% paper/appendices/data.tex — provenance of every raw source (data/raw/README.md, MANIFEST.json)
\section{Data appendix}\label{app:data}
\begingroup\sloppy

All raw files were downloaded on 2026-10-27 by \texttt{scripts/01\_download\_data.py}, run under
\texttt{python3 -I} so that no code is loaded from the download directories, and processed by
\texttt{scripts/02\_build\_calibration\_targets.py}, which reads them only through pandas, openpyxl and xlrd.
\texttt{data/raw/MANIFEST.json} logs, for each of the 36 attempted downloads (35 successful, one HTTP 404), the
URL, UTC timestamp, HTTP status, content type, byte count, SHA-256 digest, number of rows and first and last
observation, together with a \texttt{not\_downloaded} block giving the reason for every item that was sought but
not obtained. Re-running the download script regenerates the manifest; a changed digest flags a revised vintage.

\paragraph{FRED.} Twenty-four series were requested from
\url{https://fred.stlouisfed.org/graph/fredgraph.csv?id=<SERIES>&cosd=1900-01-01} (no key required) and 23 were
obtained: the BLS labour-share index (\texttt{PRS85006173}, 1947:Q1--2026:Q2), the PWT labour-compensation share
(\texttt{LABSHPUSA156NRUG}, 1950--2023), BEA income shares (\texttt{A4002E1A156NBEA}, \texttt{W270RE1A156NBEA}),
real GDP, consumption and investment (\texttt{GDPC1}, \texttt{PCECC96}, \texttt{GPDIC1}, \texttt{A191RL1Q225SBEA}),
interest rates and prices (\texttt{DFII10}, \texttt{DGS10}, \texttt{DFF}, \texttt{TB3MS}, \texttt{CPIAUCSL}), BEA
fixed-asset stocks and depreciation (\texttt{K1TTOTL1ES000}, \texttt{M1TTOTL1ES000}, 1925--2024), the S\&P 500
level (\texttt{SP500}, ten-year licence window only, 2016-10-28 to 2026-10-26), ICE BofA option-adjusted spreads
(\texttt{BAMLC0A0CM}, \texttt{BAMLH0A0HYM2}), the SLOOS tightening series (\texttt{DRTSCILM}), the H.8 bank credit,
assets and residual series (\texttt{TOTBKCR}, \texttt{TLAACBW027SBOG}, \texttt{RALACBW027SBOG}) and the Z.1 NYSE
composite level (\texttt{BOGZ1FL073164003Q}). The Wilshire 5000 (\texttt{WILL5000INDFC}) returned HTTP 404 and is no
longer distributed by FRED. Two restrictions matter for the paper: since April 2026 FRED serves only three years of
the ICE BofA indices (``Copyrighted: Pre-Approval Required''), so the 2008 and 2020 spread peaks are not in our files
and appear in the ledger only as readings of the chart descriptions in the Federal Reserve's November 2020
\emph{Financial Stability Report}; and the S\&P 500 level is limited to ten years, which is why long-run equity
moments come from the Shiller file.

\paragraph{Distributional Financial Accounts.} \url{https://www.federalreserve.gov/releases/z1/dataviz/download/zips/dfa.zip}
(895,897 bytes, archive dated 2026-09-15); 26 CSV files and three data dictionaries were extracted, covering
1989:Q3--2026:Q2 for the groups TopPt1, RemainingTop1, Next9, Next40 and Bottom50. Only
\texttt{dfa-networth-shares.csv} is used.

\paragraph{Fernald TFP.} \url{https://www.frbsf.org/wp-content/uploads/quarterly_tfp.xlsx} (252,789 bytes, vintage
``Produced on October 01, 2026''), sheets \texttt{annual} (1947--2025) and \texttt{quarterly} (1947:Q1--2026:Q2);
columns \texttt{alpha}, \texttt{dtfp} and \texttt{dtfp\_util}, in 100$\times$ (annual) or 400$\times$ (quarterly)
log changes. The author asks that FRBSF Working Paper 2012-19 be cited.

\paragraph{Shiller data.} The current \texttt{ie\_data.xls} linked from \url{https://shillerdata.com/} (1,675,264
bytes; monthly 1871-01 to 2026-10, with the October 2026 price being the 2 October close and the last two CPI
values estimated by the file's own note). Columns used: price, dividend, CPI and the ten-year yield; annual real
total returns compound monthly price plus one twelfth of the dividend. The legacy Yale copy ends in 2023-09 and was
not used.

\paragraph{SEC filings.} Capital expenditure is the XBRL concept \texttt{PaymentsToAcquirePropertyPlantAndEquipment} (namespace \texttt{us-gaap}) from
the SEC XBRL company-concept API (\url{https://data.sec.gov/api/xbrl/companyconcept/CIK<cik>/us-gaap/PaymentsToAcquirePropertyPlantAndEquipment.json},
User-Agent header required) for Microsoft, Alphabet, Amazon and Meta, supplemented by Amazon's Q4 2025 earnings
release (purchases of property and equipment \$131,819 million, twelve months 2025) and Meta's FY2025 exhibit 99.1
(\$72.22 billion including finance-lease principal). Microsoft's calendar-2025 figure is derived as the fiscal-2025
total minus the three fiscal quarters reported in 10-Qs, plus the two subsequent quarters
($16.745+17.079+19.394+29.876=83.094$ billion); it is the only derived component of the \$376.1 billion total.

\paragraph{EIA and open working papers.} Electric Power Monthly Table 5.6.A
(\url{https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_6_a}) is saved as HTML; 2026 values
are preliminary. The Flod\'en--Lind\'e IIES Seminar Paper 654 (1998) and Gorton--Metrick NBER WP 15223 (2009) PDFs
are stored in \texttt{data/raw/papers/} and were read for Table~4 and for the haircut-index sentence respectively.

\paragraph{Not obtained.} WID.world US wealth shares (client-side charts, no CSV endpoint), SCF 2022 microdata (the
bulletin was read for median and mean net worth but publishes no Gini), the Census BTOS data page (JavaScript only;
values taken from the Census story of 26 May 2026), the FDIC Quarterly Banking Profile equity ratio (not on the
summary page; the H.8 book ratio is used instead), the Cooley--Prescott (1995) chapter text (HTTP 403), the
Storesletten--Telmer--Yaron full text (HTTP 404; abstract only) and the OpenAI pricing page (HTTP 403; the
developer-docs mirror was read). Each omission is recorded in the manifest and marked \textsc{not obtained} in the
ledger rather than filled from memory.
\endgroup