🇬🇧 English · 🇪🇸 Español

Quarterly National Accounts

puremacro.fetch builds a quarterly national accounts panel from the OECD's SDMX service and — the part that usually gets hand-rolled in every notebook that touches one — does the work that comes after the fetch: putting countries on a common price reference year, scoring the accounting identities, and decomposing growth into contributions.

Nothing on this page except qna_panel, qna_labor and qna_countries touches the network. A live panel and a frozen CSV of one behave identically.

from puremacro.fetch import (
    qna_panel, qna_countries, qna_rebase, qna_identity, qna_contributions,
)

panel = qna_panel(["USA", "JPN", "DEU"], start="1995",
                  output=True, income=True, real=True)

Notebook 40 (notebooks/40_quarterly_national_accounts.py) works through all of it on a frozen six-country panel.

The three products of a build

qna_panel returns current-price levels, implicit deflators and — with real=True — volume measures, tied by

nominal = real x deflator / 100

component by component. Every transform below preserves that identity exactly.

Three approaches to GDP

The OECD publishes three separate measurements of the same quantity, from three separate source systems. qna_panel reaches all of them:

flag	approach	identity	registry
(always)	expenditure	Y = C + G + I + X - M	QNA_COMPONENTS
output=True	output (production)	Y = Σ VA_j + (D21 - D31) + YA1	QNA_ACTIVITIES
income=True	income	Y = D1 + B2A3G + (D2 - D3)	QNA_INCOME

Two caveats the registries encode:

• Two of the fifteen value-added columns are memo items, not addends. va_mfg sits inside va_ind, and va_services aggregates seven columns already listed. QNA_VA_ADDITIVE names the ten that sum to the total; QNA_VA_MEMO names the two that do not. Adding every va_* column double-counts about a third of the economy.
• Income columns exist only in current prices. There is no volume measure of compensation of employees, so they carry no deflator and no _real column even under real=True.

Coverage is not uniform. Of the OECD's 49 reference areas, 46 appear in the by-activity output flow — the United States is absent from it entirely, since the US industry accounts are a separate BEA release — and 40 in the income flow. A country that does not publish an approach reads NaN.

assets=True and durability=True — two splits of the same totals

Neither adds an approach; each opens up a total the expenditure block already carries, and both are joined with their own deflators.

flag	splits	into	registry
assets=True	inv (gross fixed capital formation)	inv_equip, inv_struct, inv_dwell, inv_ipp	QNA_ASSETS
durability=True	household consumption	cons_dur, cons_semidur, cons_nondur, cons_serv	QNA_DURABILITY

Two things to know before using them:

• The durability split is a different institutional sector. The headline cons_hh is S1M — households plus NPISH. The durability columns are S14, households only. They will not sum to cons_hh, and the gap is NPISH, not an error.
• Several countries publish these splits unadjusted while their headline aggregates are adjusted — Mexico, Japan and Turkey among them. qna_meta reports this in sa_detail, separately from sa, and sa="x13" adjusts them here. This is the reason sa_detail exists as its own column.

Durables are the part the national accounts book as consumption but that macro theory treats as household capital (Cooley & Prescott 1995; Gomme & Rupert 2007), which is why the split is worth having rather than a curiosity.

labor=True — the labour input behind those flows

The same accounts measure the labour that produced the output: employment and hours worked, each split into employees and the self-employed, on the domestic concept — labour in resident production units, which is the concept GDP is measured on, so gdp_real / hours is a productivity measure rather than two different populations divided by each other.

column	what it is
emp, emp_employees, emp_selfemp	persons, thousands
hours, hours_employees, hours_selfemp	hours worked, millions

Registered in QNA_LABOR, with the scales in QNA_LABOR_UNITS. The block carries no price, so it gets no deflator and no _real column — real=True does not change that.

The ISIC breakdown: labor_activities=True

qna_panel(..., labor=True, labor_activities=True) — or qna_labor(..., activities=True) — returns every stem again per ISIC activity, so hours gains hours_agri and hours_public and the block goes from six columns to eighteen. The two activities are registered in QNA_LABOR_ACTIVITIES: agriculture (A), where most of the labour is self-employed and most of the year-to-year output is the weather, and public administration, education and health (OTQ), whose value added the SNA defines as compensation plus consumption of fixed capital, so its measured productivity growth is near zero by construction rather than by finding.

Subtracting the two is the whole point:

market_hours = p["hours"] - p["hours_agri"] - p["hours_public"]

That turns a whole-economy Y/H into the market sector — the concept the United States publishes as its nonfarm business sector, and the only basis on which it can be compared with anyone else. The whole economy (_T) is always requested alongside the parts, because a part without its whole is not usable, and every activity of one unit of measure is resolved as a single seasonal family, since a source-adjusted total minus a raw part is not a subtraction of anything. It costs nothing extra — the same request, three activities instead of one — and 34 reference areas publish hours for all three.

labor_activities requires labor=True, and raises if it is off rather than returning a panel without the columns you asked for.

For the annual counterpart — which is the only place the United States and Japan have a national-accounts denominator at all, because the quarterly flow returns zero rows for both — see ana_by_activity in puremacro.fetch.oecd_ana_activity.

Two ratios are the point of the split:

• emp_selfemp / emp — the share of the workforce whose labour income the accounts book inside surplus_mixed rather than comp_emp. This is the missing input for the labour-share correction at the end of this page.
• hours / emp — average hours per worker, the margin that carries a German recession (2008–10: employment −0.3%, hours −3.7%) where a Spanish one is carried by heads instead (−9.9% against −9.8%).

Coverage is the raggedest of any block on this page, and worth checking before dividing one of these columns by another:

• 38 of the 49 reference areas publish heads and 34 publish hours. The United States, Japan, Argentina, Brazil, China, Colombia, Indonesia, India, Saudi Arabia and Turkey publish nothing in this flow at all — at any level of activity, so no choice of key recovers them. Canada publishes hours without heads; Australia, Switzerland, Korea, Russia and South Africa publish heads without hours.
• Chile reports hours per week and Costa Rica at an annual rate, both labelled exactly like everyone else's quarterly figure — a Chilean quarter reads ~41 hours per worker and a Costa Rican one ~2,157, against ~430 for a normal one. Both are put back on a quarterly basis, and qna_meta's hours_scale records the factor (13.0, 0.25, or 1.0 for untouched). Detection is by the level: a country is only rescaled if its median hours / emp is outside 150–1000 and a candidate factor lands it inside 250–700. Every observation the 31 same-basis countries have ever published sits in 304–572, so the band cannot fire on a country that merely works short weeks. Canada publishes hours but no heads, so nothing can be checked and its hours are left alone. hours_rescale=False returns the numbers exactly as the OECD sends them.
• Ten reference areas publish the block with no adjusted variant at all, several of them while publishing adjusted aggregates. sa="x13" adjusts them here, and the meta column sa_labor reports who did the adjusting, separately from sa (headline aggregates) and sa_detail (asset and durability splits).
• Korea adjusts the employment total at source but not its two components. Heads and hours are each resolved as one family, so Korea falls back to the raw series for all three and the decomposition still adds up; the price is that sa_labor reads none for Korea under the default. Taking the adjusted total with raw parts would put a 1.1pp seasonal artefact straight into emp_selfemp / emp.

qna_labor() — the labour block without the accounts

qna_panel(..., labor=True) joins the labour block onto the national accounts, which means it also downloads the expenditure block and, under sa="x13", seasonally adjusts it. When the labour series are all you want, that is three extra SDMX round-trips per chunk of ten countries and a lot of X-13 you did not ask for:

from puremacro.fetch import qna_labor

lab = qna_labor(["DEU", "MEX", "KOR"], start="1995", sa="x13")

It returns a long frame — code, date, variable, value, sa_source — carrying the same columns (six by default, eighteen with activities=True), in the same units, with the same hours correction and the same family-level seasonal adjustment. Two differences from the joined route are deliberate:

• sa_source is per series, not per country. A country adjusted at source for heads but not for hours gets an honest label on each, where qna_meta's sa_labor can only report mixed for the block as a whole.
• It does not depend on the expenditure block at all. In the joined route the labour rows are filtered to the countries the expenditure flow returned, so a country publishing labour but no expenditure is dropped and a failed expenditure request takes the labour rows with it. That is the right behaviour when you asked for a national-accounts panel and the wrong one when you asked for employment.

This is what build_panel.build_all uses to fill the labour gaps its local workbook does not cover.

qna_countries() — ask the source what it carries

Queries the SDMX availability endpoint for a dataflow's reference areas, so a panel covers what the source actually supports rather than a hand-typed list that goes stale the next time the OECD onboards a country.

Country groupings (OECD, EA20, G7, …) are dropped by default — see QNA_AGGREGATES — because a panel that silently mixes OECD in with USA double-counts every aggregate it touches. Falls back to a frozen list when the endpoint is unreachable: it never raises and never returns empty.

qna_rebase(panel, year) — one price reference year

The OECD references each country's volumes to that country's base year: 2017 for the United States, 2018 for Mexico, 2020 for Spain. Raw deflator columns are therefore indices against different years, and comparing them across countries as levels compares nothing.

qna_rebase rescales each country's deflators to the requested year and its volumes by the same factor, so nominal = real x deflator / 100 still holds exactly. This is a re-referencing — one scalar per country, leaving every growth rate and every chain link untouched — not a re-basing, which a published chain-linked volume does not permit at all.

qna_identity(panel) — score the identities, separately

Returns one row per country. The residuals are reported apart from each other because they are different things:

column	what it is
nominal_*	statistical discrepancy in current prices. Often forced to zero — a presentation choice. Where it is not zero, independent seasonal adjustment of each series is a large part of the reason, visible as a sign-alternating residual.
real_*	chain-linking gap in volume terms. Not a data error: chain-linked volumes are not additive away from the reference year, and the gap widens with distance from it.
output_*, income_*	each approach scored against its own flow's GDP
crossflow_output, crossflow_income	the disagreement between flows

That last row matters. The OECD publishes GDP separately in each flow and the figures do not always agree — Japan's output-flow GDP differs from its expenditure-flow GDP by up to 0.61%, Germany's income-flow GDP by up to 1.77%. Scoring an approach against a different flow's GDP would charge that disagreement to the approach's own components, so APPROACH_GDP carries gdp_output and gdp_income as their own columns and the disagreement gets its own report. Across the full 49-country panel the output identity closes to a median 0.001% of GDP and the income identity to 0.003%.

The income residual is the one with a name and a literature: the GDP–GDI statistical discrepancy, which for the United States runs to ±2% of GDP and is informative about the business cycle in its own right (Nalewaik 2010). Most European offices force it to zero.

qna_contributions(panel) — decompose growth

Real GDP growth, split into what each component contributed, with previous-period nominal weights:

g_t = Σ_i ω_{i,t-1} g_{i,t},    ω_{i,t-1} = (P_i Q_i)_{t-1} / (P Q)_{t-1}

This is the calculation chain-linking requires, and one that needs all three products of a build at once: volumes for the growth rate, current prices for the weight. Imports enter negatively. annualise=True rescales the parts along with the aggregate. What the weights do not span is returned as an explicit residual column rather than spread over the components.

A note on the labour share

comp_emp / gdp is the unadjusted labour share, and the adjective is load-bearing: the income of the self-employed is not in D1 at all but inside surplus_mixed. A country with a large self-employed sector reads low for reasons that have nothing to do with how its employees are paid. See Gollin (2002).

The correction needs a split of the workforce, which is what labor=True supplies. With all three blocks on, the panel carries every column puremacro.labor_share.gollin_adjusted_ls asks for:

panel = qna_panel(["ESP", "ITA", "DEU"], start="1995",
                  output=True, income=True, labor=True)

gollin_adjusted_ls wants	panel column	from
compensation_employees	comp_emp	income=True
mixed_income	surplus_mixed	income=True
value_added	va_total	output=True
employment_employees	emp_employees	labor=True
employment_self	emp_selfemp	labor=True

Before this, no single source in puremacro carried all five.

qna_capital() and qna_tfp() — growth accounting

puremacro.capital turns the panel into the two inputs a production function needs and the residual they leave over. The package could already estimate production functions — korv_gmm fits a capital-skill-complementarity CES — but it could not build the capital series they take as given.

from puremacro.fetch import qna_panel
from puremacro.capital import qna_capital, qna_tfp

panel = qna_panel(["DEU", "ESP", "FRA", "ITA"], start="1995",
                  assets=True, labor=True, income=True, real=True)
cap = qna_capital(panel)          # perpetual inventory, one stock per asset
tfp = qna_tfp(panel, cap)         # Solow residual against hours

Two choices are load-bearing:

• Depreciation converts geometrically, 1-(1-δ)^(1/4), never δ/4. The linear shortcut compounds to less than δ over four quarters, so it under-depreciates and biases the steady-state stock up — by 5.3% for equipment and 8.5% for IPP.
• The aggregate is a Törnqvist index of capital services, not a sum of stocks. Chain-linked volumes are not additive away from their reference year, so Σ K_i depends on which year the OECD chose; qna_rebase the same panel and it moves by up to 1%. Aggregating growth rates with rental-cost weights is invariant to machine precision. Use aggregate="sum" only for the fixed-base publishers (ARG, IDN, MEX, ZAF), where additivity holds by construction.

Read k0_sensitivity before quoting a level. The initial stock has to be assumed and the assumption decays at the asset's own depreciation rate — fast for equipment (δ=0.13) and IPP (0.20), not for structures (0.03) or dwellings (0.011). Measured on this data, a ±50% error in K₀ is still worth 7–15% of the aggregate level at the end of a 30-year panel, while the same error moves four-quarter growth by under 0.06pp a year. Growth rates are usable; levels are an assumption wearing a number's clothes.

Coverage: 34 of 49 reference areas publish all four asset classes as volumes, 33 also in current prices (Colombia does not, so it has no deflators and is refused a services index), and 29 additionally publish the hours qna_tfp needs.

Offline use

The four transforms never touch the network, so the usual pattern is to fetch once, freeze, and work from the snapshot:

panel.reset_index().to_csv("qna.csv", index=False)          # freeze
flat = pd.read_csv("qna.csv")
panel = (flat.assign(date=pd.to_datetime(flat["date"]))
             .set_index(["code", "date"]).sort_index())     # thaw

One catch: qna_meta() reads panel.attrs, which no CSV round trip preserves. Freeze qna_meta(panel) as its own file if you need the provenance — which you do if you plan to rebase, since it is what records each country's base year. tools/gen_notebook_data_qna40.py is a worked example of the whole pattern.