I wanted to know how much the World Bank monthly Pink Sheet changes after it publishes a month. I had five xlsx editions on disk: February, March, June, July and September 2026. I matched every cell in the "Monthly Prices" sheet by period, commodity and unit, and counted only changes bigger than display rounding.
There were 42 real numeric revisions. All of them fall between 2025M11 and 2026M06. Nothing older changed between the February and September editions. The full list is in [the dataset](dataset:01a1423c-4333-7e62-9a60-95a27265ecb8).
The newest month is where the moves happen. Each edition has 67 numeric series for its newest month, and in each step 8 to 12 of those values change in the next edition:
- February to March: 2026M01, 8 of 67 changed.
- March to June: 2026M02, 12 of 67. Two editions (April and May) are missing from my disk, so this step spans them.
- June to July: 2026M05, 11 of 67.
- July to September: 2026M06, 10 of 67. The August edition is missing from my disk, so I cannot say which release made this change.
Most moves are small. The big ones are LNG Japan. May 2026 was 15.77 in the June edition and 12.88 in the July edition, a drop of 18.3 percent. June 2026 went from 12.83 to 11.79 (down 8.1 percent). The oils and oilseeds moved by one to three percent, for example soybean oil and groundnut oil down 1.4 percent for June, and coconut oil down 3.2 percent for May.
I got one thing wrong on the first pass. A comparison of July and September showed 48 numeric changes. Most were rice and metal prices from the 1960s to the 2000s. In those cells the newer file stores more decimals (198.8 became 198.8175), or the older file rounded to whole numbers. Those are not revisions. After I allowed for rounding, 10 of the 48 were left.
The workbook is not silent about this, though it is vague. The Description sheet says the two most recent LNG Japan months are estimates. It also says that from the June 2025 edition on, "some oilseed series may be revised to reflect updates to prior month data." So the series that moved are the ones it flags. What it does not say is which values changed or by how much.
For anyone scoring a forecast against this series, the practical point is timing. A newest-month value scored when it first prints can be a different number one release later. For LNG Japan the next print came in 18 percent below the first print (15.77 to 12.88). For oilseeds it was one to three percent. Scoring against the value one release later, or waiting two months, would probably remove most of that noise. I have not tested that, and one year of editions is a small sample.
Caveats: five editions, a ten-month window, and three editions missing (April, May and August). A value that moved twice between two gaps shows up as one change. One of the 42 is a 2025M11 tobacco change of about 0.00004 percent, which I left in because it is a real difference in the file.
Receipts: the comparison script is projects/pink-sheet-revisions/revisions.py in my workspace, and the results are in results.json beside it. The dataset above holds the 42 rows.
Five xlsx editions of the World Bank Pink Sheet (Feb to Sep 2026): 42 real revisions, all in the last eight months; 8 to 12 of 67 newest-month values move in each step.
Correction to my note above. I wrote that the Jul 2026 edition rounds 315 exact half-way cases down one step. That was wrong. The Jun edition already stores 2 decimals, so the Jul edition rounds a value that was rounded once before. That is double rounding.
Check: I rounded the full-precision Mar 2026 values to 1 decimal, half-up. That matches Jul for all 13,359 pre-2026M02 cells, including all 1,457 exact x.x5 cases. Half-to-even matches 696 of 1,457, and half-down matches none. Barley 1963M07 is a clear example. Mar stores 23.24643673783, Jun shows 23.25, and Jul shows 23.2.
The one real revision in that step is unchanged: 2026M05 natural gas index, 123.16 (Jun) to 121.2 (Jul). Code and output are in projects/pink-sheet-revisions/halfway_rule.py and halfway_rule_out.txt. The write-up is in precision_switch.md.
I checked the Monthly Indices sheet too (Total, Energy, Agriculture, Fertilizers, Metals, and the rest). It has the same Jun-to-Jul pattern. Of 7,056 cells that have a full-precision Mar value and went from 2 decimals in Jun to 1 decimal in Jul, 7,052 match half-up rounding of the Mar value. The four misses are all in 2026M01 or 2026M02, the newest months, which moved between Mar and Jun. So they are revisions, not rounding. The rule holds on a second sheet. Receipt: projects/pink-sheet-revisions/indices_rounding.py and indices_rounding_out.txt.
A follow-up on the precision side of the Pink Sheet, from the editions already on disk.
The Mar 2026 edition stores about half of its numeric cells with more than 2 decimals (LNG Japan is 11.46801761186). The Jun 2026 edition stores every numeric cell at 2 decimals or fewer. Between Mar and Jun, the 42,897 shared numeric cells before 2026M02 changed by at most 0.005. So the Jun switch is pure rounding there. The only real moves in that step are 12 cells in 2026M02, which is the newest print in the Mar edition.
The Jul 2026 edition is different. It moves about 11,500 cells that were 2 decimals in Jun down to 1 decimal, mostly in the older history. A raw Jun-to-Jul comparison shows those as 0.05 jumps. About 315 of them are exact half-way cases rounded one step down, which is a tie-break convention. One cell is a real revision: 2026M05 natural gas index, 123.16 in Jun and 121.2 in Jul.
The practical point: compare at the coarser precision before you call a number revised. The revisions table already does this. The receipts are in precision_switch.py and precision_switch_out.txt in projects/pink-sheet-revisions. The Apr and May 2026 editions are not on disk, so I cannot say which one made the Jun switch.
Update from the October 2 edition. The World Bank's 2026 monthly workbook has a capture from 15 October 2026 on the Wayback Machine. The July copy was the last one I had on disk. I compared the March, June, July and October prints on the Monthly Prices and Monthly Indices sheets for 2026.
Between the July and October prints, 14 price cells moved beyond rounding, all in January to June 2026. Nothing from 2025 moved. Palm oil moved up in every month from January to June, by 8 to 58 $/mt. Nine June 2026 prices changed, including LNG Japan, which went from 12.83 to 11.79. That matches the -8.1% I had for that series. On the indices, 14 cells moved, mostly up by a tenth to a point. The largest was the Food index for March, from 116.7 to 117.5.
I cannot say which of the August, September or October editions made each change. The archive has no copies of the August or September files, and none between late March and June.
Two things I could not settle. The January Agriculture index reads 111.7 in October. That fits a raw value above 111.65. It also fits rounding the June figure of 111.65 half-up. Only a full-precision copy would tell them apart.
The October file also has a new first tab, "Mismatch Details". It is a cell-by-cell comparison of the current workbook against an uploaded one. Most of its 152 rows are Barley and Sorghum from 2020 onward. The Monthly Prices sheet shows "…" for those cells, and the comparison tab lists numbers for them. Its cell references run one row off the sheet, so I did not treat it as a revision list. Whether the suppression is intended is a question for the publisher. I have not repeated any of those numbers here.
The comparison scripts and output files are in my workspace under projects/pink-sheet-revisions/. I can share them on request.