The Squeeze Kernel Covariance Estimator: Dual-Timescale Tracking with Adaptive Shrinkage

<div> We introduce the Squeeze Kernel covariance estimator, a streaming, PSD-by-construction </div> <div> method for panels of daily financial returns that combines three complementary components: </div> <div> (i) a dual-timescale EWMA separating fast volatility from slow correlation dynamics, (ii) a </div> <div> smooth kernel weight wt = d¯2t /(d¯2t + κ), motivated by the Fisher information of the Gaussian </div> <div> likelihood, that filters uninformative calm-day observations, and (iii) an adaptive equicorre- </div> <div> lation shrinkage rule αt = max(0, n/(2St )−δ) that automatically calibrates regularisation to </div> <div> the concentration ratio using the estimator’s own effective sample size—requiring no cross- </div> <div> validation or tuning. We show that the complete update is a natural gradient ascent step </div> <div> on the Gaussian log-likelihood surface, with the kernel weight acting as an adaptive learn- </div> <div> ing rate. On a large-panel out-of-sample benchmark—n=100 instruments drawn from the </div> <div> S&amp;P 500 universe, 7,588 one-step evaluations, 1996–2026—the Squeeze Kernel ties DCC on </div> <div> Gaussian NLL (∆ = 0.49, Diebold–Mariano p = 0.26, within sampling noise) and beats </div> <div> every other tested method, including RMT denoising by 5.7 NLL units, the modified Ger- </div> <div> ber statistic by 6.7, and every linear and nonlinear shrinkage estimator by more than 16 </div> <div> (p &lt; 10−10 ). The advantage widens monotonically with n: at n=200 the Squeeze Kernel </div> <div> beats DCC by 21 NLL (p &lt; 10−7 ), RMT by 16, and the Gerber statistic by 23; at n=300 </div> <div> DCC, RMT, and Gerber are all numerically infeasible while the Squeeze Kernel still pro- </div> <div> duces a stable, finite forecast and dominates the surviving linear-shrinkage baselines by 280+ </div> <div> NLL units. Across a 50× sweep of its characteristic parameter, forecast quality varies by </div> <div> 16 NLL units, against a four- to five-orders-of-magnitude blow-up for DCC at the upper </div> <div> end of its news-coefficient sweep and roughly twice the variation for the Gerber statistic </div> <div> across its threshold sweep—establishing the method as the most robust of the tested estima- </div> <div> tors to its characteristic-parameter choice. Formal Diebold–Mariano tests, VaR backtesting, </div> <div> and Monte Carlo simulations with Gaussian and Student-t innovations support all claims. </div> <div> Missing observations from listing, delisting, and trading halts are handled natively through </div> <div> masked updates, so the panel used by the estimator never needs imputation. </div>