Neural network approach revolutionizes time series clustering with statistics

A new paper from Ángel López-Oriona and Ying Sun proposes a fundamentally different way to cluster time series data. Traditional methods like K-means, K-medoids, or hierarchical clustering require users to choose an algorithm, specify cluster shapes, and often manually select the number of clusters—a brittle, heuristic-heavy process. The authors' approach, called amortized neural clustering, replaces these steps with a neural network trained on simulated time series. The network learns to map statistical features—primarily autocorrelations and quantile autocorrelations—directly to optimal cluster partitions. This data-driven affinity structure eliminates the need for explicit objective functions or iterative heuristics, and in one version, the model autonomously determines the correct number of clusters from the data itself.

Comprehensive empirical studies show the method achieves competitive or superior clustering accuracy compared to conventional techniques, even when those techniques are given the true number of clusters—a significant advantage in real-world scenarios where the cluster count is unknown. The researchers validated the framework on financial time series of stock returns, demonstrating its practical utility. By reducing reliance on algorithm selection, calibration, and ad-hoc decisions, this approach opens new possibilities for automated, adaptive clustering of temporal data across scientific and industrial domains, from sensor networks to economic forecasting.