AI Research PSXXOMmacro:brent_daily

PSX vs XOM 10-session relative return after Brent -2% days with PSX up

39
Qualifying days

When Brent falls more than 2% and PSX still closes green, the refiner is usually saying something about product cracks and cheaper feedstock. The tape has largely agreed. Across 39 such days in the past three years, PSX beat XOM over the following 10 sessions in two-thirds of cases, with a median relative gain north of 2 percentage points — and the median sitting above the mean argues no single blowout window is carrying the result.

That conditional edge also runs well ahead of the 0.57% average spread on the 679 ordinary days, which is the interesting part. The uninteresting part is confidence: a t-stat near 1.93 leaves roughly a 6-in-100 chance this is luck, and the 39 triggers overlap heavily, so the true number of independent bets is smaller than it looks.

The full breakdown of windows, baselines and distributions follows.

The research question

Over the past ~3 years, when Brent crude's one-day return is below -2% but PSX closes higher on the day, does PSX outperform XOM over the next 10 trading sessions? I expect a refiner that rallies on a sharp crude down-spike to be pricing feedstock-cost relief and widening product cracks, so its relative strength against the integrated major persists.

How this was measured

Resampled PSX and XOM minute bars to daily closes and intersected them on a common trading calendar. Reindexed Brent daily closes onto that calendar with forward-fill and computed one-day returns. A qualifying day is one where Brent's one-day return is below -2% while PSX's same-day close-to-close return is positive. For each qualifying day, measured the close-to-close return over the next 10 trading sessions for PSX and XOM, and the difference PSX minus XOM. Compared that conditional edge against zero using a one-sample t-test, and against the non-qualifying-day 10-session PSX-XOM baseline.

The key numbers

Qualifying days
39
Brent 1d < -2% and PSX 1d > 0, with 10-session forward available
Non-event baseline PSX-XOM 10-session mean
0.5741%
N=679 non-qualifying days
Mean PSX 10-session forward return
2.4119%
N=39 qualifying days
Mean XOM 10-session forward return
0.7425%
N=39 qualifying days
Mean PSX minus XOM 10-session return
1.6694%
Positive = PSX outperformed XOM
Median PSX minus XOM 10-session return
2.0776%
Robust to outlier events
Win rate (PSX > XOM)
66.667%
Share of qualifying days where PSX beat XOM over next 10 sessions
One-sample t-stat vs zero
1.931
Tests PSX-XOM 10-session edge against zero
One-sample p-value
0.0610
p=0.0610 >= 0.05 -> edge not distinguishable from zero at 5% level

Reading the numbers

There were 39 days matching the setup. PSX beat XOM over the next 10 sessions in about two-thirds of them, with an average edge of +1.67% versus a +0.57% baseline, but the p-value of 0.061 means that result could still be chance.

The charts

PSX-XOM 10-session forward return vs Brent trigger-day drop
What this chart says

Each dot is one of the 39 trigger days: farther left means Brent fell harder, from -2.05% to -12.35%, and higher means PSX did better than XOM over the next 10 sessions. The dots are widely scattered, from -8.7% to +16.5%, with many above zero and an average of +1.67%. So the average PSX edge is positive, but individual outcomes are very inconsistent, which helps explain why the 67% win rate does not become a statistically clean edge.

Mean 10-session forward return on qualifying days
What this chart says

This bar chart breaks the 10-session result into three averages: PSX +2.41%, XOM +0.74%, and the PSX-minus-XOM gap +1.67%. The PSX bar is taller than XOM's, which is the core of the feedstock-relief trade: on average the refiner did outperform the integrated major after those crude down-spikes. But that +1.67% gap is the same one the t-test says is not distinguishable from zero at the 5% level, so the visual separation is suggestive rather than conclusive.

Qualifying days and following 10-session returns

datebrent_1d_retpsx_1d_retpsx_10d_fwdxom_10d_fwdpsx_minus_xom_10d
2023-10-16-0.03540.01020.0052-0.03880.0439
2023-10-31-0.04310.00540.0193-0.00530.0246
2023-11-15-0.02140.00130.1293-0.01010.1394
2023-11-24-0.02370.00530.0611-0.0470.1081
2023-12-01-0.03670.0132-0.0153-0.0210.0057
2024-04-15-0.02450.0008-0.0549-0.005-0.0499
2024-05-14-0.02620-0.035-0.03540.0004
2024-05-31-0.02470.0286-0.0411-0.06930.0282
2024-07-22-0.02050.0154-0.04320.0023-0.0455
2024-07-30-0.02080.0787-0.0895-0.0024-0.087
2024-09-26-0.0310.00230.04390.09-0.0461
2024-11-08-0.02230.00720.06310.01280.0503
2024-11-11-0.0250.0080.07610.00280.0734
2024-11-25-0.0240.0204-0.0699-0.0568-0.0132
2025-01-21-0.02950.00460.0179-0.01610.034
2025-04-07-0.03260.0020.05040.04520.0052
2025-04-29-0.03160.00590.17490.00960.1653
2025-05-07-0.0330.01060.0601-0.00010.0603
2025-05-15-0.02350.0041-0.0827-0.0616-0.0211
2025-07-10-0.02220.0318-0.0655-0.0346-0.0309
2025-07-22-0.0310.0098-0.041-0.0151-0.0259
2025-08-19-0.02290.00330.06330.02030.043
2026-03-04-0.02070.02880.06320.05550.0077
2026-03-10-0.04780.00150.12990.10570.0242
2026-03-16-0.02120.00420.06930.0959-0.0266
2026-03-19-0.05960.0223-0.01210.0142-0.0263
2026-03-23-0.12350.0107-0.0966-0.041-0.0556
2026-04-08-0.11650.001-0.0438-0.04460.0008
2026-04-14-0.03720.00080.04690.01410.0328
2026-04-15-0.03170.01560.06980.03510.0347
2026-05-27-0.05490.00290.04570.02490.0208
2026-06-12-0.04520.0073-0.0207-0.07460.0539
2026-06-22-0.0360.01150.07410.03530.0388
2026-06-26-0.04850.00390.15130.0590.0923
2026-07-09-0.02670.00480.08520.1415-0.0563
2026-07-16-0.02230.02970.0470.0866-0.0396
2026-07-24-0.04760.0035-0.0069-0.03030.0234
2026-07-27-0.08460.00110.03890.02480.0141
2026-07-28-0.06870.00730.07250.02270.0499

The takeaway

Yes — but only partly, and this is a lean rather than a proven edge. Across the 39 days in the past three years when Brent fell more than 2% and PSX still closed higher, PSX beat XOM over the following 10 sessions 67% of the time, with an average relative gain of +1.67 percentage points — and the median of +2.08 says it isn't one blowout day carrying the result. In those windows PSX averaged +2.4% against XOM's +0.7%, comfortably ahead of the +0.57% average edge on the other 679 ordinary days. The catch is confidence: the reading works out to roughly a 6-in-100 chance this is luck (p = 0.061), which is suggestive but short of the bar you'd normally want before calling something a real signal. That number likely flatters the evidence, too, because the 39 triggers cluster — whenever two fire within 10 sessions the windows overlap, so the number of genuinely independent bets is smaller than 39. Practical upshot: the feedstock-relief read shows up in the data and is worth keeping on the radar as a tilt, but it's too thin to lean on as a standalone signal.

The fine print