Motivated by limitations in applying existing unsupervised classification methods to economic data partitioning, a nonparametric, deterministic, and robust data-driven partitional hard cluster analysis, derived from a novel similarity measure, is introduced. The main objective of the proposed method in the present study is to support financial investors in the portfolio diversification process by providing a less arbitrary approach to quantify similarity levels between investment alternatives (pairwise) as well as revealing the clustering structure (whole sample) and data patterns through time. This is especially useful during periods of turmoil (e.g. financial contagion episodes, such as the Global Financial Crisis of 2007-2008), when investment alternatives tend to become more similar and, therefore, harder to distinguish between themselves. An algorithm is constructed in order to run the proposed white-box method, which dynamics may be readily interpreted through a clear data visualisation. The conceptual benefits and caveats of the proposed method is compared to the well-established and most used cluster method (i.e. the k-means and two of its variations, using the k-means++ algorithm) applied to the most popular benchmark data (i.e. Fisher’s Iris dataset). Moreover, empirical results applied to a period of 15 years of real-world time series of the most relevant economies worldwide provide statistically significant evidence that the clustering structure of international stock markets effectively changes according to market conditions.