Unclear performance advantage of Kendall-based methods at q=0.2 for American stocks

Determine whether Kendall-based generalized correlation coefficient approaches—specifically Kendall correlation and Kendall ICVC—outperform Pearson-correlation–based Random Matrix Theory cleaning schemes (Rotationally Invariant Estimator, regularized RIE Γ, RIE + identity rescaling, eigenvalue clipping, and Pearson-based ICVC) in terms of out-of-sample portfolio risk for American equities when the dimension-to-sample ratio q = N/T equals 0.2, across the minimum-variance, omniscient, mean-reversion, and random long-short strategies (noting that Kendall-based methods appear superior for the random long-short case).

Background

The paper benchmarks Kendall-based generalized correlation matrices and Pearson-correlation cleaning schemes (RIE, RIE Γ, RIE + Id, Clipped, ICVC) for constructing Markowitz portfolios across several regions and sample regimes (q ∈ {0.5, 1, 2}). Kendall-based methods consistently show lower out-of-sample risk in many settings.

To probe the robustness of these findings as data availability increases (smaller q), the authors extend the analysis to q ≈ 0.2 for American stocks. While Kendall retains strong performance in some contexts, the authors explicitly note uncertainty about Kendall’s overall advantage at q=0.2, except for the random long-short strategy, signaling an unresolved comparative performance question at that sampling ratio.