Global macro

Code: pipeline/tipsheet/compute/global_macro.py, publish/global_macro.py.

Bundles: macro/global/* (23 files, the largest about 89 KB). Each bundle’s meta carries its source, licence, unit, method and start dates.

Sources

• FRED, through the Data bank dataset fred_series (33 series, refreshed daily by the 06:00 fred update). For each series and date the latest fetch is used.

  What	FRED ids	Publisher	Starts
  Treasury yields	DGS3MO, DGS2, DGS10, DGS30	Federal Reserve H.15 / Treasury	1981, 1976, 1962, 1977
  Real yield, breakeven	DFII10, T10YIE	H.15 / Treasury	2003
  Fed balance sheet, TGA	WALCL, WDTGAL, WTREGEN	Federal Reserve H.4.1	2002-12-18
  Overnight reverse repo	RRPONTSYD	New York Fed	2003 (full-time facility from 2013-09-23)
  Broad dollar, USD/INR	DTWEXBGS, DEXINUS	Federal Reserve H.10	2006, 1973
  Brent, WTI	DCOILBRENTEU, DCOILWTICO	EIA	1987, 1986
  CPI, core CPI	CPIAUCNS, CPILFENS (CPIAUCSL kept as a check)	BLS	1913, 1957
  PCE, core PCE	PCEPI, PCEPILFE	BEA	1959
  Jobs	UNRATE, PAYEMS	BLS	1948, 1939
  Fed funds, M2	FEDFUNDS, M2SL	Federal Reserve H.15, H.6	1954, 1959
  Financial conditions	NFCI, STLFSI4	Chicago Fed, St Louis Fed	1971, 1993
  Equity risk (derived only)	SP500, VIXCLS	S&P Dow Jones Indices, Cboe	2016, 1990

• FPI flows, through the IndiaDataHub loader in compute/flows.py:

  • NSDL daily net investment in US dollars (equity, and the total of all instruments), from 2021.
  • RBI’s monthly portfolio investment in India (US$, balance-of-payments basis), from April 2000.

• IMF World Economic Outlook, through the Data bank dataset imf_datamapper: annual real GDP growth and average consumer-price inflation for the world, advanced and emerging groups, and nine economies.

Licences

• Levels are published only for public-domain series. US federal government works (Federal Reserve Board, Treasury, BLS, BEA, EIA) are public domain. FRED’s notes for every published series carry no third-party copyright; this was checked series by series through the FRED API on 2026-10-01. The Chicago Fed and St Louis Fed indices carry no restriction in FRED’s notes and are published with attribution.
• Oil: EIA’s reuse page says its data are public domain. EIA presents Brent and WTI spot as its own series with no third-party restriction.
• S&P 500 and VIX are copyrighted (FRED’s notes name S&P Dow Jones Indices and Cboe). Only weekly returns, drawdown and the VIX’s percentile are published, never a level.
• IMF: published with attribution. The IMF’s data terms allow reuse with attribution. The Data bank’s source note says “personal-research use only”, which is stricter than the IMF’s own terms. The bundle says so, and the owner should confirm.

Method

• Spreads: 10y-2y and 10y-3m are computed from the constant-maturity yields on the same day.
• Net liquidity: Fed total assets minus the Treasury General Account minus overnight reverse repos, all at Wednesday levels, in US$ billion.
  • It is a market shorthand, not an official measure.
  • A second column uses the week-average TGA (WTREGEN), the variant most often quoted. The two differ by a median of $9.5 bn.
  • Holiday Wednesdays: ON RRP is taken from the last value within 6 days, because the previous overnight operation is still outstanding.
  • Before the standing facility (2013-09-23): a Wednesday with no operation counts as zero.
• US CPI inflation: year on year from the not-seasonally-adjusted CPI-U, which is how BLS reports its headline. PCE uses BEA’s seasonally adjusted index, as BEA reports it.
  • All year-on-year rates are by calendar month, so a missing month gives a null rather than a shifted comparison.
• Weekly views: weeks end on Friday.
  • Each column takes its last value in the week.
  • The row is dated by the latest actual observation, so no row is dated after its data.
  • The H.4.1 bundle keeps its Wednesday dates.
• India link (descriptive only):
  • Weekly, from 2021: weekly FPI net investment (US$ million), weekly changes in net liquidity, the US 10-year (bp) and the broad dollar (%).
    • z-scores are taken against the trailing 156 weeks (minimum 52).
    • Rolling 52-week correlations (minimum 40) are between weekly FPI equity flows and each change.
  • Monthly, from 2003: RBI’s monthly FPI against 12-month changes, with z-scores against the trailing 60 months, and 36-month rolling correlations with one-month changes.
  • Nothing uses future data.
• IMF table: five years around the build year. Years from the build year on are flagged as projections. India’s figures are on a fiscal-year basis in the WEO.

Checks (2026-10-01, on real data)

• 10y-2y: our spread equals FRED’s T10Y2Y exactly on 12,576 of 12,579 days since 1976. The three exceptions (1990-11-21, 1991-01-29, 1995-11-29) are off by 0.01–0.02 points.

• 10y-3m: our spread matches T10Y3M on all but 2 of 11,188 days (1991-01-29 and 1999-10-01, up to 0.10 points).

• Breakeven: T10YIE equals DGS10 minus DFII10 on all 5,940 days.

• CPI levels: FRED’s CPIAUCNS matches the BLS public API (series CUUR0000SA0) on all 31 months it returned, January 2024 to August 2026.

  • Year-on-year rates computed from BLS’s own levels match ours for March to August 2026: 3.26, 3.81, 4.25, 3.53, 3.36 and 3.40%.

• CPI against BLS headline prints: our headline and core rates round to the published figures every time:

  Month	Headline	Core
  Jun 2022	9.1
  Sep 2022	8.2	6.6
  Sep 2023	3.7	4.1
  Dec 2024	2.9	3.2
  Jan 2025	3.0	3.3

  The comparison is stored in the bundle meta. Seasonally adjusted rates differ from these by up to 0.14 points over the last five years.

• Fed balance sheet: it peaks at $8,965 bn on 2022-04-13, which agrees with the published H.4.1 peak.

• Published files: every macro/global file is parsed as strict JSON (no NaN or Infinity) at publish time, and is under 150 KB.

Caveats

• October 2025 CPI does not exist. BLS published none because of the federal shutdown, so that month and October 2026 have no year-on-year rate.

• The correlations between FPI flows and global drivers are weak and unstable:

  Window	Range
  Weekly, rolling 52 weeks	−0.40 to +0.35
  Monthly, rolling 36 months	−0.58 to +0.34

  Full-period weekly correlations are between −0.02 and −0.09. The strongest monthly link is with the dollar (−0.33 over 246 months). Read these as description, not cause.

• NSDL dates FPI trades one day after the exchanges. Weekly sums absorb most of the lag.

• The H.10 dollar and rupee rates lag. They are published weekly on Mondays, so the latest week can be missing. The rupee here is a New York noon rate; use macro/india for RBI/FBIL rates.

• S&P 500 history on FRED covers only 10 years. Its drawdown is measured from the highest close since 2016.

• Revisions: NFCI and STLFSI4 are revised back through history every week, and payrolls are revised for two months after release.