Identification of latent clusters of investment products based on risk factors using machine learning methods in the Russian stock market

Subject. This article examines the latent structure of risk profiles of retail investment products. Objectives. The article aims to determine stable clusters based on multi-factor risk metrics that reflect investors’ potential annual losses. Methods. The analysis is based on a rolling sample of 60-day windows for each fund, which captures the temporal dynamics of risk. For each window, the following indicators are calculated: volatility, Value at Risk (VaR), Conditional VaR (CVaR), intraday range, trading volume, and a liquidity metric (days-to-liquidate). Features are standardized and processed using machine learning methods: k-means clustering and principal component analysis (PCA). Silhouette score, Hopkins statistic, and inter-cluster distances are calculated to assess clustering quality. Results. The analysis shows a high degree of clusterability in the data. The k?means model with five clusters reveals stable groups of observations differing in the depth of potential losses and liquidity. Inter-cluster distances ranged from 2.15 to 9.67 in the normalized feature space, indicating strong differentiation in risk profiles. Conclusions and Relevance. Clustering based on multidimensional risk factors enables the identification of hidden risk structures that are not observable through single-metric analysis. The application of unsupervised machine learning expands the possibilities for empirical analysis and contributes to a more accurate segmentation of investment products by their potential loss levels. The proposed approach can be used for the automated typology of investment products within internal risk management models, as well as for developing universal risk indicators. The results can assist financial institutions and regulators in auditing and refining existing risk rating scales, including their adjustment to behavioral and market-specific factors.