GNPR Distances for Clustering Time Series by Distribution and Dependence

Summary

This article presents the generic non-parametric representation (GNPR) distance for comparing time series using both distributional and dependence information. The motivation is that correlation or other familiar similarity measures can make series appear alike even when their return distributions, and therefore their risk characteristics, differ. GNPR is contrasted with generic parametric representation, which the article describes as suitable for normally distributed variables, while GNPR can also distinguish non-normal distributions.

The examples use synthetic correlated random walks with normal, Laplace, and Student-t distributions, then compare GNPR with distance correlation, Spearman’s rho, and GPR. The reported illustrations show GNPR separating distribution clusters that the other measures do not clearly distinguish; all compared methods separate the example’s correlation clusters. A tuning parameter controls the balance between distribution and dependence information, so its useful setting depends on the data. These are generated-data demonstrations rather than evidence of improved investment performance, and the article notes that even GNPR may not perfectly distinguish similar distributions, such as normal series with different standard deviations.

Key ideas

• Correlation alone can miss differences in return distributions that matter for risk comparisons.
• GNPR compares time series using both distributional and dependence information.
• Synthetic examples show GNPR separating several non-normal distribution clusters.
• A tuning parameter controls the balance between distribution and dependence information.
• The demonstrations use generated data and do not establish trading performance.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.