The study is devoted to testing the hypothesis about the possibility of creating a diversified portfolio of stocks by clustering them according to risk and sustainability indicators. Within the framework of this problem, three clustering methods were used: the K-means method, the Gaussian mixture model, and hierarchical clustering by the Ward method. We also selected three criteria for stock clustering: Beta index, volatility, and ROE. According to the data from 2018 to 2023, clusterization was carried out for each period with the selection of the optimal number of clusters, and portfolios of financial assets were created according to predefined rules. Alternatively, stock portfolios were created for each year without prior clustering according to the rules of Markowitz portfolio theory. After that, each portfolio was tested over the next year. The results based on the metrics of profitability and risk show that the hypothesis has a right to exist. Portfolios created with pre-clustering outperformed the classic options in terms of profitability. The problem has scope for further investigation. In particular, a different set of features for stock segmentation or the use of clustering methods in combination with other machine learning or neural network algorithms can also produce high-quality results.