Learning Opportunity Just Shared to the PRMIA Risk Library
For your continued learning, we are pleased to share two whitepapers researched by Bloomberg with lead author Jan Dash.
Stable Reduced-Noise "Macro" SSA-Based Correlations for Long-Term Counterparty Risk
This paper introduces a methodology from geophysics, Singular Spectrum Analysis (SSA), to obtain stable, noise-cleaned correlations for long term risk (e.g. counterparty risk). SSA is applied to time series to smooth them in a robust manner. The SSA-smoothed time series are then used to obtain the correlations. These are called “macro” correlations because they are determined with macroscopic time scales. Stable correlations are desirable to suppress noise from short time scales that make risk measures unstable. If correlations move around, risk measures also move around, making business decisions difficult. SSA-based correlations ameliorate this business problem.
Introduction to Noise-reduced Correlations Using Singular Spectrum Analysis
This paper summarizes new results for estimating correlations for use in risk management. These estimates have better behavior than traditional estimation approaches from both a business standpoint and a technical standpoint. We smooth time series using Singular Spectrum Analysis (SSA) and compute correlations based on these smoothed series. We demonstrate that SSA-based correlation estimates have less noise than standard correlation estimates between unsmoothed series using: the signal-to-noise ratio, and distances from noise using polynomials generalizing the z-score and random matrix theory constructs. New useful analytic estimates for all eigenvalues of a random matrix are described. SSA-based correlations also enjoy superior time stability. Technical aspects are given in four accompanying papers, including extensive analyses of time stability and the noise-reduction tests described in this short paper.
These offerings are designed to support our community in their ongoing learning journey and are shared for your benefit.