Clustering Behavioral Regimes in the Iranian Capital Market Using Self-Organizing Maps (SOM): An Unsupervised Learning Approach (A Proof-of-Concept Study)

Abstract

This study aimed to evaluate, as a proof of concept, the capability of a Self-Organizing Map (SOM) to identify and characterize latent behavioral regimes in multidimensional Iranian capital-market data through an unsupervised learning framework. This quantitative, computational, exploratory, and applied study used multivariate financial observations representing dimensions such as return, volatility, trading activity, market conditions, and selected firm-level characteristics. After data cleaning, assessment of missing and extreme values, temporal alignment, and scaling, the observations were entered into a two-dimensional 6 × 6 SOM implemented in Python using the MiniSom library. The model used a Gaussian neighborhood function, an initial learning rate of 0.50, and 1,000 training iterations. Cluster extraction was based on the topological organization of the trained map and U-Matrix patterns. Model performance was assessed using training and quantization error, intra-cluster distance, visual inspection of the U-Matrix, and cluster-validity criteria including the Silhouette coefficient and Davies–Bouldin index. The network training error decreased from 0.482 to 0.067, representing an approximate reduction of 86.1% and indicating substantial convergence during learning. U-Matrix analysis revealed nonuniform inter-neuron distances and distinguishable topological regions, supporting the existence of heterogeneous latent structures. Four final clusters were extracted. Cluster 1 contained 175 observations and showed the greatest internal cohesion with a mean intra-cluster distance of 0.23; Cluster 2 contained 152 observations with a mean distance of 0.31; Cluster 3 included 108 observations with a mean distance of 0.44; and Cluster 4 included 85 observations and exhibited the greatest heterogeneity with a mean intra-cluster distance of 0.51. The findings indicate that SOM can provide a meaningful unsupervised representation of heterogeneous behavioral regimes in Iranian capital-market data and may serve as a useful exploratory foundation for subsequent regime-based forecasting, portfolio management, and risk-monitoring studies.