Transparency is a feature. This page explains what covariance shrinkage does and why — in plain terms. The precise formulas and numeric conventions are documented in our internal methodology; the summary here is deliberate, not an omission.
The plain sample covariance is unbiased, but it gets noisy and unreliable when you only have a modest number of periods relative to the number of assets. In that regime it tends to overstate the biggest sources of variation and understate the smallest, and the resulting matrix can become unstable — nearly impossible to invert cleanly for tasks like portfolio optimization. Shrinkage fixes this by blending the noisy sample estimate with a simpler, more stable target, so the final estimate keeps most of the real structure while shedding much of the noise.
This is the well-known Ledoit-Wolf (2004) estimator, which pulls the sample covariance toward a structured target by a data-driven, optimal amount chosen from the data itself — you do not have to guess how much to shrink. The target used here is the scaled-identity target: a clean, well-behaved stand-in that treats every asset as having the same variance and no cross-correlations, sized to match the overall scale of the sample.
The blend is a weighted average of the two. The shrinkage intensity is bounded between fully trusting the sample covariance at one extreme and collapsing entirely onto the stable target at the other; in practice it lands somewhere in between, leaning more toward the target the noisier the sample looks. Read a higher intensity as "the sample estimate was noisy, so we relied more on the stable target," and a lower intensity as "the sample was trustworthy enough to mostly keep as-is."
A useful property: the shrinkage preserves total variance exactly. It does not change how much overall variability the estimate attributes to the assets — it only redistributes that variability across the different directions of risk, damping the extremes that the raw sample tends to exaggerate. The result is always a symmetric matrix.
