# Data choices: us-inflation-components story **Question.** US inflation rose in 2026, to 3.7% in September. Which parts of the basket account for it, and is it broad, as in 2022, or narrow? Checked before choosing the angle: monthly contributions for January 2019 to September 2026 by five components. The data gave a clear answer: energy accounts for 99.8% of the rise from February to September; everything else contributed 2.382 and 2.384 percentage points respectively, both rounding to 2.38. A "what has risen most since 2020" ranking was considered and dropped, as a second story. **Measure.** CPI-U, US city average, **not seasonally adjusted**, 12-month percent change: the BLS headline measure. Relative importances are defined on unadjusted indexes, so contributions must use them too. Monthly seasonally adjusted moves (for example "+0.6% in September") are not used. **Contributions, not weights.** A component's contribution to the 12-month change, in percentage points, uses the BLS method (): relative importance at the start of the window, updated from the preceding December's table by the component's price change relative to all items, multiplied by the component's price change over the window. We split each window at its December link month, using the applicable weights for each segment, as BLS advises; expenditure weights were updated biennially through 2022 and annually from 2023: the first part uses the old basket; the second starts from the December table on the new basket and is scaled by the all-items change already accumulated. The December table used for year *Y* is BLS's `(Y-1).htm` (or the archive's `(Y-1).txt`), which uses the weights in force from January *Y*; its header names the weights (for example "Table 1 (2024 Weights) ... December 2025"). **Checks (in `calculate.mjs`, reported in `statistics.json`).** - The five contributions sum to the all-items change within 0.001 percentage points in every month (the script fails above 0.02). - Carrying each December basis table forward 12 months reproduces BLS's own "old weights" December table for 2021–2025 within 0.001 (the script fails above 0.01). Some old-weights pages omit special aggregates such as energy; those rows are simply not checked. **Components.** Energy (`SA0E`); food (`SAF1`, which excludes alcoholic beverages; alcohol sits in "other goods"); commodities less food and energy commodities (`SACL1E`, "other goods"); shelter (`SAH1`); services less energy services (`SASLE`) minus shelter ("other services"). Their December relative importances sum to 100.000 ± 0.001 in every year 2017–2025. Gasoline (`SETB01`) is shown as part of energy, never added to it. "Everything except energy" is all items minus energy's contribution. **Weight shown in chart 4.** Relative importance at September 2025, the start of the 12-month window, on the 2025 basket (2023 expenditure weights), as BLS's method uses. The 2026 basket (2024 weights, from the December 2025 table) puts energy at 6.383%, slightly lower than the 6.4% shown. **Dates.** The February 2026 comparison month is the last before the oil-price shock that EIA dates to 28 February 2026. January 2026 gives the same story (all items 2.39%, energy −0.01 points). Because February 2026 falls inside the rent-imputation window (below), the post also compares with September 2025, which is outside it: energy's contribution rose from 0.18 to 1.35 points while everything else fell from 2.84 to 2.38, so the finding does not depend on the February baseline. Between February and September everything except energy did not stay flat: it peaked at 2.70 points in May 2026. The June 2022 comparison is the series' maximum all-items rate in 2019–2026. **Gap.** BLS did not publish October 2025 for most series (2025 lapse in appropriations; the API returns `-` with footnote "Data unavailable due to the 2025 lapse in appropriations"). Gasoline, new vehicles and used cars have October values (non-survey data); they are kept in the extract but no window that starts or ends in October 2025 is computed, because all items is missing. Charts break their lines there; nothing is plotted as zero. **Known measurement issue, not corrected.** BLS carried October 2025 rents forward from April 2025 (rent and owners' equivalent rent unchanged in October), and said the effect would resolve in April 2026. Shelter's 12-month contributions for windows ending November 2025 to March 2026 are therefore probably understated; windows ending April 2026 or later span both the October 2025 carry-forward and the April 2026 catch-up, so, per BLS's description, the September 2026 figure should not be affected. This is stated in the post, attributed to BLS, and not quantified here. A BLS Monthly Labor Review article tested alternative imputations (); it was not read for this post. **Not used.** Seasonally adjusted series; chained CPI (C-CPI-U); PCE price index (the Federal Reserve's target measure, which weights shelter less and would give different shares); regional indexes.