For twenty-three years, EIA published a weekly petroleum analysis every Wednesday. The essay stopped in January 2024. The tables carried on alone until 29 October 2025, and then they stopped too. We rebuilt that final edition from the published series to see whether an independent platform could stand where it stood.
Replication run 10 August 2026 · every number below is reproducible against the live server
The exercise is deliberately unflattering to us: take a report written by the agency that publishes the data, and try to reproduce it exactly, from the same series, without seeing it first. No interpretation, no rounding to hide behind — the same eight headline numbers, plus their week-on-week and year-ago changes.
The report engine was asked for the standing dashboard as of 27 October 2025, and told to read nothing published after that date. Here is what came back, beside what EIA printed:
| Metric | EIA published | EnergyScope | Difference | Dateline |
|---|---|---|---|---|
| Crude oil stocks | 416.0 mb | 415.966 | −0.034 | 24 Oct |
| Total gasoline stocks | 210.7 mb | 210.738 | +0.038 | 24 Oct |
| Distillate stocks | 112.2 mb | 112.189 | −0.011 | 24 Oct |
| Propane stocks | 105.681 mb | 105.681 | exact | 24 Oct |
| Retail gasoline | $3.035 | 3.035 | exact | 27 Oct |
| Retail diesel | $3.718 | 3.718 | exact | 27 Oct |
| Heating oil, residential | $3.596 | 3.597 | +0.001 | 27 Oct |
| Propane, residential | $2.435 | 2.434 | −0.001 | 27 Oct |
Eight of eight, inside EIA's own rounding. The week-on-week and year-ago changes match too — propane stocks +2.506 on the week and +4.151 on the year, exactly as printed.
The differences are not errors — they are resolution. EIA publishes stocks rounded to 0.1 million barrels. The underlying series are reported in thousands of barrels, which is what we hold and what the engine returns. Where the table above shows −0.034, that is the report we are reproducing having less precision than the data it was built from.
One detail worth pausing on: the two datelines. TWIP always carried two — supply data as of the Friday-released Wednesday week-ending, retail prices as of the Monday survey. We did not encode that convention anywhere. The engine was given a single as-of date, 27 October, and told to take the last observation at or before it for each series. The supply rows landed on 24 October and the retail rows on 27 October by themselves, because that is what the data does. The convention was never a convention; it was a consequence.
The engine reads nothing published after the as-of date, so a report generated for a past week reproduces what was knowable then. That is what makes the comparison fair rather than a lookup.
What it does not do is reconstruct the original vintage. We serve today's values for those historical weeks, and EIA revises. For weekly petroleum series the revisions are small — the match above is itself evidence of that — but a claim of true point-in-time replay would require storing every vintage as published, which we do not. That limitation is stated here for the same reason it is stated on our methodological background page: the reader should hear it from us.
The same call runs against this week. In Excel it is one formula:
=ES.Report("eia_weekly") ' the current standing dashboard
=ES.Report("eia_weekly", "2025-10-27") ' the final TWIP edition, rebuilt
It returns the table — stocks, supply, retail and spot prices, each with its week-on-week and year-ago change — into the sheet, where an analyst can put their own columns beside it. And because the workbook knows when each series was last updated, the sheet can tell you the new WPSR has landed before you refresh, rather than after.
What we are not claiming. TWIP was written by the agency that runs the surveys, with access to context — refinery outage detail, survey colour — that we would be inferring. We are not the official publication returning. We are showing that the numerical spine of it can be rebuilt from published data, verified against the original, and kept running afterwards.
Every figure here comes from one call against the public server. Point any Arrow Flight client at it, ask for the same preset and date, and compare:
action: report
body: {"preset": "eia_weekly", "as_of": "2025-10-27"}
Everything above is built from EIA's own published series, so it can be checked against the originals rather than taken on trust (all links verified 2026-08-10):
The original is still online — read them side by side: EIA's final This Week in Petroleum tables, 29 October 2025. That page is where the "EIA printed" column above comes from; it is EIA's own archive, not our transcription of it. If our numbers and theirs ever disagree, we want to know: [email protected].
How the signals sitting on top of this data are defined and scored: Methodology & Calibration. Where the constructs come from: Methodological background.