methodology
Peer Shrinkage
A GNG covariance step that pulls the estimated correlations between stocks toward their peer groups, so a short or noisy price history carries less weight.
Peer shrinkage is a covariance-cleaning step that nudges the estimated correlations between stocks toward what their business similarity suggests. Stocks that GNG's peer engine identifies as close peers are assumed to share a meaningful correlation, so a short or noisy return history does not get the final word. On GNG it sits on top of the base covariance estimate and is off by default.
The method builds a target correlation matrix from peer groups. Stocks linked as peers inside your universe form a cluster; within a cluster, every pair gets the cluster's average peer-similarity score as its target correlation (held between 0.10 and 0.85), and pairs in different clusters get a target of 0. The shrunk correlation is then a weighted average: at the default intensity of 0.3, 70% of the sample correlation plus 30% of the peer target. Each stock's own volatility is left unchanged.
Peer shrinkage helps most for stocks with short histories, such as recent listings, and in small samples where two close competitors happen to show a low measured correlation. It encodes a sensible prior, that companies in the same business tend to move together, without ignoring the data. The trade-off is that the target depends on the quality of the peer groups: if two companies are linked as peers but trade very differently, the step adds bias, which is why the intensity is kept modest.
Example with illustrative numbers: two regional banks show a sample correlation of only 0.20 over a short window, while their peer group's average similarity sets a target of 0.60. At an intensity of 0.3, the shrunk correlation is 70% of 0.20 plus 30% of 0.60, or 0.32. A bank and a software company in different peer groups with a sample correlation of 0.50 would move to 70% of 0.50, or 0.35.
On GNG, peer shrinkage is an optional step in the Portfolio Optimizer's covariance estimate. It uses each stock's top 25 peers from GNG's peer engine, the same peer data the optimizer's Tax-loss harvest mode uses to suggest replacement holdings. It is off by default, so standard runs rely on Ledoit-Wolf shrinkage, plus Marchenko-Pastur denoising when you switch that on.
Ledoit-Wolf Shrinkage is the base estimate this step adjusts, and Correlation explains the quantity being shrunk.
Related terms
More methodology terms
Get full access to GNG Research
Create a free account to access portfolio tracking, advanced tools, and more.