What does Sizing Systems and Return Minimisation conclude, in numbers, about “Shrinkage Estimates of Brand Offsets: How Far to Trust Two Measured Garments”?
Answer: Chapter 4 uses empirical Bayes shrinkage to cut the mean squared error of brand offsets to 0.808 of the raw means, and shows that with one garment measured per brand the "largest-running" brand on a page is overstated by 1.3922 cm on average.
Scope of the numbers: Its brands, shops, prices and body dimensions are illustrative; all calculations are generated by its programs; the book is not individual business advice, deals only with the measurement of garment sizes, and makes no health judgement. Marked boxes separate published evidence, runnable Python calculations (21 programs), classical sources and the author's hypotheses.
Data table
The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.808
| estimator | MSE (cm^2) | RMSE (cm) | ratio to raw |
| raw | 0.6937 | 0.8329 | 1.000 |
| eb | 0.5604 | 0.7486 | 0.808 |
| oracle | 0.5527 | 0.7434 | 0.797 |
| n_b | RMSE raw | RMSE eb | weight k at true tau |
| 1 | 1.2738 | 1.0815 | 0.709 |
| 2 | 0.9061 | 0.8322 | 0.830 |
| 3 | 0.7309 | 0.6908 | 0.880 |
| 5 | 0.5740 | 0.5542 | 0.924 |
| 10 | 0.4015 | 0.3946 | 0.961 |
| mean estimated tau: 1.967 (true 2.0) |
Show the full program
# Chapter 4. p41: Shrinking brand offsets. J = 30 brands, true offsets
# u_b ~ N(0, 2.0^2) cm; the store measures n_b garments of label M per
# brand (n_b cycles 1, 2, 3, 5, 10), each with sd 1.28 cm (style 0.8 and
# tape 1.0 combined). Raw mean versus empirical-Bayes shrinkage
# u_hat = k_b * ybar_b, k_b = t2 / (t2 + s2 / n_b),
# where t2 is estimated by moments, max(0, mean(ybar^2 - s2/n_b)).
# Mean squared error over 2,000 fixed-seed replications. Illustrative.
import math
import random
J, TAU, SD, REPS = 30, 2.0, 1.28, 2000
NS = [(1, 2, 3, 5, 10)[j % 5] for j in range(J)]
rng = random.Random(20261112)
err = {"raw": [], "eb": [], "oracle": []}
by_n = {n: {"raw": [], "eb": []} for n in (1, 2, 3, 5, 10)}
t2s = []
for _ in range(REPS):
u = [rng.gauss(0.0, TAU) for _ in range(J)]
yb = [u[j] + rng.gauss(0.0, SD / math.sqrt(NS[j])) for j in range(J)]
t2 = max(0.0, math.fsum(yb[j] ** 2 - SD ** 2 / NS[j]
for j in range(J)) / J)
t2s.append(t2)
for j in range(J):
v = SD ** 2 / NS[j]
eb = t2 / (t2 + v) * yb[j] if t2 > 0 else 0.0
orc = TAU ** 2 / (TAU ** 2 + v) * yb[j]
err["raw"].append((yb[j] - u[j]) ** 2)
err["eb"].append((eb - u[j]) ** 2)
err["oracle"].append((orc - u[j]) ** 2)
by_n[NS[j]]["raw"].append((yb[j] - u[j]) ** 2)
by_n[NS[j]]["eb"].append((eb - u[j]) ** 2)
mse = {k: math.fsum(v) / len(v) for k, v in err.items()}
print("estimator MSE (cm^2) RMSE (cm) ratio to raw")
for k in ("raw", "eb", "oracle"):
print("%-9s %.4f %.4f %.3f"
% (k, mse[k], math.sqrt(mse[k]), mse[k] / mse["raw"]))
print("n_b RMSE raw RMSE eb weight k at true tau")
for n in (1, 2, 3, 5, 10):
r = math.sqrt(math.fsum(by_n[n]["raw"]) / len(by_n[n]["raw"]))
e = math.sqrt(math.fsum(by_n[n]["eb"]) / len(by_n[n]["eb"]))
k = TAU ** 2 / (TAU ** 2 + SD ** 2 / n)
print("%3d %.4f %.4f %.3f" % (n, r, e, k))
print("mean estimated tau: %.3f (true %.1f)"
% (math.fsum(math.sqrt(t) for t in t2s) / REPS, TAU))Single source and citation
Fan, Y. (2026). Anthropometric Principal-Component Sizing, Vanity-Size Drift and Cross-Border Return-Rate Minimisation (Version 1.0.0), Chapter 4: Shrinkage Estimates of Brand Offsets: How Far to Trust Two Measured Garments. 5D Library. DOI pending
DOI pending · monograph page goes live after publication · 5DL-B-215 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026215,
author = {Fan, Yafu},
title = {Anthropometric Principal-Component Sizing, Vanity-Size Drift and Cross-Border Return-Rate Minimisation},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}Other chapter answers in this book
- Chapter 1 · A Label Is Not a Measurement: Brand Dispersion of the Same "M" as a Random Effect
- Chapter 2 · A Finer Size Chart or a Brand Residual: the Error Budget of Five Sizing Policies
- Chapter 3 · A Two-Dimensional Principal-Component Grid of Height and Girth: What a One-Axis Label Throws Away
- Chapter 5 · Vanity-Size Drift: Tracking a Moving Brand Offset with a Kalman Filter
- Chapter 6 · The Expected Cost of Returns: Sizing Up, Cross-Border Freight and Ordering Several Sizes
- Chapter 7 · Sources East and West: the Palace Tailors of the Zhou li and Quetelet's Average Man
- Chapter 8 · Closing the Loop: a Buyer's Direct-Message Size Recommender, a Return-Reason Code Table and a Pre-Registered Test