3-2-1 crack spread (3:2:1, 321) vs refinery utilization, 1990–2026: the link that changed sign in 2020

Latest observation

Summary

The record in six numbers

What the tight-market story says

The standard reading joins the two series in one sentence. When refineries run close to their limit, product markets are tight, and a tight product market is what a wide crack spread measures; when the fleet is idle, products are plentiful and the margin is thin. Eco3min’s own guide to refining margins as a driver of oil profits states it in that form, a utilization rate held above roughly 90% for months pointing to a tight market conducive to high margins, and the utilization dataset page repeats it: high utilization tends to coincide with wide 3-2-1 crack spreads because both reflect tight product markets. The reading is intuitive, it is what the summer of 2026 looks like, and this page does not caricature it.

It is also, on the record, a description of the last six years. Between the first joint week of November 1990 and the end of 2019, 1,385 weeks, the level correlation of utilization and the crack is −0.177, and it is negative in each decade taken separately: −0.255 in the 1990s, −0.185 in the 2000s, −0.081 in the 2010s. The 350 weeks since January 2020 return +0.560. Put together, the 1,735 weeks give 0.025, a number that says the two arms cancel rather than that the two series are unrelated. Two arms that cancel is a regime statement, and separating regimes is what the regime view of commodities over the economic cycle is for.

Two facts from the file frame everything below. First, the 13 weeks in which utilization fell by 7 points or more from the week before, all of them named hurricanes or freezes, saw the crack rise in the same week in 12 cases, by a median of $2.59 against a median of $0.003 in the other 1,622 consecutive weeks. Second, the weeks in which the fleet ran hardest before 2020 were not wide-margin weeks: the 99 weeks at or above 96% utilization between 1993 and 2019 had a median crack of $4.83, and the highest reading among them, $22.19, dates from the week of August 7, 2015.

Two frequencies, two signs

The 0.025 of the full sample is the sum of two co-movements that run in opposite directions and that dominate at different frequencies. Week to week, the relationship is negative and always has been. On the 1,635 consecutive week pairs of the file, the correlation of the change in utilization with the change in the crack is −0.137; it is −0.091 in the 1990s, −0.239 in the 2000s, −0.151 in the 2010s and −0.039 since 2020. The mechanism is the one the outage weeks make visible: a plant that stops does not stop the cars, so the product it no longer makes has to be priced out of the market, and the margin on every other barrel widens. The response is immediate. The cross-correlation of the weekly changes peaks at lag zero; at every lead or lag from one to eight weeks it stays between −0.072 and +0.103.

Quarter to quarter, the relationship is positive and always has been, weakly before 2020 and three times more strongly since. On non-overlapping quarterly averages the correlation of the changes is +0.139 on the 117 quarters before 2020 and +0.363 on the 27 quarters since; on annual averages it is +0.120 and +0.662. This is the other arm: over a season, refiners run harder when product demand pulls and ease off when it does not, and the margin moves the same way. Before 2020 the fleet had room to answer that pull, so the extra runs capped the margin; since 2020 it has had less, and the same pull shows up in both series at once.

The level correlation is where the era matters most, because levels carry both arms plus the trend of each series. Before 2020 utilization drifted up through the 1990s and down through the 2000s while the crack did the opposite, which is why the level correlation of 1990–2019 is −0.177 on 1,385 weeks; its block-bootstrap 95% interval, on 13-week blocks, runs from −0.305 to −0.050, entirely below zero. Since 2020 the level correlation is +0.560 on 350 weeks, interval +0.345 to +0.678; a null of no association, obtained by resampling the two series independently in blocks, spans −0.329 to +0.295 on that sample size. Without the COVID year, on the 298 weeks since January 2021, it is +0.411, interval +0.103 to +0.510. The Spearman rank correlations are −0.198 before 2020 and +0.578 since.

The dated turn is in the file’s own columns. Take each week’s utilization minus the mean of the same ISO week over the sample, the seasonal gap, and each week’s crack minus the median of the 52 previous observed weeks, the trailing deviation, and correlate the two over rolling 156-week windows. The correlation bottomed at −0.554 in the window ending October 25, 2019, crossed zero in the window ending May 8, 2020, the week in which operable capacity reached its 18,976 thousand barrels a day peak, and reached +0.581 in the window ending September 25, 2020. It has been positive on every one of the 332 windows since. Before 2020 it was negative on 59.4% of 1,178 windows, and its longest positive run lasted 154 windows, from the window ending December 14, 2001.

The thirteen drops

Thirteen weeks in the file show utilization falling by 7 points or more from the preceding week, on consecutive observations. Every one of them is a hurricane or a freeze on the Gulf Coast, where close to half of US capacity sits, except the first, in December 1990, for which the file records no cause. In 12 of the 13 the weekly-average crack rose in the same week; the median rise is $2.59, the mean $4.95, and the largest, in the week ending September 2, 2005 after Hurricane Katrina’s landfall on August 29, is $30.72, from $11.12 to $41.84, the largest one-week increase of the entire record. The week of Harvey, ending September 1, 2017, added $10.39 on the weekly average and saw a daily high of $36.59. The peak within eight weeks of the drop arrives, in median, two weeks after it, and in 5 of the 13 cases in the drop week itself.

Two features of the table matter for reading the utilization print. The margin moves before the print does: the EIA reports the week ending Friday on the following Wednesday, and by then the spot legs have traded through the storm, which is why the crack’s response is complete in the drop week in 5 of the 13 cases and the weekly cross-correlation shows nothing beyond lag zero. And the size of the drop does not set the size of the response: Rita’s 16.9-point drop, the largest with Harvey’s, added $3.12 to a margin that Katrina had already tripled four weeks earlier, and the two Uri weeks together removed 27.1 points of utilization for a combined $4.01. The threshold itself is a choice, disclosed: at 5 points the table has 22 weeks and a median same-week rise of $1.21; at 10 points it has 8 weeks and $3.53. Each of those drop weeks is also a week in which crude stayed in the tanks, which is why the weekly US crude inventory dataset reads as the counterpart of the utilization print.

The gradient by decade, and what it concedes

The strongest objection to the headline is the era. The median crack of all weeks was $8.00 before 2020 and $23.01 since, so a comparison of the fullest weeks across the two eras, $4.98 against $43.34, carries that shift as well as anything about utilization. The objection is right about the level and wrong about the order. Within 1990–2019 the bucket at or above 95% is the lowest of the five, below the 85 to 90% bucket by $4.64 and below the 90 to 95% bucket by $1.83; within 2020–2026 it is the highest, above the 85 to 90% bucket by $21.60 and above the 90 to 95% bucket by $15.64. The era moved every bucket up; it did not have to reverse their order, and it did.

The concession that remains is the decade table. Inside each pre-2020 decade the gradient is not negative but flat: in the 1990s the 85 to 90% bucket reads $3.84 and the at-or-above-95% bucket $3.62; in the 2000s, $7.59 and $7.23; in the 2010s, $18.12 and $18.60. The negative gradient of the pooled 1990–2019 sample comes from composition, because the 1990s, the decade of the thinnest margins, supplied 80 of the 172 weeks at or above 95%. What is robust across every cut is the absence of any positive gradient before 2020: in no pre-2020 decade did the fullest weeks earn more than the 85 to 90% weeks by more than $0.48. Since 2020 the gap is $21.60, and it is $20.91 on the 2021–2026 sample that leaves out the COVID year, on the same 28 weeks.

Three more disclosures belong here. Moving the era boundary to January 2019 gives pre and post medians of $4.84 and $42.20 for the fullest bucket, on 33 post weeks; moving it to January 2021 gives $4.98 and $43.34 on 28. Removing 2011–2013, when a WTI discount to seaborne crude inflated the Cushing-based crack without any product scarcity, a bias the anatomy of the 2022–2023 refining episode documents, moves the pre-2020 level correlation from −0.177 to −0.052 and leaves the fullest-bucket median at $4.98. And the seasonal gap computed on a trailing five-year norm rather than the full-sample norm, the real-time variant, agrees with the full-sample version against the detrended crack in the 2000s (−0.011 against −0.062), the 2010s (−0.251 against −0.250) and the 2020s (+0.274 against +0.311), and disagrees in the 1990s, where data gaps and the decade’s utilization uptrend make both versions unstable (+0.290 against +0.151). We report both rather than the flattering one.

Align any week against Katrina, Harvey or 2026

The table above quotes thirteen weeks. The file has 1,735, and the reader’s own week of interest is unlikely to be on the list. The module below draws the utilization series as a navigator; pointing at any week, or moving with the arrow keys, and pressing Enter or clicking adds it as an episode, and the fourteen presets add the thirteen drop weeks and the record crack week of August 28, 2026. Up to three episodes are aligned on a common axis from eight weeks before to twelve weeks after: utilization in level, and the crack as its change from the week before the aligned week, which is the only transform the module applies. The default view aligns Katrina, Harvey and the 2026 record week: two margins that jumped when throughput collapsed, and one that set a record while throughput stood at 98.0%.

The module reads the published CSV directly. If it does not load, the aligned-episodes chart above is the static equivalent of its default view, with Uri and Elliott added.

What each utilization bucket was followed by

The natural benchmark for a refining margin is the margin itself six months and a year later. The table classifies each week by its utilization bucket and reports the median change of the weekly-average crack 26 and 52 calendar weeks ahead, in dollars a barrel, the interquartile range at 52 weeks and the share of positive 52-week changes, on the weeks for which the forward week exists in the file. A drawdown column is not reported: a spread that crosses zero has no meaningful percentage drawdown. Windows overlap, so adjacent weeks share most of their forward path and the effective sample is far smaller than the counts.

The below-80% rows are the demand-shock weeks of 2020 and the outage weeks, followed by the recoveries that ended them; the 2020–2026 row of that bucket, 41 weeks, is almost entirely the spring of 2020. The post-2020 fullest bucket has 12 weeks with a completed 52-week window and an interquartile range of $45.48, which is a caveat, not a finding. After the 13 drop weeks the median 26-week change is −$0.87. Past distributions are not predictive of future outcomes. Regime-conditional statistics describe historical patterns, not expected returns.

Levels to watch

Era, decade and frequency tables

Median weekly-average crack, dollars a barrel, by utilization bucket and period, with the number of weeks. The buckets are left-closed: 90 to 95% contains 90.0 and excludes 95.0.

Correlation of utilization and the crack by period and frequency. Weekly changes use consecutive weeks only; quarterly and annual figures use non-overlapping calendar averages; the seasonal-gap column correlates the utilization gap to its ISO-week mean with the crack’s deviation from its trailing 52-week median.

Historical turning points

Week ending August 28, 1998, the utilization record. Utilization printed 100.5%, the highest of the file, on 15,761 thousand barrels a day of capacity; the weekly crack was $2.45, in the 7th percentile of the record to that date. The fleet was running above nameplate and earning almost nothing on the Harbor margin: the 1990s in one week.

Week ending September 2, 2005, Katrina. Utilization fell from 97.1% to 86.8% and the weekly crack rose from $11.12 to $41.84, a one-week increase of $30.72 that no other week of the record approaches; the next largest, in the week of April 29, 2022, is $16.93. Four weeks later Rita took utilization to 69.8% and the crack was $21.29.

Weeks ending April 3 and April 17, 2020, the demand collapse. Utilization fell to 75.6% and then 67.6% while the weekly crack read $6.96 and $10.08; the low of the year, $4.76, came in the week ending May 29. Both series fell together, and this is the first large episode of the file in which they did: the demand arm, not the supply arm.

Week ending May 20, 2022. Utilization 93.2%, weekly crack $60.49, daily high $71.74, the record until 2026; the highest weekly average of 2022, $67.49, was the week before, ending May 13. Operable capacity stood at 17,941 thousand barrels a day, 1,035 below the May 2020 peak.

Week ending August 28, 2026, the current observation’s neighbourhood. Utilization 98.0%, the highest since the 98.1% of the week ending July 1, 2005, and a weekly crack of $68.03, the highest of the record, with the daily record of $73.84 falling on August 31 in the following week. Gross inputs were 17,661 thousand barrels a day, below the 18,243 of August 10, 2018, on a capacity base 949 smaller than in May 2020. The two series stand at their joint extreme for the first time in the file.

Methodology

The crack is the site’s daily 3-2-1 convention recomputed from the three EIA spot legs, with the product legs converted from dollars a gallon to dollars a barrel at 42 gallons, then averaged over the EIA reporting week ending Friday. The utilization rate is the EIA’s weekly percent utilization of refinery operable capacity, joined on the same Friday. A panel week exists when both series exist; 123 of the 1,858 utilization weeks are dropped because at least one spot leg has no trading day in the week, 122 of them between 1993 and 2006. Weekly changes are computed on consecutive observed weeks only.

Buckets and eras. Five utilization buckets with left-closed bounds at 80, 85, 90 and 95%. The era boundary is the first weekly observation of 2020, January 3; it was chosen before the rolling correlation was computed, and the rolling correlation’s own zero crossing, May 8, 2020, falls 18 weeks after it. Sensitivity to boundaries of January 2019 and January 2021 is reported in the gradient section. Outage weeks are consecutive observations with a utilization change of −7.0 points or less; the threshold sensitivity is reported in the drops section. The Hurricane Georges week of October 2, 1998 shows a 10.4-point fall against the previous observation, but that observation is three weeks earlier because the product legs have a gap, so it is not flagged. Correlations are Pearson unless stated; block-bootstrap intervals resample 13-week blocks 2,000 times, and the null band resamples the two series independently. Quadrants classify each week by the sign of the seasonal gap and of the trailing deviation: the same-sign share, hot-wide plus slack-thin, is 51.1% of 1,333 weeks before 2020 and 57.7% of 350 since, a reminder that at weekly resolution the two series are close to independent in either era. Forward changes are the crack 26 and 52 calendar weeks ahead minus the crack that week, on the weekly grid, NaN when the forward week is absent or not yet observed.

Reproducibility. The file and the three scripts that build it, compute every statistic on this page and audit each number against the file are the deliverables of this study. In Python:

The PADD 3 utilization rate, available weekly from June 2010, was used as a control and is not republished: its level correlation with the national rate is 0.922 on 849 weeks and its weekly-change correlation 0.881, so the national series carries the Gulf Coast outages that the drop table lists. The site’s record of every US refinery shutdown since 1990 and its list of closures since the January 2020 capacity peak document the denominator; this page takes the weekly operable capacity as published.

Sources and references

• US Energy Information Administration, Weekly Petroleum Status Report, weekly percent utilization of refinery operable capacity (WPULEUS3), operable crude distillation capacity (WOCLEUS2) and gross inputs to refineries (WGIRIUS2), 1990–2026.
• US Energy Information Administration, spot prices: Cushing, OK WTI (RWTC), New York Harbor conventional gasoline regular (EER_EPMRU_PF4_Y35NY_DPG), New York Harbor No. 2 heating oil (EER_EPD2F_PF4_Y35NY_DPG), daily, 1986–2026.
• US Energy Information Administration, Refinery Capacity Report, annual, and “Copyrights and Reuse”, public domain statement.
• National Hurricane Center, tropical cyclone reports for Lili (2002), Ivan (2004), Katrina and Rita (2005), Gustav and Ike (2008), Harvey (2017) and Ida (2021); US Department of Energy, “Comparing the impacts of the 2005 and 2008 hurricanes on US energy infrastructure”, 2009.
• Borenstein, S., Cameron, A. C. and Gilbert, R., “Do gasoline prices respond asymmetrically to crude oil price changes?”, Quarterly Journal of Economics, 112(1), 1997, on the product-price side of the margin.
• Kilian, L., “Not all oil price shocks are alike: disentangling demand and supply shocks in the crude oil market”, American Economic Review, 99(3), 2009, for the identification logic of supply and demand shifts used in the sign discussion.
• Granger, C. W. J. and Newbold, P., “Spurious regressions in econometrics”, Journal of Econometrics, 2(2), 1974, on why level correlations of trending series are reported here next to their change counterparts.
• Eco3min Research, US 3-2-1 crack spread dataset and US refinery utilization rate dataset, the two series this page joins.

Limitations

• Geography. The utilization rate is national; the crack is a New York Harbor product basket against Cushing crude. A Gulf Coast outage moves the national rate through the weight of PADD 3 and the Harbor margin through the product market, but a Midwest or West Coast disruption can move one without the other.
• Gross margin, not profit. The 3-2-1 excludes operating cost, natural gas, catalyst, capital and the crude slate of any actual plant, and the EIA-spot construction differs by a basis from the RBOB-futures crack. The 20 April 2020 negative-crude day is inside the week ending April 24, 2020 and affects that week’s average and high.
• Ex-post seasonal norm. The ISO-week mean is computed on the full sample, so the seasonal gap of any week uses later data. The trailing five-year variant is reported in the gradient section and is unstable in the 1990s.
• Composition across decades. The pooled pre-2020 gradient is negative partly because the 1990s supplied 80 of the 172 weeks at or above 95%; the decade table is the honest version and shows a flat gradient inside each pre-2020 decade.
• Small post-2020 buckets. The fullest bucket since 2020 has 28 weeks, and 12 with a completed 52-week window; the era has 27 quarters and 7 calendar years.
• Data gaps and revisions. 123 utilization weeks have no matching crack, almost all in 1993–2006; two gross-inputs values, for the weeks ending August 13, 1993 and May 24, 2002, are absent from the EIA file and are derived from the utilization identity. The EIA revises weekly data and the capacity base annually; the file is regenerated from the source on each update rather than appended.

Frequently asked questions