to a single clustering technique, but instead should provide the facility to explore the data using different techniques.

A Proposal of a Methodological Framework with Experimental Guidelines to Investigate Clustering Stability on Financial Time Series

“…Random Matrix Theory Asymptotics [2], High Dimension Moderate Sample Size Asymptotics (HDMSS) and High Dimension Low Sample Size Asymptotics (HDLSS) [3]. Besides lacking from these theoretical properties for now, clusterings obtained are much dissimilar from one algorithm to another when applied on financial time series [4], [5]. Even worse, adding small noise to a given sample and applying the same algorithm on the original one and its perturbated version yields different clusters.…”

Clustering high‐frequency financial time series based on information theory

“…34 Clustering is often employed to overcome such difficulties by grouping homogeneous stocks, thus reducing the dimensionality of the assets. 35,36 F I G U R E 3 Holt-Winters double exponential smoothing of comovement probability of BMY and LLY up to each week…”

Toward a generic representation of random variables for machine learning

“…Yet, practitioners and researchers pinpoint that state-of-the-art results of clustering methodology applied to financial times series are very sensitive to perturbations (Lemieux et al, 2014). The observed unstability may result from a poor representation of these time series, and thus clusters may not capture all the underlying information.…”