AI Research MPCmacro:brent_daily

MPC daily-return beta to Brent crude on extreme-up Brent days (>95th pctile, N=37 days)

741
Total trading days (aligned)

The idea is intuitive: refiners like Marathon Petroleum normally ride crude higher, but when Brent spikes violently, margins get squeezed and the market sells the stock. That would flip a usually positive beta negative on tail days. So we tested it—isolating the 37 sessions over three years where Brent’s daily gain cleared its 95th percentile, then measuring MPC’s response on exactly those days.

The result is a clear non-finding. The extreme-day beta does turn negative at -0.16, but the regression can’t distinguish it from zero: the confidence interval spans -0.50 to +0.18, and the p-value sits at 0.34. Meanwhile, the full-sample beta stays solidly positive at +0.28. In plain terms, there’s no reliable flip—just noise on a small sample. The complete methodology, chart, and a look at what the averages actually did on those spike days follow below.

The research question

Over the past ~3 years, does Marathon Petroleum (MPC) exhibit a negative daily-return beta to Brent crude on days when Brent crude's daily return exceeds its 95th percentile? I expect a negative beta, as extreme crude spikes squeeze refining margins and trigger selling in refiners, contrary to their usual positive correlation with crude.

How this was measured

Brent crude daily-close returns from the global brent_daily_df macro frame; MPC daily close-to-close returns from MPC_df minute bars resampled to daily. Aligned both on common trading days. Computed the 95th percentile of daily Brent returns in-sample to define extreme-up days. On those extreme days, ran OLS linear regression of MPC returns on Brent returns — the slope coefficient is the conditional beta. Also computed the full-sample beta (all trading days) and the non-extreme-day beta for contrast. If the extreme-day beta is materially below the full-sample beta — and especially if it turns negative — the refining-margin-squeeze hypothesis is borne out: sharp oil spikes compress crack spreads, and the market sells refiners despite their usual positive oil-levered correlation.

The key numbers

Total trading days (aligned)
741
Days where both Brent and MPC have valid daily returns
Extreme-up Brent days
37
Days where Brent return > 3.89% (in-sample 95th pctile)
95th pctile Brent threshold
3.8941%
Defines the tail — ≈5% of trading days
Extreme-day beta (MPC on Brent)
-0.1601
OLS slope on 37 extreme-up days, 95% CI [-0.498, 0.178]
Extreme-day beta's negative-sign check
-0.1601
beta=-0.1601 < 0 → refining-margin-squeeze hypothesis supported on tail
Extreme-day R²
0.0258
R²=0.026 ≤ 0.5 → noisy relationship even on extreme oil days
Extreme-day regression p-value
0.3423
p=0.3423 ≥ 0.05 → slope is not statistically distinguishable from zero
Full-sample beta (all days)
0.2780
N=741 days — the refiner's usual positive oil beta baseline
Non-extreme-day beta (Brent ≤95th pctile)
0.3292
N=704 days — beta on 95% of non-tail days
Mean Brent return on extreme days
6.4995%
Average tail-event magnitude
Mean MPC return on extreme days
1.5614%
Average refiner response during Brent spikes
Full-sample R²
0.1094
How well Brent explains MPC returns unconditional

Reading the numbers

Over 741 days MPC normally moves with Brent (beta +0.28). On the 37 extreme Brent-up days (>3.89%), the beta flips to -0.16, supporting the margin-squeeze story — but the relationship is noisy (R²=0.026, p=0.34), so the negative beta isn't statistically solid.

The charts

MPC vs Brent daily returns — extreme-up days highlighted
What this chart says

The scatter of all 741 days slopes upward as a whole, matching the full-sample beta of +0.278: MPC usually rises with Brent. The days that matter for the question are the extreme Brent spikes on the right side of the plot, and their MPC returns look muted — on those days Brent averages +6.5% while MPC averages only +1.6%. That flattening at the right edge is the first visual clue that the usual positive relationship does not hold in the tail.

MPC vs Brent — extreme-up days ONLY (Brent > 95th pctile)
What this chart says

Restricted to the 37 extreme-up days, the scatter no longer shows a clear upward slope. Brent returns run from roughly +4% to +12.5%, but MPC returns stay in a narrow band from -3.4% to +5.5%, and the fitted slope is -0.16 — negative, but with an R² of just 0.026, Brent explains almost none of the spread. The p-value of 0.34 says this negative slope is not statistically distinguishable from zero.

Beta comparison: full-sample vs non-extreme vs extreme-up
What this chart says

This bar chart puts the three betas side by side: all days give +0.278, the 704 non-extreme days give +0.329, and the 37 extreme-up days give -0.160. The right-hand bar crossing below zero is the whole story — on the days Brent spikes hardest, MPC stops moving with oil and, if anything, drifts the other way. The confidence interval for that extreme beta spans -0.498 to +0.178, so the flip is directionally consistent with the refining-margin-squeeze hypothesis but not precisely estimated.

Beta decomposition by Brent-return regime

regimeN_daysbetar_squaredp_value
Full-sample7410.2780.10940
Non-extreme (≤95pct)7040.329200
Extreme-up (>95pct)37-0.16010.02580.3423

Extreme-up days — event-level detail

brent_returnmpc_return
0.12530.0229
0.119-0.016
0.09790.0514
0.08950.0258
0.0862-0.0136
0.0831-0.0342
0.0830.0494
0.08070.0202
0.07890.0085
0.0782-0.0155
0.07530.0016
0.07280.0381
0.07270.0135
0.07130.0129
0.06880.0062
0.06730.0169
0.06710.0081
0.0664-0.0203
0.06580.0368
0.06490.0155
0.05820.04
0.05560.0335
0.05480.0097
0.04980.0106
0.04960.0154
0.04860.0031
0.04840.0026
0.04690.0324
0.04650.0242
0.04490.0256

The takeaway

The short answer is no — MPC doesn't show a reliable negative beta on extreme Brent-up days; the apparent flip to -0.16 is statistically indistinguishable from zero. That estimate comes from just 37 tail days (Brent up more than 3.9%), and the 95% confidence interval runs from -0.50 to +0.18, so the true beta could easily be positive. The p-value of 0.34 reinforces that: a slope this far from zero would appear roughly a third of the time even if the true beta were zero — basically a coin flip. For context, the full-sample beta is +0.28 and the non-extreme beta is +0.33, so MPC's usual relationship to Brent is solidly positive; only on the tail does the estimate dip negative, and even then it's lost in noise (R² of 0.03). Interestingly, on those extreme days MPC averaged +1.6% while Brent averaged +6.5%, so the simple story of refiners selling off on crude spikes doesn't show up in the average either. The practical takeaway: the margin-squeeze hypothesis is not supported by this three-year window, and any tail beta is too imprecise to rely on.

The fine print