VLO vs XLE after low oil-beta / falling-Brent screen (N=27 signals)
When VLO's 20-day beta to Brent fell into the bottom quintile of its prior year while crude was selling off, VLO went on to beat XLE over the following 20 sessions by an average of 7.5% — and it won 81.5% of the time. That is a large relative edge from a screen that fired on just 27 of 448 eligible days.
The setup, not energy-sector drift, appears to be doing the work: on the other 421 days the same spread averaged only 2.7%. The catch is that these signals cluster, sharing most of their underlying return days, and the year split is lopsided — a coin flip in 2024, a perfect record in 2025.
The full methodology, distribution and caveats are laid out below.
Over the past ~3 years, when VLO's 20-day realized beta to Brent crude is in the bottom quintile of its prior 252-session distribution while Brent crude's 20-day return is negative, does VLO outperform XLE over the next 20 trading sessions? I expect a refiner to decouple from falling crude when downstream cracks are driving margins, so VLO leads the oil-beta sector until crude sensitivity mean-reverts.
How this was measured
Daily closes for VLO and XLE were built from minute bars, and Brent crude daily values were aligned to common trading days by forward-fill. VLO's 20-session realized beta to Brent was estimated as the rolling covariance of VLO returns with Brent returns divided by the rolling variance of Brent returns. The bottom-quintile condition compares the current beta with the 20th percentile of the prior 252 sessions of beta, excluding the current session. The screen required that current beta to be at or below that percentile and Brent's 20-session return to be negative. Forward performance was measured as the close-to-close 20-session VLO return minus the corresponding XLE return.
The key numbers
Reading the numbers
Over the test, 27 days fit the setup. On those days VLO beat XLE by 7.50% on average over the next 20 sessions, with a 10.07% median and an 81.5% win rate. That is 4.80 percentage points better than non-signal days; the p-value says that gap is unlikely to be random noise.
The charts
This histogram shows how VLO-XLE performed over the next 20 sessions on the 27 qualifying signal days. The results are not one-way: they run from -4.9% at the worst to +18.35% at the best, so some signals did lag. But the average is +7.5% and the median is +10.1%, which means the typical signal landed well ahead, supporting the idea that VLO tends to beat XLE after this setup.
Each dot is one signal day, with VLO's realized beta to Brent on the horizontal axis and the forward VLO-XLE return on the vertical axis. The beta readings span about -0.12 to +0.22 and average roughly 0.04, which is the low-sensitivity zone the screen targets. Forward returns range from -4.9% to +18.35% and average +7.5%, showing that even inside this narrow low-beta group the outcome varies widely, but the center of the cloud is positive.
The main line is VLO's rolling 20-session beta to Brent, and the second line is the prior 252-session 20th percentile cutoff that defines the bottom quintile. VLO's beta averages 0.44 and peaks at 1.54, so the stock often moves meaningfully with crude; the cutoff averages 0.16. The screen fires when the beta line drops below that cutoff, which is the low-beta half of the condition, and the end of the chart shows beta at 0.133 versus a 0.1702 threshold, still below the line. The deepest dip reaches -0.1191, illustrating how far these decoupling episodes can go when Brent is also down.
Qualifying signal summary by calendar year
| Year | Signals | Mean VLO−XLE | Median VLO−XLE | VLO beat rate |
|---|---|---|---|---|
| 2,024 | 11 | 0.0243 | 0.0231 | 0.5455 |
| 2,025 | 16 | 0.1099 | 0.1263 | 1 |
The takeaway
Yes — VLO did beat XLE on this screen, and the average gap was big: across 27 qualifying days the forward 20-session VLO-minus-XLE spread averaged +7.5%, with a median of +10.1% and VLO ahead in 81.5% of cases. The screen is what's doing the work, not just energy-sector drift — on the other 421 eligible days the same spread averaged only +2.7%, so the setup added roughly 4.8 percentage points of relative performance, and that difference is statistically solid (Welch p ≈ 0.0014). Almost all of the gain came from VLO's own move (+11.3% average) while XLE managed just +3.8%. The honest caveat is that the headline t-stat of 5.7 flatters the evidence: the 27 signals aren't 27 independent bets, since adjacent signals share 19 of 20 return days, and crudely discounting for that overlap puts the real confidence somewhere closer to 'strong lean' than 'near-certain'. The year split clinches the caution — 11 signals in 2024 delivered only +2.4% on average with a 55% win rate, essentially a coin flip, while all 16 signals in 2025 worked, averaging +11.0% with a perfect record. Practical read: the decoupling thesis is real and directionally right, but the result is really a 2025 story, and one that only triggers on about 6% of trading days.
The fine print
- The 27 signals are not independent: overlapping 20-session forward windows mean the effective sample is far smaller than it looks, so the p-values overstate confidence.
- Everything hinges on 2025 — 2024's 11 signals were a coin flip (+2.4% average, 55% hit rate), so this may be one favorable regime rather than a durable rule.
- XLE holds VLO as a constituent, so the benchmark isn't independent of the thing being tested, which flatters the differential somewhat.
- The realized beta uses same-day Brent returns; a live screen would be a session behind, and results are specific to this ~3-year window.