An efficient alternative to CDKM for size-constrained clustering on financial data analysis

k -means is a well-known method that partitions data into clusters to support downstream tasks such as financial market risk forecasting, quantitative portfolio optimization, and fraud detection. However, the traditional k -means solutions are fragile on credit data due to highly imbalanced cluster sizes, which hinder meaningful analysis on credit policy rules. Therefore, we propose a size-constrained k -means framework (denoted as SCKM) from an alternative perspective that integrates a reward term to prevent highly imbalanced cluster sizes. Specifically, we reformulate the objective into a trace maximization form with a reward term. Then we develop a reward-guided alternating descent method to optimize the objective. Finally, we speed up the framework by introducing an improved version that pre-stores intermediate results. Experiments on 6 real-world credit datasets and 1 non-financial dataset show that SCKM controls the cluster size imbalance to 5.28%–7.93% while achieving competitive Sum of Squared Errors (SSE) among compared methods. It also improves clustering efficiency, reducing runtime by 75.1%–91.3% over CDKM. In a personal lending scenario, the credit rating standard developed by SCKM is effective.