Subdivided Clustering for Enhanced Predictive Accuracy

In the regression-based prediction problem using clustering methods, it can be questionable to follow the conventional clustering method that considers every observation in data to obtain well-partitioned clusters. Unlike the classification problem, when aiming to predict a target value from a temporal perspective, it might not be efficient to focus on partitioning the entire data. However, it could be more efficient to focus on finding a single cluster right next to a prediction target for accurate prediction. Thus, this study proposes an approach to obtain this single cluster, which recursively partitions a cluster from the previous stage of clustering, thereby enhancing the prediction accuracy of existing clustering methods. Empirical validation using the Standard and Poor’s 500 index with economic and financial variables across different time intervals demonstrated the efficacy of the proposed approach. Enhancing predictive accuracy and stability, the proposed approach based on the fundamental principles of prediction would facilitate comprehensive applicability in prediction problem.