Skip to content
Canli Capital

Measurements

Ledoit wolf effective sample

Ledoit wolf effective sample is one of 21 measurements this engine publishes in full. It is regenerated from a real run rather than transcribed, and it is shown here with its own claim boundary so the number and its limits arrive together.

The measurement

Any non conservative
yes
Schema
canli.alphac-ledoit-wolf-effective-sample.v1

Baskets

Nested beyond what this page renders. Read it in the artifact.

Non conservative halflives

  • 21
  • 63
  • 126
  • 252
  • 504
  • 720

Production setting

Conservative
no
Delta star as coded
0.081609
Delta star reestimated on effective window
0.147988
Delta star rescaled to effective T
0.114219
Effective sample rows
514.437641
Halflife bars
720
Inverse vol weights are invariant
yes
Max ex ante vol shift
0.038344
Mean ex ante vol shift
0.003608
Median ex ante vol shift
0.002617
Rescale clipped at one
no
Rows passed in production
720
Understatement factor pure T
1.399587

The approximation. ledoit_wolf_cc computes delta* = clip((pi - rho)/(gamma*T), 0, 1) with T = the rows PASSED. Production passes the full unweighted 720-row window while the matrix being shrunk is the EWMA covariance. Documented in the function's own docstring as the closest admissible sample.

The channel the error travels. NOT position sizing. The constant-correlation target shares S's diagonal, so this shrinkage cannot change a variance and inverse-vol weights — what the production `rank` allocator sizes on — are mathematically invariant to it. The error reaches the book only through quantities that use the off-diagonal: the overlay's ex-ante vol sqrt(w' Sigma w), which scales the whole book, and the MVO allocator, which is not the production default.

Verdict. NOT CONSERVATIVE AT ANY HALFLIFE, BUT IMMATERIAL AT THE PRODUCTION SETTING (0.36% mean ex-ante vol error). It becomes MATERIAL at a short halflife: 4.46% mean and 38.1% worst case at 21 bars. Fix it BEFORE shortening the covariance halflife, not after.

What this does not do. It does not change the live path. Production runs the legacy 720-bar covariance halflife, and any change to it is gated by the live-change ceremony and is the owner's. This measures what a future halflife decision would be walking into.

Why it is never conservative. The documented reasoning — that the unweighted window is the closest admissible sample — assumed the EWMA's effective sample could EXCEED the window. It cannot: ewma_cov is windowed at cov_window_bars, so its effective sample is bounded by the window at every halflife. Measured here it peaks at 635 rows against a T of 720 and falls to 514 at the production setting, so delta* is understated everywhere rather than only at short halflives.

Why it was conservative. A larger true T means a smaller true delta*, so using 720 when the estimator's effective sample EXCEEDS 720 over-shrinks slightly — and over-shrinking is the safe direction for a covariance estimate. The concern is the opposite case.

Worst case in ladder

Conservative
no
Delta star as coded
0.018308
Delta star reestimated on effective window
0.148516
Delta star rescaled to effective T
0.217521
Effective sample rows
60.598693
Halflife bars
21
Inverse vol weights are invariant
yes
Max ex ante vol shift
0.380773
Mean ex ante vol shift
0.044601
Median ex ante vol shift
0.031196
Rescale clipped at one
no
Rows passed in production
720
Understatement factor pure T
11.881444

Check it yourself

Every figure above is read from /glassbox/ledoit_wolf_effective_sample.json, the artifact the engine wrote. Nothing on this page is typed by hand: the page is generated from that file, so a figure that moves in the artifact moves here and a figure that is not in the artifact cannot appear here at all.

All 21 measurements · The glass box · The research library