DSA210 / Data & provenance

Know what is in the panel.

The 2015–2024 snapshots are versioned, checked and downloadable. “No missing values” is never inferred from a total row count.

Coverage, before any imputation

Non-missing values out of 120 months per country.
Country	Rows	CCI	Retail	Unemployment	Inflation
Germany	120	120	120	120	120
Spain	120	120	120	120	120
Turkey	120	120	120	120	120
UK	120	120	0	69	71

The UK is not silently completed

UK retail is absent for every month; it is excluded from retail-target models. UK unemployment is observed for 69 months and inflation for 71. Missing controls are imputed from training-fold medians with missingness indicators. No CCI or retail target is imputed.

A working data dictionary

Field	Unit	Definition / scope
CCI	Index, long-term average 100	Official OECD CCICP, amplitude adjusted; consumer-confidence survey index.
sunshine_hours	hours per month	Five-city fixed-population-weighted monthly sunshine; seconds converted to hours.
sunshine_capital_hours	hours per month	London, Berlin, Madrid, Ankara snapshots; sensitivity analysis only.
temperature_2m_mean	°C	Population-weighted city monthly average of daily means.
temperature_2m_max / min	°C	Weighted mean of city monthly daily-temperature extrema; not a national extreme.
precipitation_sum / rain_sum	mm per month	Weighted city monthly totals.
precipitation_hours	hours per month	Weighted monthly duration from source snapshot.
retail_index	Volume index, 2021=100	Eurostat snapshot (G47, seasonal/calendar adjusted per notebook); UK unavailable.
unemployment_rate	%	Eurostat control snapshot; some UK months missing.
inflation_yoy	% year-on-year	HICP control snapshot; some UK months missing. Raw percentages are retained.
covid_dummy	0 / 1	1 from March 2020 through June 2022; a coarse control, not a policy-effects estimate.
month_sin / month_cos	unitless	Cyclical calendar controls; deterministic, no fitted parameters.

Take the evidence with you

Research panel ↗480 rows · observed targets, weather and economic controls Model scores ↗Pooled out-of-fold metrics Every prediction ↗Actual vs predicted, model, country and test year Robustness tests ↗Effects, HAC intervals, p-values and adjusted q-values Calendar splits ↗Exact training/test boundaries Country metrics ↗Check pooled results by country Annual metrics ↗Inspect every test year Source manifest ↗URLs, filters, input SHA-256 checksums and environment Research report ↗Turkish research paper

What the snapshot cannot recover

Source vintage matters

CCI was retrieved on 31 October 2026. Weather, retail and macro inputs are versioned snapshots. Their exact source retrieval timestamps and the raw Eurostat responses were not retained. These are revised historical values, not a real-time vintage.

Country weather is a proxy

Five selected cities do not represent every household. Fixed population weights are documented in the notebook, but their reference year was not recorded. Open-Meteo uses reanalysis/model estimates; no new station observations were collected.