# Gauge 1: The Lipstick Test — locked pre-registration

Locked on **2026-10-01**, before downloading analysis data or computing primary
statistics. Machine-readable parameters: `config/gauge01.toml`.

NBER chronology checked 2026-10-01 at
https://www.nber.org/research/business-cycle-dating. The official search-indexed
page identifies February 2020 as the latest peak and April 2020 as the latest
trough. Direct page access returned HTTP 403; the official indexed page was
available. No later dated peak is included. The nine peak/trough pairs are:

| Peak | Trough |
|---|---|
| 1960-04 | 1961-02 |
| 1969-12 | 1970-11 |
| 1973-11 | 1975-03 |
| 1980-01 | 1980-07 |
| 1981-07 | 1982-11 |
| 1990-07 | 1991-03 |
| 2001-03 | 2001-11 |
| 2007-12 | 2009-06 |
| 2020-02 | 2020-04 |

## Data and definitions

BEA NIUnderlyingDetail, monthly, seasonally adjusted, all years: U20403 (quantity
indexes, Table 2.4.3U), U20404 (price indexes, Table 2.4.4U), U20405 (current
dollars, Table 2.4.5U). Validate table names and select exactly one normalized
description per table: Cosmetic/perfumes/bath/nail preparations and implements;
Clothing and footwear; Jewelry and watches; Personal consumption expenditures;
Goods. Save line numbers/codes; ambiguous or absent lines stop the pipeline.
This tests the grouped small-luxury category, not fragrance separately.

Also fetch FRED USREC, CUUR0000SEGB02 (NSA cosmetics CPI, begins December 1977),
CPIAPPSL (apparel CPI), UNRATE and their metadata. Monthly data begin in 1959-01
(report actual start). Gaps, nonpositive BEA quantity/price indexes and a latest
BEA month more than six months old are hard failures. Cosmetics nominal share
outside 0.1–2% is a warning. No imputation, interpolation or invented values.

For peak t0 and category quantity Q, charts use I(h)=100 Q(t0+h)/Q(t0),
h=-12,...,24. The trend g=[ln Q(t0)-ln Q(s)]/(t0-s), where
s=max(t0-36 months, first observation), with actual months in the denominator.
A(h)=100[ln Q(t0+h)-ln Q(t0)-g*h], h=0,...,24. Gap G(h) is cosmetics A(h)
minus clothing A(h). The primary statistic is mean G(12) across the nine peaks.

## Registered tests

- **H1 primary:** mean cosmetics-minus-clothing A(12)>0; one-sided placebo test.
- **H1b:** slope of (y_cosmetics-y_clothing) on y_total<0 with intercept;
  OLS/Newey–West HAC, four lags. Quarter level is the mean of all three monthly
  quantity observations, y=100 delta ln(quarter level); all complete quarters.
- **H1c:** mean cosmetics-minus-jewelry A(12)>0; one-sided placebo test.

Placebo candidates require a full 36-month trend, at least 24 future months, and
distance **>=24 months** from every real peak. Compute each candidate statistic
once. Each draw selects as many distinct candidates as the included real peaks,
without replacement. Draw 5,000 subsets with numpy.random.default_rng(42), or
enumerate all subsets if their count is <=5,000. P=(1+number of draw means >=
observed)/(1+draw count). Repeated subsets across draws are allowed.

Report greater-direction binomial sign tests with p0=.5 and a 10,000-draw
bootstrap 95% interval across peaks for H1/H1c, labeled **rough (n=9)**, as
supplementary summaries, not additional primary tests. Benjamini–Hochberg across
H1, H1b, H1c at FDR .05; report raw/adjusted p, effect sizes and 95% intervals.
Also estimate cosmetics, clothing and jewelry own betas on total growth with
intercept and four HAC lags. No correlations of trending levels.

Price channel is descriptive: apply the same abnormal-path formula to price
indexes, cosmetics minus total at horizon 12 for each peak, and correlate YoY
BEA cosmetics price growth with BLS cosmetics CPI growth from 1979 onward.

## Robustness, reported only

Recompute H1 placebo p, sign count and bootstrap interval (draw size equals peaks
used), excluding 2020, excluding July 1981 (double dip), excluding 1960; horizons
6/18/24; trends 24/48 months; raw gaps (g=0); goods as comparison; pre-2000 and
2000-on peaks separately. Fake dates retain exclusion of **all real peaks**, full
variant trend window and at least 24 future months. Intensity regression:
monthly delta(100 nominal cosmetics/nominal total) on monthly delta UNRATE,
intercept, HAC 12 lags. No robustness variant is promoted to primary.

## Presentation and index reading

Five final SVG/PNG charts: nine recessions with independent axes; gap scoreboard
and placebo distribution; own betas/HAC intervals; long-run nominal share; tilt.
Okabe–Ito palette, white background, USREC grey shading and source notes.
Exploration: log quantities, nominal share, cosmetics BEA/BLS YoY prices.

Small-luxury tilt = cosmetics YoY percent quantity growth minus clothing YoY
percent quantity growth, in percentage points. Report latest full-history
percentile, mean over USREC=1 months and last 120 months of history. Site reading
id small-luxury-tilt, title Small-luxury tilt, label small luxuries, heart layer,
pp unit. Direction compares latest with closest one-year-earlier observation:
z=change/SD of all historical one-year changes; |z|<.5 steady, otherwise sign.
Interpretation is generated directly from numerical results.

Reports include primary first, every registered test and BH p, all peak G(12),
own betas, robustness, price channel, line identities, vintages, QA, caveats with
evidence, deviations and chart list. Preserve immutable raw bytes and sanitized
metadata. Confirmatory stages require locked=true and the locked files committed
before the first confirmatory run. Stage hashes establish idempotence; reruns from
cached raw must reproduce results JSON excluding timestamps.

## Deviations

- **2026-10-01 — Acquisition compatibility, before confirmatory statistics.**
  Use BEA's working `/api/data/` endpoint and validate U20405 by its official
  Table 2.4.5U personal-consumption-expenditures title plus the existing units
  checks; the authentic catalog omits “current dollars” from that title. Accept
  only two verified source-label aliases: “Cosmetic / perfumes / bath / nail
  preparations and implements” for the registered cosmetics label (line 139;
  DCOSRA/DCOSRG/DCOSRC), and “Jewelry and watches (part of 119)” for the registered
  jewelry label (line 63; DJRYRA/DJRYRG/DJRYRC). Preserve the original API labels.
  Literal prompt labels fail exact matching against the official responses.
  The intended product lines, fixed config, hypotheses, horizons and test
  procedures are unchanged.
