Every source in this database publishes some numbers that are wrong, and the interesting ones are wrong in ways that look plausible. This page is the record of what has been found, when, and what was concluded.
No published value is ever altered. The right repair for a base break is splicing a series, which changes numbers a statistical office stands behind — not a decision a pipeline should make silently. What this system does instead is detect defects the sources publish, record them with a date, and warn on the specific cells they ruin. The data you get is the source's data; what is added is knowing where not to trust it.
Deterministic rules run over the whole database after every ingest. "Open" means nobody has reviewed it yet, not that it is wrong.
| Check | Open | Raised |
|---|---|---|
| frequency_gap | 817 | 1,473 |
| base_break | 333 | 441 |
| outlier_value | 146 | 239 |
| outlier_unindexed | 32 | 32 |
| multiple_reference_period | 0 | 28 |
| stale_breakdown | 14 | 15 |
| zero_index | 7 | 12 |
| rate_vs_index | 3 | 3 |
| cross_source | 0 | 2 |
| breakdown_begins | 0 | 1 |
2,246 flags in total across 1,871 series, last run 2026-09-11 00:57.
Reviewed and written up. Each is a dated entry in the repository, so it survives the database being rebuilt from scratch.
| Finding | Found | Status | Series | Periods | Detected by |
|---|---|---|---|---|---|
lby-cpi-rebase-2025 | 2026-08-22 | confirmed | LBY.CPI_INDEX.M, LBY.CPI_CP*_INDEX.M, LBY.CPI_YOY.M +3 more | 2025-01-01 | base_break, cross_source, multiple_reference_period, manual |
| The published index contains a base break with no revision behind it: Dec 2024 is 303.7 and Jan 2025 is 101.1. The source moved to a new base going forward and left history alone, and reports 'Multiple' reference periods for the series itself. Found independently three ways: the month-over-month cliff, a 68pp disagreement with the World Bank's published annual inflation for 2025, and by eye. The IMF's own derived series divide across the break rather than handling it -- published MoM for Jan 2025 is -66.71% and YoY reads about -66% for the whole of 2025 -- so the artefact is in the source data, not in anything computed here. | |||||
dji-cpi-2013-discrepancy | 2026-09-01 | confirmed | DJI.CPI_INDEX.M, DJI.CPI_YOY.M, DJI.CPI_MOM.M | 2013-01-01 | cross_source, breakdown_begins |
| Resolved 2026-09-02, and it is a base break in the IMF's monthly index at January 2013. The disagreement was the symptom: the monthly levels imply -12.56% for 2013 where the World Bank publishes +2.71%, 15.3pp apart and one of only two country-years out of 4,994 to disagree by more than 2pp. What settles it is not the headline but the DIVISIONS. The index steps 99.41 -> 84.31 in January 2013, -15.2%, and then climbs normally from the new level, giving +1.3% for 2014 -- ordinary inflation on either side of one cliff. And all twelve COICOP divisions BEGIN in that same month: there is no December 2012 for any of them. A new collection starting and the old headline continued through it without rescaling is exactly what that looks like, and the World Bank's +2.71% is then the correct figure for the year. Worth keeping: base_break never saw this. At -15.2% it sits under the -20% threshold, and lowering the threshold to catch it would take the raw hit count from 130 to 216 -- most of them real -- for this one case. The threshold stays where it is and cross_source covers the band beneath it, which is the argument for having two checks that look at different evidence. | |||||
imf-multiple-base-declared | 2026-09-02 | confirmed | KNA.CPI_*, LBR.CPI_*, MDA.CPI_* +1 more | whole series | multiple_reference_period |
| The IMF reports reference_period 'Multiple' for every CPI series of these four countries -- 13 each for St Kitts and Nevis and Libya, 1 each for Liberia and Moldova. That is the source declaring the series is not on a single base, not something inferred here, so it is confirmed on the source's own authority. Where in the history the base changes is NOT established: the base_break check finds a cliff in Moldova but none in the other two, which means either the change was small enough to pass under -20% or it predates our history. | |||||
wbg-conflict-volatility-2024-2025 | 2026-09-02 | suppressed | WBG.CPI_* | whole series | base_break, outlier_value |
| Both shape checks are defeated here, and by the same thing. The single-month outlier rule requires a month to sit far outside its neighbours WHILE THEY AGREE with each other; in Gaza in 2024 and 2025 consecutive months differ by 28% legitimately, so two of its months clear that bar without being defects. The threshold was left where the rest of the database wants it and this territory named here instead: moving a number until one country disappears is not a finding. The check fires eight times for West Bank and Gaza between Jul 2024 and Nov 2025, and this is the check misreading real data rather than a defect. A rebasing happens once and is permanent; these move in both directions and never settle -- the headline index runs +9.2%, +14.8%, +3.0%, -32.6%, -14.1%, +14.0%, +30.7%, +23.6% across nine months, and WBG appears in both the 'level returns' and 'level stays down' populations. That is a wartime economy, and the numbers are the source's best account of it. Suppressed rather than confirmed: nothing here is wrong with the data. | |||||
lso-cpi-rebase-2009-03 | 2026-09-02 | confirmed | LSO.CPI_* | 2009-03-01 | base_break |
| Four Lesotho series drop together in March 2009 and stay down: the headline index goes 69.03 -> 50.93 and averages 51.53 over the following six months, 0.75x its pre-fall level. Permanent, simultaneous across series, and the only break month this country has -- the rebasing signature on all three counts. No source announcement was found; the conclusion rests on the shape of the data alone. | |||||
gin-cpi-2007-01-month-scrambled | 2026-09-02 | confirmed | GIN.CPI_* | 2007-01-01 | base_break, outlier_value |
| NOT a base break: January 2007 is one bad month across ELEVEN Guinean series at once, and the giveaway is that it is wrong in both directions. Six fall and five rise -- headline 25.15 -> 13.63 -> 25.99, communication 12.0 -> 23.5 -> 11.8, transport 32.4 -> 12.8 -> 30.0 -- and every one of them returns to trend the following month, within a few percent of where it was in December. A rebasing is permanent and moves one way; this does neither. Something went wrong with the whole month's return, which is a larger and more specific fault than a transcription error in one cell, though the cause is still not established. | |||||
lka-cpi-cp05-2024-04-tenfold | 2026-09-02 | confirmed | LKA.CPI_CP05_* | 2024-04-01 | outlier_value |
| One month multiplied by ten, not a rebasing. Sri Lanka's furnishings index reads 186.34 in March 2024, 1,875.00 in April, and 187.40 in May, and carries on from there -- 188.4 by July. The two months either side agree with each other to 0.6%, so there is no new base: there is one wrong cell, off by a factor of ten, in a series that is otherwise continuous. The earlier reading of this as a base change was drawn from base_break, which only detects falls and therefore lands on the month AFTER the defect, and it had the practical consequence backwards: levels either side of April 2024 are directly comparable. | |||||
zmb-cpi-outlier-months | 2026-09-02 | confirmed | ZMB.CPI_CP01_*, ZMB.CPI_CP02_* | 2024-02-01, 2022-03-01 | outlier_value |
| Two isolated bad months, neither a rebasing. Food reads 4,755.04 in February 2024 between 464.53 in January and 486.52 in March -- ten times its neighbours, which agree with each other to 4.7% -- and the series carries on climbing steadily afterwards, 491.4, 497.6, 504.8. Alcohol and tobacco has a smaller one of the same shape: 327.44 in March 2022 between 263.83 and 266.86. Both were found by looking at the months either side rather than at the fall, which is why the first pass filed the food one as a base change in March 2024, a month late and with the wrong consequence. | |||||
caf-cpi-suspended-2021-2024 | 2026-09-02 | confirmed | CAF.CPI_* | whole series | frequency_gap |
| The Central African Republic publishes nothing between December 2021 and November 2024 -- 34 consecutive months across every CPI series -- and then resumes. The gap is a fact of the source response, not a parse failure: the surrounding months are present in the same request. WHY publication stopped is not established here. | |||||
dji-cpi-suspended-2023-2026 | 2026-09-02 | confirmed | DJI.CPI_* | whole series | frequency_gap |
| Djibouti publishes nothing between July 2023 and March 2026 -- 31 consecutive months across every CPI series -- and then resumes, reaching July 2026. Note this country also carries an 2013 base break (see dji-cpi-2013-discrepancy); the two are separate and neither explains the other. | |||||
bhs-cpi-suspended-2015-2018 | 2026-09-02 | confirmed | BHS.CPI_* | whole series | frequency_gap |
| The Bahamas publishes nothing between December 2015 and January 2018 -- 24 consecutive months across every CPI series -- and then resumes, currently reaching April 2026. The gap is in the source response rather than in the parse. | |||||
cog-cpi-suspended-2018 | 2026-09-02 | confirmed | COG.CPI_* | whole series | frequency_gap |
| The Republic of the Congo publishes nothing between August 2018 and February 2019 -- five consecutive months across every CPI series -- and then resumes, currently reaching May 2026. Short enough that it could be a reporting delay rather than a suspension; the gap itself is not in doubt. | |||||
yem-cpi-ends-2015 | 2026-09-02 | confirmed | YEM.CPI_* | whole series | frequency_gap |
| Yemen is not a suspension and the gap check's wording is slightly wrong about it. There is an eleven-month hole from December 2014, the series resumes for a single month in December 2015, and that is the newest observation in the database -- so the check's 'the series resumes afterwards' is technically true and misleading. Publication stopped in 2015, during the war. Nothing since has been withdrawn; the source simply has no more to send. | |||||
mdg-cpi-2010-01-month | 2026-09-02 | confirmed | MDG.CPI_* | 2010-01-01 | outlier_value |
| January 2010 is wrong across four Madagascan divisions at once, all of them too HIGH and all returning the next month: recreation 71.7 -> 150.9 -> 73.1, furnishings 74.8 -> 117.6 -> 76.8, clothing 64.3 -> 79.2 -> 64.8, health 54.9 -> 69.5 -> 56.3. December and February agree with each other to within 3% in every case, so there is no step and no new base -- one month's return is bad. Consistent direction across four divisions points at the month rather than at the categories. | |||||
bdi-cpi-2025-09-month | 2026-09-02 | confirmed | BDI.CPI_* | 2025-09-01 | outlier_value |
| September 2025 is wrong across four Burundian divisions and, like Guinea's January 2007, it is wrong in BOTH directions: communication 134.3 -> 248.5 -> 133.3 and education 191.7 -> 270.8 -> 195.8 spike, while recreation 263.4 -> 134.2 -> 263.5 and restaurants 294.5 -> 196.6 -> 303.5 halve. Every one of the eight surrounding months agrees with its partner to within 3%. Two of these look like the same pair of numbers swapped between divisions, which is a guess; that the month is unusable is not. | |||||
bdi-cpi-isolated-outlier-months | 2026-09-02 | confirmed | BDI.CPI_CP06_*, BDI.CPI_CP10_*, BDI.CPI_CP12_* | 1999-05-01, 2000-01-01, 2007-01-01, 2007-09-01 | outlier_value |
| Four scattered bad months in Burundi's older history, each in a single division and each with the neighbours agreeing to within 5%: education reads 9.14 in May 1999 between 59.42 and 59.42 -- the most extreme cell in the database, a sixth of its own neighbours; health 36.9 in January 2000 between 28.5 and 29.7; miscellaneous 55.3 in January 2007 and 55.6 in September 2007, both between values near 42. Separate from the September 2025 event, which is a whole month rather than one series. | |||||
sdn-cpi-isolated-outlier-months | 2026-09-02 | confirmed | SDN.CPI_CP01_*, SDN.CPI_CP02_*, SDN.CPI_CP05_* +3 more | 1996-04-01, 1996-07-01, 1996-09-01, 1997-01-01, 1997-11-01, 2001-04-01, 2003-04-01, 2019-02-01 | outlier_value |
| Nine isolated bad months across eight dates, more than any other country, and the concentration in 1996-97 is not a coincidence: that is Sudan's hyperinflation, when the index doubles inside a year and the published values carry one decimal place at levels around 5, so the data is at its least robust exactly where it moves fastest. The rule used here does not simply ask for a big move -- it requires the months either side to AGREE, which they do in every case here to within 15% while the flagged month sits 20-55% off the midpoint. Examples: furnishings 44.9 -> 21.5 -> 51.6 (Apr 1996), education 38.1 -> 30.0 -> 39.3 (Nov 1997), food 67.1 -> 52.2 -> 69.3 (Apr 2003), communication 975.8 -> 1,423.5 -> 1,043.8 (Feb 2019). | |||||
syc-cpi-2017-01-outliers | 2026-09-02 | confirmed | SYC.CPI_CP07_*, SYC.CPI_CP08_* | 2017-01-01 | outlier_value |
| Two Seychelles divisions are wrong in January 2017 and in opposite directions: transport reads 131.2 between 101.8 and 104.1, communication 66.8 between 98.0 and 98.0. Communication is the cleaner case -- its neighbours are identical to three figures, so there is nothing for a real movement to hide in. Both return completely the following month. | |||||
cmr-cpi-2025-07-outliers | 2026-09-02 | confirmed | CMR.CPI_CP06_*, CMR.CPI_CP07_* | 2025-07-01 | outlier_value |
| Two Cameroonian divisions are wrong in July 2025, in opposite directions, against neighbours that are identical: health reads 132.1 between 102.1 and 102.1, transport 100.0 between 132.1 and 132.5. The two wrong values are each other's neighbours' levels, which is what a swap between two divisions looks like -- a reading the numbers suggest and do not prove. Either way both months are unusable and both series continue undisturbed. | |||||
imf-isolated-outlier-months | 2026-09-02 | confirmed | BGD.CPI_CP09_*, BHR.CPI_CP09_*, COD.CPI_CP07_* +5 more | 2019-12-01, 2025-05-01, 2017-01-01, 2009-12-01, 2014-06-01, 2014-07-01, 2019-04-01, 2012-07-01 | outlier_value |
| Eight single bad months in eight countries, one division each, with nothing tying them together beyond the shape and the source. Kept as one entry rather than eight near-identical ones, with every case named so none is hidden behind a summary: Bangladesh recreation 149.90 in Dec 2019 between 98.1 and 98.6; Bahrain recreation 101.7 in May 2025 between 64.3 and 67.6; DR Congo transport 98.1 in Jan 2017 between 130.5 and 131.2; Congo communication 111.2 in Dec 2009 between 72.7 and 78.2; Djibouti alcohol and tobacco 158.8 in Jun 2014 between 80.7 and 89.9; Gabon communication 100.3 in Jul 2014 between 135.0 and 138.8; Pakistan restaurants 237.3 in Apr 2019 between 119.5 and 120.8, almost exactly double; Chad housing 60.2 in Jul 2012 between 84.0 and 89.4. In every case the two neighbours agree with each other far more closely than either agrees with the flagged month, and the series continues from where it left off. | |||||
mus-cpi-divisions-absent-1999-2002 | 2026-09-02 | confirmed | MUS.CPI_CP08_*, MUS.CPI_CP10_*, MUS.CPI_CP11_* | whole series | zero_index, frequency_gap |
| Three Mauritian divisions have no real data before mid-2002, and the source records the absence two different ways in the same series -- which is why it took two checks to see one event. Communication, education and restaurants all read exactly 0.00 from July 1999; education and restaurants then stop being published at all from February 2001 to June 2002, while communication carries on reading 0.00 until June 2002. All three start at a real level in July 2002 (56.99, 44.11, 30.16). The zeros are placeholders, not price levels -- an index of zero means the basket is free -- and the source agrees with that reading: it publishes no month-over-month rate for July 2002, which is precisely the month a rate against zero would be infinite. 74 of the 80 zero index cells in the database are these three series. | |||||
zwe-cpi-divisions-zero-2009 | 2026-09-02 | confirmed | ZWE.CPI_CP08_*, ZWE.CPI_CP10_* | 2009-01-01 | zero_index |
| Zimbabwe's communication and education indices read exactly 0.00 for January, February and March 2009 and then start at 3.59 and 3.49 in April. The country's headline index is published for those same three months (1.96, 1.90, 1.84), so this is not the office being shut: two divisions have no number while the aggregate does. That is the quarter the Zimbabwe dollar was abandoned, and the zeros are placeholders for prices that could not be collected in it -- a reading the shape supports and no source statement here confirms. As in Mauritius, the source publishes no rate for the first real month. | |||||
pan-cpi-2023-withdrawn-not-suspended | 2026-09-02 | confirmed | PAN.CPI_INDEX.M, PAN.CPI_MOM.M, PAN.CPI_YOY.M | 2023-07-01 | frequency_gap |
| Not a suspension, and this database can prove it. The gap check sees three missing months from August 2023 and calls it a pause in publishing; the revisions table says the IMF PUBLISHED August, September and October 2023 -- 99.73, 99.87, 99.68 -- and then withdrew all three in vintage 12. They are the first withdrawals the tombstone code ever recorded on live data. The year-over-year series lost six months rather than three, August to October 2023 and again August to October 2024, because those later months compare against the withdrawn ones: the derivation showing through in the revision record exactly as the arithmetic predicts. Why the source retracted the quarter is not established. The gap check should eventually consult `revisions` and say 'withdrawn' where a tombstone covers the hole, rather than 'suspended'. Recorded here rather than built now: it is one country, and the sentence being wrong is a smaller problem than a check that reaches into a table it has never needed. | |||||
bdi-cpi-division-steps-2024 | 2026-09-02 | confirmed | BDI.CPI_CP07_*, BDI.CPI_CP08_* | 2024-05-01, 2024-11-01 | base_break |
| Two Burundian divisions step down in 2024: transport 183.20 -> 119.30 in May (-34.9%, six-month average 0.78x) and communication 203.30 -> 125.40 in November (-38.3%, 0.66x). Transport is the odd one -- it sits near 128 for six months and then jumps to 211.2 in November 2024 and keeps climbing to 231, ABOVE where it started, so whatever happened in May was undone in November rather than carried forward. That is a third shape and it is not a rebasing either. Burundi has more flags than any other country in this register (see also the two outlier entries); its series need a closer look than one pass. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
bhr-cpi-cp09-2020-03 | 2026-09-02 | confirmed | BHR.CPI_CP09_* | 2020-03-01 | base_break |
| Bahrain's recreation and culture index falls 94.50 -> 71.20 in March 2020 (-24.7%) and keeps falling to 57.5 by November, averaging 0.65x. Unlike the others in this class it does not step and stop -- it declines for eight more months and then recovers to 77.3 by December 2020. A base change does not do that, and the month is the one every recreation index in the world moved in, so the balance of evidence here favours a real collapse in this category rather than a re-reference. Kept confirmed rather than suppressed because that is a judgement about the world, not a measurement from the data. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
dji-cpi-division-steps | 2026-09-02 | confirmed | DJI.CPI_CP02_*, DJI.CPI_CP05_*, DJI.CPI_CP06_* | 2014-05-01, 2015-01-01 | base_break |
| Two different things in one country. Furnishings and health both step down in January 2015 -- 120.01 -> 95.97 and 126.25 -> 99.55 -- and stay at 0.81x and 0.79x, landing within a few points of 100 and moving barely at all for the following year, which is what a division put on a 2015 base looks like. Alcohol and tobacco is not that: 108.15 -> 80.66 in May 2014 with a six-month average of 0.97x, in a series that swings between 80 and 159 across 2014 and carries an isolated bad month of its own in June (see imf-isolated-outlier-months). That series is noisy throughout and no base change is visible in it. Separate from the January 2013 headline break, which has its own entry. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
ecu-cpi-cp04-2026-05 | 2026-09-02 | confirmed | ECU.CPI_CP04_* | 2026-05-01 | base_break |
| Ecuador's housing index steps 127.69 -> 101.29 in May 2026, -20.7%, and that is the newest month held -- there is no 'afterwards' to look at yet, so the shape test the other entries rest on cannot be run. Landing at 101.29 is consistent with a re-reference, and it is the only thing pointing that way. Worth returning to once a few more months arrive: this is the one case in the register that a later fetch will settle by itself. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
est-cpi-cp10-2014-09 | 2026-09-02 | confirmed | EST.CPI_CP10_* | 2014-09-01 | base_break |
| Estonia's education index steps 211.90 -> 162.10 in September 2014 (-23.5%) and holds at 162.10 exactly for four months before resuming. September is when an education index reprices, and holding a value flat between repricings is normal for this division, so the shape is less informative here than elsewhere. Estonia is also the one country in this class with no usable base check: it declares 1997A and the series does not reach back that far. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
eth-cpi-cp11-2001-01 | 2026-09-02 | confirmed | ETH.CPI_CP11_* | 2001-01-01 | base_break |
| The steepest in this class: Ethiopia's restaurants and hotels index goes 35.82 -> 13.34 in January 2001, -62.8%, and stays at 0.39x -- around 14 for the following year. December 2000 is the FIRST month of the series, so there is exactly one observation on the old footing and nothing to characterise it with. A single orphaned value ahead of a clean start is more consistent with a splice than with a 63% fall in restaurant prices, but with one point there is nothing to test. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
gab-cpi-cp08-2015-05 | 2026-09-02 | confirmed | GAB.CPI_CP08_* | 2015-05-01 | base_break |
| Gabon's communication index steps 152.96 -> 96.58 in May 2015 (-36.9%) and settles at 0.56x. It is the one series in this class that fails the base check: Gabon declares 2018A and this division averages 94.8 across 2018, not 100, so its published history is not on the base the source states for it. That is independent evidence of a splice, and it is the strongest single piece of evidence in this group. The series also carries a bad month in July 2014 (see imf-isolated-outlier-months), which is a separate fault. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
geo-cpi-cp04-2020-12 | 2026-09-02 | confirmed | GEO.CPI_CP04_* | 2020-12-01 | base_break |
| Georgia's housing and utilities index drops 127.34 -> 99.35 in December 2020, holds near 100 for exactly three months, and returns to 137.2 in March 2021 -- above where it started. A clean three-month rectangle in one division, which is what a temporary subsidy or price cap looks like in an index, and not what a base change looks like. The level RETURNS, which rules a base change out: a rebasing is permanent. That leaves a run of months the series recovers from -- too long for the single-month outlier rule, which needs the neighbours either side to agree, and too short to be a new base. Whether the run is bad data or a real episode is not established. | |||||
gin-cpi-division-steps | 2026-09-02 | confirmed | GIN.CPI_CP01_*, GIN.CPI_CP08_* | 2010-02-01, 2022-08-01 | base_break |
| Guinea's food index steps 49.62 -> 28.48 in February 2010 (-42.6%, 0.60x) and its communication index 325.55 -> 104.40 in August 2022 (-67.9%, 0.32x, landing on 100 and staying within a point of it for a year). Food is the case where the headline argument carries real weight: food is the heaviest division in a low-income basket, a 43% fall in it could not leave the headline at +0.9% for the month, and the arithmetic does not work for a real price fall. Separate from January 2007, which is a bad month across the whole country. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
grc-cpi-cp01-1973-09 | 2026-09-02 | suppressed | GRC.CPI_CP01_* | 1973-09-01 | base_break |
| The oldest flag in the register and an artefact of precision, not a defect. Greek food in September 1973 reads 1.36 against 1.72 in August -- but this is a 2020-based index fifty years before its base, so the values are around 1.7 and are published to two decimals. One step of the last digit is 0.6% at this level, and the whole flagged movement is four points of a quantity that has three significant figures. The series returns to 2.05 within six months and to 2.2 by January 1974, which is 1973-74 Greek inflation doing what it did. Suppressed: the check is right that the ratio is large and wrong that it means anything. | |||||
ken-cpi-cp08-2010-08 | 2026-09-02 | confirmed | KEN.CPI_CP08_* | 2010-08-01 | base_break |
| Kenya's communication index steps 118.16 -> 90.42 in August 2010 (-23.5%) and stays at 0.75x, drifting slowly down to 87 over the following year rather than sitting flat. A level that keeps moving after the step is a point against a re-reference, which usually lands somewhere and stops. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
kos-cpi-division-steps | 2026-09-02 | confirmed | KOS.CPI_CP07_*, KOS.CPI_CP10_* | 2008-10-01, 2014-10-01 | base_break |
| Two different things. Transport falls 108.70 -> 83.63 in October 2008 (-23.1%) and then moves continuously -- 96.1, 89.5, 87.8, 88.5 -- which is a series still being measured, in the month the oil price halved. Education is the opposite: 138.94 -> 99.94 in October 2014 and then EXACTLY 99.94 for the following year and beyond, which is not a measurement at all but a value carried forward (see imf-indices-held-constant). Kosovo declares a 2015 base, and 99.94 from October 2014 is a division put on it early and then left. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
lbn-cpi-cp08-2014-06 | 2026-09-02 | confirmed | LBN.CPI_CP08_* | 2014-06-01 | base_break |
| Lebanon's communication index steps 99.64 -> 78.78 in June 2014 (-20.9%) and then sits between 76.0 and 76.4 for a year -- a range of half a percent, which is a series that has stopped moving rather than one being repriced. The step is from almost exactly 100, and Lebanon declares a December 2013 base, so what changed in June 2014 was not the base: the division was at its base value and left it once. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
lbr-cpi-division-dips | 2026-09-02 | confirmed | LBR.CPI_CP04_*, LBR.CPI_CP08_* | 2016-01-01, 2006-01-01 | base_break |
| Two Liberian dips that both come back, and both start in January. Housing reads 56.3, 56.4, 56.4 for January to March 2016 between 72.0 in December 2015 and 72.7 in April 2016 -- three months at 78% of the surrounding level, then straight back. Communication does the same shape in two months: 61.35, 61.35 in January and February 2006 between 84.77 and 79.2. A run of identical values that returns to trend is a placeholder, not a price path. Liberia is also one of the countries whose source declares 'Multiple' reference periods. The level RETURNS, which rules a base change out: a rebasing is permanent. That leaves a run of months the series recovers from -- too long for the single-month outlier rule, which needs the neighbours either side to agree, and too short to be a new base. Whether the run is bad data or a real episode is not established. | |||||
lca-cpi-division-steps | 2026-09-02 | confirmed | LCA.CPI_CP08_*, LCA.CPI_CP09_* | 2022-01-01, 2016-06-01 | base_break |
| Saint Lucia's communication index steps 101.45 -> 80.66 in January 2022 and then holds between 79.0 and 81.5 for a year; recreation steps 148.61 -> 117.09 in June 2016 and continues down to 107 over the following months. The communication case steps from almost exactly 100 against a declared January 2018 base, so the pre-step level is the base value and the step is a departure from it rather than a re-reference to it. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
mrt-cpi-cp02-2020-08 | 2026-09-02 | confirmed | MRT.CPI_CP02_* | 2020-08-01 | base_break |
| Mauritania's alcohol and tobacco index steps 170.93 -> 108.69 in August 2020 (-36.4%) and then keeps falling -- 89.1, 80.9, 77.1, 74.1, 71.3, 63.0, 60.4 -- to a six-month average of 0.46x and well past that afterwards. The step is the start of a sustained decline rather than a level change, which is the shape a re-reference does not have. This is the strongest 'the prices really moved' case in the group. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
mus-cpi-cp02-2018-04 | 2026-09-02 | confirmed | MUS.CPI_CP02_* | 2018-04-01 | base_break |
| Mauritius' alcohol and tobacco index steps 100.24 -> 79.20 in April 2018 and then holds between 79.0 and 80.3 for a year. It also fails the base check: Mauritius declares 2023A and this division averages 104.5 across 2023 rather than 100, so its history is not on the declared base -- the same signal as Gabon. Separate from the three divisions that were not collected before 2002 (see mus-cpi-divisions-absent-1999-2002). The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
sdn-cpi-division-steps | 2026-09-02 | confirmed | SDN.CPI_CP07_*, SDN.CPI_CP08_*, SDN.CPI_CP09_* +1 more | 2006-10-01, 1996-12-01, 1997-07-01 | base_break |
| Sudan's clearest case is October 2006, where transport and communication step together -- 136.33 -> 101.57 and 135.65 -> 101.07, both -25.5% to the decimal -- and then sit at 100 for a year. Two divisions moving by the identical percentage in the same month, landing on 100, against a declared 2007 base: that is a re-reference of both to the coming base and nothing else fits. Education in July 1997 is a smaller version, 46.46 -> 35.95 and stays at 0.79x. Recreation in December 1996 is not in this class at all -- 5.86 -> 3.43 and back to 5.61 within six months, during the hyperinflation, in a series published to two decimals at levels near 5. Sudan also carries nine isolated bad months (see sdn-cpi-isolated-outlier-months). The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
sle-cpi-cp10-2018-09 | 2026-09-02 | confirmed | SLE.CPI_CP10_* | 2018-09-01 | base_break |
| Sierra Leone's education index steps 45.58 -> 34.69 in September 2018 (-23.9%) and then climbs slowly and steadily -- 34.7, 34.9, 34.9, 35.0, 35.2, 35.4, 35.7 -- which is a series being measured from a new level rather than parked at one. September is an education repricing month, as in Estonia. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
ssd-cpi-volatility-2019-2020 | 2026-09-02 | suppressed | SSD.CPI_* | 2019-11-01, 2020-06-01 | base_break |
| South Sudan has no divisions -- only a headline index -- so every argument the other entries make from one division against the aggregate is unavailable here, and the check's country-level reasoning reduces to the shape alone. The shape is a series that swings both ways and recovers: 46.57 -> 31.11 in November 2019 and back to 47.3 by May 2020; 47.29 -> 37.72 in June 2020 and back to 55.5 by November. Rises of the same size are in the same history (+34% from June to October 2020) and only the falls are flagged, because the check is asymmetric by design. This is the West Bank and Gaza pattern in a smaller economy, and the same ruling follows: the numbers are the source's account of a currency in trouble, not a defect. | |||||
sur-cpi-division-steps | 2026-09-02 | confirmed | SUR.CPI_CP02_*, SUR.CPI_CP04_*, SUR.CPI_CP06_* | 2009-04-01, 2016-04-01 | base_break |
| Suriname is the one country here where the source's own metadata names the date. It declares 'APRIL 2016 - JUNE 2016=100', and the housing index steps 130.38 -> 96.50 in APRIL 2016 -- onto the declared base, in the declared month. That is a re-reference with the history left unscaled, confirmed by the source's own label rather than inferred from shape, and it is the most solidly established base break in this register after Libya's. The April 2009 pair (alcohol and tobacco 63.04 -> 43.84, health 55.08 -> 44.06) is a separate event seven years earlier that the declaration does not cover, sharing only the month. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
syc-cpi-division-steps | 2026-09-02 | confirmed | SYC.CPI_CP03_*, SYC.CPI_CP10_* | 2009-01-01, 2010-01-01 | base_break |
| Two January steps a year apart, and only one of them stays. Clothing goes 73.33 -> 53.82 in January 2009 (0.77x) and then climbs continuously back to 62.2 by September -- the Seychelles rupee was floated at the end of 2008 and this is a series moving hard in both directions around it. Education goes 80.44 -> 64.29 in January 2010, holds at 72.2 for three months, and is back to 80.44 -- its exact former value -- by May, which is a hole in the series rather than a price path. Neither is a base change. Separate from the two bad months in January 2017. The level RETURNS, which rules a base change out: a rebasing is permanent. That leaves a run of months the series recovers from -- too long for the single-month outlier rule, which needs the neighbours either side to agree, and too short to be a new base. Whether the run is bad data or a real episode is not established. | |||||
tcd-cpi-cp12-2015-01 | 2026-09-02 | confirmed | TCD.CPI_CP12_* | 2015-01-01 | base_break |
| Chad's miscellaneous goods index steps 151.76 -> 115.02 in January 2015 (-24.2%) and then sits between 115.0 and 118.5 for a year. It fails the base check as Gabon and Mauritius do: Chad declares 2022A and this division averages 105.6 across 2022 rather than 100. A January step in a series that is not on its own declared base is the re-reference signature, though it is landing near 115 rather than near 100, which a plain re-reference would not do. The step is permanent, confined to one division, and not a revision: these values were published this way. What the data CANNOT settle is whether the division was put on a new base by itself or its prices really fell, because both produce exactly this shape and nothing held here separates them -- the series is verified at 100 in the reference period the source declares, which is true either way. The country's headline index barely moves in the same month, which is suggestive and not conclusive: these are mostly light divisions, so a real fall would leave only a small mark on the aggregate. What IS certain is the consequence, because the source's own rates are this index restated -- the published month-over-month and year-over-year figures equal the index ratios at all 388,805 and 370,534 overlapping cells, checked 2026-09-02. So if the step is a re-reference, the source's own rates divide across it, and the reader has to be told. | |||||
zwe-cpi-cp01-2023-07 | 2026-09-02 | suppressed | ZWE.CPI_CP01_* | 2023-07-01 | base_break |
| Zimbabwe's food index falls 966.37 -> 670.02 in July 2023, and unlike every other entry in this class the country's HEADLINE falls with it, -15.3% in the same month. That is the test the other cases fail, passed: a division and the aggregate moving together is a price level moving, not one series being re-referenced. The run into it is +26% then +104% in the two preceding months, which is the Zimbabwe dollar in mid-2023, and the series ends two months later at 622.35. Suppressed: nothing here is wrong with the data. Zimbabwe separately publishes two divisions as 0.00 in 2009 (see zwe-cpi-divisions-zero-2009). | |||||
imf-indices-held-constant | 2026-09-02 | watching | LBR.CPI_CP10_*, BDI.CPI_CP10_*, SLB.CPI_CP08_* +6 more | whole series | manual |
| Found while working the base-break queue and recorded rather than checked, because measuring it did not support a rule. 109 runs across 74 series publish an index that does not change AT ALL for two years or more. Liberia's education index is exactly 100.00 for 157 consecutive months -- thirteen years -- and Burundi's is 95.00 for 98. The reason there is no check: 82 of the 109 runs are communication and education, which is exactly where administered and regulated prices sit, so a flat index in those divisions is often the correct answer rather than a placeholder. A rule firing 109 times in the two divisions where flatness is normal would be the queue crying wolf, which this file is explicit about not doing. The subset worth returning to is the 19 runs pinned at exactly 100.00: a series sitting on its base value and never leaving it is a different claim from one holding a fixed tariff. The series named here are the longest runs, so a reader of any of them is told; the other 65 series are not, which is a gap this entry does not close. | |||||
imf-gbr-cpi-is-cpih | 2026-09-02 | confirmed | GBR.CPI_INDEX.M, GBR.CPI_CP*_INDEX.M, GBR.CPI_YOY.M +3 more | whole series | manual |
| The series this database holds as the United Kingdom's CPI, from the IMF, is the ONS's CPIH -- a different index on a wider basket, including owner occupiers' housing costs and council tax. Found while building the ONS connector, by comparing the IMF's GBR series against both of the ONS's published indices. The index is identical to CPIH at every one of 6,019 overlapping monthly cells across all thirteen concepts (the all-items aggregate and twelve divisions, 463 months each), and matches the ONS CPI at 1,882 of the same 6,019. The gap is not rounding: CP04 (housing, water and fuels) runs up to 16.1 index points apart from the CPI, which is roughly what owner occupiers' housing costs are worth, and the all-items index up to 2.2 points. No value here is wrong -- every number is a number the ONS published -- but the NAME is, and a reader comparing 'UK CPI' from this database against the ONS's headline gets a different figure with nothing to explain it. The series ids are public identifiers and are not being renamed (ARCHITECTURE section 10); the UK's actual CPI now arrives under GBR.CPI_ONS_* from ons.mm23, alongside CPIH under GBR.CPIH_ONS_*. Separately, the IMF's published year-over-year figure for GBR is the ratio of its own ROUNDED index at all 451 months, and agrees with the rate the ONS actually publishes in only 302 of them -- so where both exist, the ONS series is the one to quote. | |||||
gbr-gas-cap-2023-07 | 2026-09-02 | suppressed | GBR.CPI_ONS_CP0452_*, GBR.CPI_ONS_CP04521_*, GBR.CPIH_ONS_CP0452_* +1 more | 2023-07-01 | base_break |
| A real fall in a regulated price, not a change of base. UK gas drops 25.2% in July 2023 (225.90 -> 168.90; natural gas 25.8%) and stays down, which is the shape base_break is built to suspect. Three things say it is a price movement. First, it dilutes correctly up the weighted hierarchy: gas -25.2%, its parent electricity/gas/other fuels -15.1%, housing and fuels -4.1%, all items -0.4%. A re-referencing does not do that -- the ONS would re-reference the parents with the child, and the aggregate would not move at all. Second, the ONS's own published month-over-month figure agrees, and for this source that is independent evidence rather than arithmetic: it is computed from unrounded levels and disagrees with the ratio of the published index at 31.7% of cells. Third, the level is FLAT either side (228.0, 228.0, then 169.3, 169.3, 169.3), which is what an administered price resetting looks like and not what a re-referencing looks like. The 'four series break together' line in the flag counts GAS and natural gas twice, once for CPI and once for CPIH -- one category, not four. | |||||
gbr-council-tax-1991-04 | 2026-09-02 | suppressed | GBR.CPIH_ONS_CP049_* | 1991-04-01 | base_break |
| An administered price cut, and the same shape as the 2023 gas cap: flat, one step, flat. Council tax and rates reads 52.6 for January, February and March 1991, 37.1 in April, then 37.0 for the months after -- a 29.5% fall with no drift on either side. A re-referencing would rescale the whole series, not step it once in the middle, and the months before and after are on the same scale as each other. The cause is not in the data but is not in doubt either: the community charge was cut by 140 pounds a head from April 1991, funded by the VAT rise announced that March. This series exists only in CPIH -- council tax is outside the CPI basket -- which is why one flag appeared here without a CPI twin. | |||||
gbr-liquid-fuels-oil | 2026-09-02 | suppressed | GBR.CPI_ONS_CP0453_*, GBR.CPIH_ONS_CP0453_* | 1991-03-01, 2026-06-01 | base_break |
| Heating oil tracking crude, twice, and in both cases a trend rather than a step. March 1991: 51.4, 50.0, 39.9, 37.5, 38.4 -- the fall continues into the following month and then partially recovers, which no change of base does. June 2026: 324.9, 264.3, 211.3, 209.1 -- three consecutive monthly falls, of which the check flags only the one that crosses -20%. Liquid fuels carry a small weight, so the dilution up the hierarchy is small but present and in the right direction (-1.1% at electricity, gas and other fuels in 1991). A base change would be one step and permanent; both of these are runs. | |||||
gbr-thin-subclass-volatility | 2026-09-02 | suppressed | GBR.CPI_ONS_CP01116_*, GBR.CPIH_ONS_CP01116_*, GBR.CPI_ONS_CP01152_* +9 more | whole series | base_break, outlier_value |
| Not defects: the tail of a distribution the thresholds were not calibrated on. Both checks fire beyond 20% -- outlier_value at 20% from the midpoint of a month's neighbours, base_break at a 20% one-month fall -- and both numbers were tuned on the IMF's COICOP DIVISIONS. The ONS publishes four levels below that, and volatility rises monotonically with depth. Measured over every monthly index cell held, as deviation from the midpoint of its neighbours: median 0.15% at the UK all-items index, 0.21% at division, 0.25% at group, 0.35% at class and 0.55% at subclass, with the 99.9th percentile rising from 1.4% to about 25%. A fixed 20% rule therefore selects roughly the top 0.1% of ordinary movement at subclass level whether or not anything is wrong; at division level the same rule was near-silent, which is where its precision came from. The individual shapes agree. Binding services and e-book downloads is flagged seven times between 2017 and 2026, spiking 21% to 67% and returning within a month or two every time -- seven events is not an event, it is a thin sample. Margarine does the same four times (109 -> 81 -> 106, and 114 -> 83 -> 106), always returning. Air and sea passenger transport are seasonal. In every case the parent category absorbs the movement and the all-items index moves by less than 1%. Neither defect the checks look for is present: a base change is permanent and these return; one bad cell happens once and these recur. Making the threshold depth-aware is the fix, and it is ROADMAP item 16. | |||||
arg-bienes-servicios-2018-08 | 2026-09-02 | confirmed | ARG.CPI_INDEC_*_BIENES_*, ARG.CPI_INDEC_*_SERVICIOS_* | 2018-08-01, 2018-09-01 | outlier_value, base_break |
| One bad month across a whole family. Every one of the 14 series in INDEC's goods/services split -- Bienes and Servicios, in all seven regions -- reads far too high in August 2018 and returns completely in September. The size runs from 1.15x the midpoint of its neighbours (Nacional servicios, 159.95 -> 191.67 -> 173.01) to 1.98x (Patagonia bienes, 143.03 -> 299.88 -> 160.17), ordered by region rather than scattered. Three things establish it is a defect rather than a price movement or a change of base. It is not permanent: every series is back on trend the following month, which no re-referencing is. It is not in the prices: the twelve COICOP chapters for the same month are undisturbed, the largest moving 1.047x, and goods and services are composed FROM those chapters -- they cannot both be right. And it is not ours: the value was read back from the live API for 147.3_ISERVICNIA_DICI_T_23 on 2026-09-02 and it returns 302.644 for August 2018, the same number this database holds. August 2018 sits inside Argentina's currency crisis, which is presumably why a wrong figure went out unnoticed; the mechanism is not recoverable from the data held. The headline index and the COICOP chapters are unaffected and can be used across the date. The base_break flag on GBA bienes in September 2018 is the same event seen as the recovery, which is what base_break does with a spike. | |||||
usa-fresh-produce-volatility | 2026-09-02 | suppressed | USA.CPI_BLS_US_*_SEFL_*, USA.CPI_BLS_US_*_SEFL01_*, USA.CPI_BLS_US_*_SEFL02_* +4 more | whole series | base_break, outlier_value |
| Seasonal produce doing what seasonal produce does, and the second independent confirmation of ROADMAP item 16. Thirty flags across sixty years on lettuce, tomatoes, potatoes, bananas, eggs, fresh vegetables and fresh whole chicken -- 1935 to 1996, scattered, no two in the same month, every one of them a spike or a trough that reverses. Neither defect the checks look for is present: a base change is permanent and these return, and a single bad cell happens once where these happen every few years. The measurement settles it. Deviation from the midpoint of a month's neighbours, over the whole history held: the all-items index has a MEDIAN of 0.13% and a worst month of 2.4% in 1,360 months; lettuce has a median of 4.32% and a worst of 45.9%, tomatoes 4.23% and 47.7%, potatoes 1.59% and 36.0%. Lettuce's typical month therefore sits 33 times further from its neighbours than the headline's does. A fixed 20% rule cannot mean the same thing for both -- on the headline it would never fire at all, and on lettuce it selects ordinary seasonality. The UK showed this at COICOP subclass depth on the same day; this is the same finding from a different source and a different classification. | |||||
usa-energy-price-collapses | 2026-09-02 | suppressed | USA.CPI_BLS_US_*_SETB_*, USA.CPI_BLS_US_*_SETB01_*, USA.CPI_BLS_US_*_SETB02_* +6 more | 1990-01-01, 2008-11-01, 2008-12-01, 2020-04-01 | base_break, outlier_value |
| Real commodity price collapses, and the shape proves it. In November 2008 unleaded regular fell 31.7% in a month (278.70 -> 190.22) and fell another 21.2% in December; midgrade, premium, gasoline all types, motor fuel, other motor fuels and energy commodities all fell together, in both the adjusted and unadjusted series, which is 16 flags for one event. That is the signature of a commodity moving, not of a base change: a re-referencing lands on one series and its own subtree, while this lands on every series that contains gasoline and on nothing that does not -- the all-items index falls 1.7% in the same month, which is what a 30% fall in a 5% weight looks like after dilution. The level also keeps falling rather than stepping once and holding. Same story for fuel oil in January 1990, which spiked on the invasion of Kuwait and gave it back, and for gasoline in April 2020. | |||||
usa-item-events | 2026-09-02 | suppressed | USA.CPI_BLS_US_*_SS62032_*, USA.CPI_BLS_US_*_SSFV031A_*, USA.CPI_BLS_US_*_SEEE03_* | 1999-05-01, 1999-06-01, 2021-01-01, 2022-11-01 | base_break, outlier_value |
| Three real events, each in one item. Admission to sporting events fell 22.6% in January 2021, when stadiums were closed and the index was being priced off whatever was still on sale. Food at elementary and secondary schools fell 24.1% in November 2022, at the end of the universal free school meals programme, and the same index had already been suspended for five months in 2020 when schools shut -- see usa-item-suspensions. Internet services fell 25.0% in May 1999 and kept falling, which is the dial-up price war; BLS had introduced the index only in December 1997. None of the three is a base change: each is confined to one item while its parents move normally. | |||||
usa-early-quarterly-cadence | 2026-09-02 | suppressed | USA.CPI_BLS_US_* | 1913-12-01, 1914-12-01, 1935-03-01, 1936-09-01, 1951-12-01, 1963-03-01, 1963-12-01, 1968-05-01 | frequency_gap |
| Not gaps in a monthly series: a monthly series that used to be quarterly. 47 items are affected and the cadence is visible in the periods themselves -- SAF116 runs 1952-12, 1953-03, 1953-06, 1953-09, 1953-12, 1954-03 and so on, a clean three-month step, before becoming monthly later. SA0L1 starts 1935-03, 1935-07, 1935-10, 1936-01: quarterly with the ragged spacing of the period. BLS files all of them as periodicity 'R' (monthly) with period codes M03, M06, M09 and M12, so they are stored as monthly series with holes, which is what they are -- the value in M03 1953 is a March observation, not a first-quarter average, and converting them to quarterly here would be exactly the frequency conversion this system does not do. The check is right that the data is absent and right that it resumes; the answer is that the source published less often then, which is not a suspension and not a defect. | |||||
usa-item-suspensions | 2026-09-02 | suppressed | USA.CPI_BLS_US_*_SS62032_*, USA.CPI_BLS_US_*_SS62011_*, USA.CPI_BLS_US_*_SS10011_* +30 more | whole series | frequency_gap |
| Individual items BLS stopped publishing and later resumed, correctly reported. 46 flags over 33 item codes from 1998 to 2025, and the gaps are real in the source: checked against the raw flat files, BLS simply publishes no row for those months. Leased cars and trucks is absent for 16 months from January 2021, when the used-vehicle market made the lease-residual pricing unusable; food at elementary and secondary schools is absent for five months from March 2020, when schools closed; admission to sporting events for six months from June 2020. The recent ones are worth watching rather than forgetting: services by other medical professionals has been absent for nine months from June 2025, and inpatient hospital services and motor vehicle body work for three from September 2025. That is a cluster in the most recent data rather than a historical curiosity, and if it grows it is a coverage story rather than a quality one. Suppressed because each is the source publishing nothing, which the check reports accurately and which no reprocessing here can change. | |||||
imf-bra-divisions-end-2019 | 2026-09-02 | confirmed | BRA.CPI_CP* | whole series | stale_breakdown |
| The IMF's COICOP breakdown for Brazil ends in July 2019 while the IMF's own Brazilian headline runs to June 2026 -- 83 months apart, in one dataset, for one country. Eleven divisions, 33 series, every one of them ending in the same month, and all eleven start in January 2002. CP01, food, is worse than stale: it has never carried a single value. Verified in the raw archive rather than inferred from what we stored -- of the twelve archived IMF snapshots, six contain a CP01 file and in all six Brazil's series block holds 559 Obs elements and not one OBS_VALUE. The same check on CP02 in the newest snapshot finds 620 of 968 carrying values, so the reader is not an artefact and the connector is not dropping anything: the source publishes the shape of a series and no numbers in it. Confirmed against the live IMF API on 2026-09-02 (item 17), which returned the same 559 empty observations. Nothing here can be repaired by refetching or reprocessing, and no other rule reports it: the series are still published every week, so last_seen_at is right that nothing was withdrawn, and the breakdown has a span rather than a gap, so frequency_gap is right to stay quiet. The answer for a reader who wants current Brazilian category detail is IBGE's own classification, which is why item 17 became the argument for national connectors. | |||||
ibge-ipca-yoy-needs-twelve-months-of-its-own-table | 2026-09-06 | suppressed | BRA.CPI_IPCA_*_YOY.M | 2019-12-01 | frequency_gap |
| 398 flags, all the same fact: IBGE splits the IPCA category breakdown into one table per era, and a 12-month rate needs twelve months of its own table before it can be computed. Table 1419 ends December 2019 and table 7060's 12-month variable begins December 2020, so no category year-over-year is published for January to November 2020 -- by either table, because the months it would compare straddle the boundary between them. It is IBGE's convention rather than an accident, and the proof is that it happens at every table start: 444 category year-over-year series begin 2012-12 (1419 starts 2012-01), 58 begin 2020-12 (7060 starts 2020-01), and the pattern repeats wherever a category enters the basket mid-table -- 10 begin 2018-05, 9 begin 2014-01. Not a parse failure, and this is what separates the two: the monthly rate and the weight for the SAME categories are continuous across the same boundary -- 463 categories in December 2019 and 456 in January 2020 for each -- while the 12-month rate has 463 in December 2019 and zero in January 2020. The categories are all there; one variant of them is not. Nothing is derived to fill it. Chaining the monthly rates would reproduce the missing figures to about 0.005 pp, but the connector contract refuses a derived id the source also publishes, and this id is published -- by 1419 before the hole and by 7060 after it. One series whose middle is ours and whose ends are the source's is the thing that rule exists to prevent. | |||||
ibge-ipca-basket-changes-2012 | 2026-09-06 | suppressed | BRA.CPI_IPCA_1106022_*, BRA.CPI_IPCA_1107097_*, BRA.CPI_IPCA_1108009_* +10 more | 2011-12-01 | frequency_gap |
| Thirteen products that were in the IPCA basket until December 2011, left it when the basket was restructured for table 1419, and came back later: cable television, a notary's fees, a driving school, an adult-education course, melon, salt cod, fermented milk, annatto, ammonia cleaner, decorative stone, rib eye, whiting and mullet. 26 flags, two variants each -- the monthly rate and the weight, which is every variant these categories have before 2012. The dates say what happened: rib eye returns January 2014, the two fish and the driving school May 2018, and the other nine January 2020 -- each at the start of a table or partway through one, never at an arbitrary month. A basket is revised and a product that stops being representative leaves it; cable television re-entering in 2020 is a consumption pattern changing, not a value going missing. The check is right that the months are absent and right that the series resume. What it cannot know is that the absence is the source's own basket rather than a source that stopped working, and no reprocessing here can fill months in which IBGE priced nothing. | |||||