Macro Edge Research

Kuwait CPI 2006–2026 — Reconstruction Methodology

How 231 CSB bulletins in five formats became one cross-validated monthly series across three base years

Version 1.0 — August 2026

Series: monthly Kuwait CPI, all items and twelve COICOP divisions, 2006-01 → 2026-06 (246 months), plus official weights and CSB's exclusion aggregates from 2016-06.

Status: MER reconstruction of the official record. Values from 2013 onward reproduce the Central Statistical Bureau's printed base-2013 figures; history before 2013 is chain-linked from the 2000 and 2007 bases and is not an official series.

1. Why a reconstruction exists

Kuwait's Central Statistical Bureau publishes consumer prices as monthly bulletins — PDF and, in most years, Excel — on a page that offers no machine-readable archive, no continuous series, and no bridge across its three base years (2000, 2007, 2013). Public aggregators carry either annual averages or short recent windows. Anyone who wants the monthly record has to assemble it from the bulletins themselves. This document records how MER did that, and every safeguard between the bulletins and the published charts.

2. Source material

The raw archive is 231 CSB issues spanning 2007-01 → 2026-06 (333 files), held permanently on disk exactly as downloaded: 102 issues exist in both PDF and Excel, the rest in PDF only — all of the pre-2014 era, plus later stretches (notably 2021–2023) where CSB posted no workbook. Five container formats appear across the archive and each has its own parser path:

Era	Container	Notes
2007–2013	PDF bulletins	some without ruling lines — parsed from word geometry
2014–2016	multi-sheet Excel workbooks	one sheet per table
2017–2020	Excel + PDF pairs	the PDFs carry base-2013 history the same era's workbooks print in base 2007
occasional	OLE2 files served with an `.xlsx` name	detected by magic bytes, not extension
occasional	Word `.docx` with embedded worksheets	worksheets extracted and parsed as Excel

Arabic text is normalised before matching (Unicode NFKC, hamza/yā/tā-marbūṭa folding, tatweel and diacritic stripping, Eastern-Arabic digit translation), and reversed-Arabic runs — an artifact of some PDF text layers — are matched against their reversal.

3. Cross-validation: every month, multiple witnesses

Each bulletin prints a trailing table of roughly thirteen months, so almost every month of the series appears in many independent issues, and from 2014 in two file formats of the same issue. The parser exploits that redundancy instead of trusting any single file:

Every (base year, month, series) cell is collected from every issue that prints it, and disagreements are resolved by majority vote. CSB does not revise CPI, so a disagreement is a capture or typesetting error, not a revision — the wrong copy is simply outvoted, typically ~12-to-1.
Where a vote ties 1-vs-1, the candidate consistent with its resolved neighbours wins: a price index moves a fraction of a percent per month, so the candidate implying a smooth path beats the one implying a spike that reverses.
The PDF and Excel of the same issue are cross-checked pair-by-pair: of 5,435 comparable cells, 5,365 agree exactly (98.7%); the disagreements trace overwhelmingly to one PDF with swapped year columns, which the workbook copy outvotes.
Two confirmed defects are quarantined by name, with the evidence recorded in the parser source: a 2014 workbook whose January column is mislabeled "JAN 2013" (CSB's own typo) and a 2018 PDF carrying one misdated column. Other bad sources are neutralised by the pipeline's ordinary defences rather than a named list — a 2015 PDF with a corrupt font map and a run of 2013 PDFs with unparseable layouts simply fail to parse, and two 2009 PDFs with one misdated column each are outvoted by the majority — and every month they cover is independently supplied by neighbouring issues' trailing tables.

4. Base years and chain-linking

CSB rebased the index in 2007 and 2013 (bases 2000 = 100, 2007 = 100, 2013 = 100). The parser tags every observation with its base — stated in the table title where present, inferred from the level (~30% apart) where not — and never mixes bases in a single series. The published level series is base 2013 throughout: official prints from 2013-01 onward, and earlier history chain-linked per division through the overlapping months of adjacent bases (headline linking factors ≈ 0.6441 for base 2000 → 2013 and ≈ 0.7786 for base 2007 → 2013). Chain-linking preserves every month-to-month change exactly as printed in its own base; what it cannot preserve is the level's official status, which is why the pre-2013 span is labeled a reconstruction.

5. External anchors

Two checks against sources that never touched this pipeline:

World Bank annual inflation for Kuwait, 2018–2025: the reconstruction's annual averages match to ~0.01pp in every year.
The June 2026 release, the newest month at time of writing, matches the figures carried in Kuwaiti press coverage of the official bulletin (headline +2.19% y/y; ex-food +1.43%).

Before 2013, the reconstruction's annual averages sit 0.1–0.6pp below World Bank figures. That is the expected vintage effect of chain-linking from later-base tables rather than averaging the original base-2000 prints, and it is disclosed rather than adjusted away.

6. Internal consistency gates

Two identities are computed on every rebuild and gate what the site publishes:

Aggregation identity: the weighted division indices are compared to the printed all-items index. In the official base-2013 era the residual stays inside CSB's own rounding; the chain-linked era does not satisfy the identity exactly (different linking factors per division — an arithmetic consequence, not an error) and is therefore excluded from any chart that depends on it.
Contributions additivity: the contributions-to-inflation stack is shown only from 2016-12, where both ends of every 12-month window are official prints; its worst residual against the printed headline is 0.09pp, inside print rounding.

7. Update pipeline

The series refreshes automatically: a fetcher checks the CSB page daily and downloads any issue not already archived (the archive is append-only — nothing is ever deleted, so the record survives source-side removals); a new issue triggers a full re-parse of the entire archive with all of the validation above; a parse failure rolls back to the previous CSVs and leaves the published page untouched. The quarantine list is versioned in the parser source, so a re-parse cannot silently re-admit a known-bad file.

8. Known limitations

The pre-2013 span is a chain-linked reconstruction, not an official series, and carries the vintage gap described in §5.
CSB publishes to one or two decimal places depending on era; the series carries the precision as printed and never manufactures digits.
Exclusion aggregates (ex-food, ex-housing, ex-food-and-housing) exist only from 2016-06, when CSB began printing them.
The 2000-base era before 2006 exists in fewer, harder bulletins and is not yet included; the published series starts at 2006-01, the earliest month the archive supports at full cross-validation strength.

Prepared by Macro Edge Research (Qutaiba Mohammad, Macro Strategist). This document describes data construction only; it is general research, not investment advice, and contains no recommendation to deal in any security. The underlying official record belongs to the Central Statistical Bureau, State of Kuwait.