USO-to-Brent 20-day tracking error conditioned on Brent realized-volatility regime
USO's tracking error to Brent doesn't just widen when oil volatility spikes. It roughly triples. When Brent's 20-day realized volatility sat in the top quintile of its prior-year distribution, USO's trailing 20-day tracking error averaged 44.2% annualized. In the calm bottom quintile, it was 15.8% — a 2.79x ratio that holds up under HAC adjustment (p = 0.0004).
The study sorts roughly three years of aligned daily returns by Brent's volatility regime, then measures how far USO drifts from the crude benchmark over the following 20 sessions. The takeaway is straightforward: a direct oil ETF hedge becomes a noisier proxy precisely during supply shocks, when you might want it most. One caveat tempers the result — the 182 high-vol days cluster into just 13 distinct episodes, so the effect is clear but not broad.
The full evidence, including dispersion and episode-level detail, is in the analysis below.
Over the past ~3 years, when Brent crude's 20-day realized volatility is in the top quintile of its prior 252-session distribution, does USO's 20-day tracking error to Brent crude widen significantly versus the bottom quintile? I expect roll-cost and benchmark mismatch to decouple the ETF from spot crude most when oil volatility spikes, so direct oil ETF hedges become noisier precisely during supply shocks.
How this was measured
USO minute bars are collapsed to daily closes. Brent daily closes are aligned to USO trading days via forward-fill, then both return series are computed. Brent's 20-day realized volatility is the annualized standard deviation of daily log returns (sqrt(252) scale). For each day, that current Brent RV is compared with the prior 252-session distribution of Brent 20-day realized volatility, using a shifted rolling quantile so the current session is never in the lookback. Days at or above the 80th percentile are labeled Top quintile; days at or below the 20th percentile are labeled Bottom quintile. USO's tracking error to Brent is the annualized standard deviation of daily (USO return minus Brent return) over the trailing 20 sessions. A Welch t-test and a Newey-West HAC regression with 20 lags compare the tracking-error distributions across quintiles.
The key numbers
Reading the numbers
On days when Brent's recent volatility sat in the top fifth of its past year, USO's 20-day tracking error averaged about 44% versus roughly 16% on the calmest days - a gap of 28 percentage points, or nearly three times as wide. The odds that gap is luck are tiny: p=0.0000 day by day, and still 0.0004 after adjusting for the fact that volatility comes in clusters.
The charts
This box plot splits every day in the sample by where Brent's 20-day realized volatility ranked against its own prior 252 sessions, then shows the spread of USO's tracking error in each bucket. The detail to fix on is that the two boxes barely touch: the calmest-quintile days run from about 9.7% to 19.3% tracking error with a 15.8% mean, while the top-quintile box starts at 24.5% and runs to 48.7% with a 44.2% mean, stretching to a 115% worst case across 182 days. The widening is not a handful of freak days sitting inside an otherwise normal range - the two regimes occupy largely separate territory, which is what the research question asked.
The bar chart reduces the same comparison to two numbers: about 15.8% average tracking error in the bottom volatility quintile versus 44.2% in the top, a ratio of 2.79 times. Note the sample imbalance behind those bars - 37 bottom-quintile days against 182 top-quintile days - because it tells you high-volatility days were the common state over this three-year window, not a rare exception. For anyone weighing a direct oil ETF as a hedge, the practical read is that the instrument's link to Brent is at its loosest exactly when a supply shock makes hedging most tempting.
Here each of the 711 daily observations is a dot, plotted with Brent's 20-day realized volatility on the horizontal axis (running from about 15% to 111% annualized, averaging 37%) against USO's tracking error on the vertical axis (from about 6% to 115%, averaging 26%). The eye-catching feature is vertical: tracking error ranges far wider than volatility does, with an upside tail reaching more than four times its own average while nothing drops below about 6%. That asymmetry - errors can blow out but not shrink in a comparable way - is the decoupling risk the question was probing, and it means a hedge sized on a normal day can be materially mismatched on a stressed one.
Tracking-error summary by Brent realized-vol regime
| Regime | N days | Distinct episodes | Mean TE | Median TE | Std TE |
|---|---|---|---|---|---|
| Bottom quintile | 37 | 7 | 0.1581 | 0.1633 | 0.0212 |
| Top quintile | 182 | 13 | 0.4416 | 0.3381 | 0.2717 |
The takeaway
Yes — when oil volatility spikes, USO's tracking error to Brent blows out roughly threefold. On days in the top quintile of Brent's prior-year 20-day realized-vol distribution (Brent RV averaging about 60.5% annualized), USO's trailing 20-day tracking error averaged 44.2% annualized; in the bottom quintile (Brent RV near 20.7%) it averaged just 15.8%. That's a 28.4-point gap and a 2.79x ratio, and it's not driven by a handful of wild days alone — though the dispersion tells you the widening is lumpy rather than a smooth shift, with top-quintile TE standard deviation of 27.2% versus 2.1% in the calm regime. Statistically this is about as clear as it gets: the odds the gap is luck are roughly 4-in-10,000 even after adjusting for the fact that overlapping 20-day windows make consecutive observations highly dependent (HAC p = 0.0004). The practical takeaway is that a direct oil ETF hedge is a noisier instrument precisely during supply shocks — if you're sizing one, budget a much wider band around the crude exposure you think you have. The one thing tempering the strength is breadth: the 182 high-vol days collapse into only 13 distinct episodes, so you're really looking at a dozen-odd shocks, not 182 independent tests.
The fine print
- USO holds WTI-linked futures exposure, not Brent spot — the gap deliberately includes roll cost, benchmark mismatch, fees and premium/discount, not pure ETF slippage.
- 182 top-quintile days come from just 13 distinct high-vol episodes (37 days from 7 low-vol episodes), so the effective sample is far smaller than the day counts imply.
- This is contemporaneous conditioning: tracking error is wide while Brent vol is already elevated, not evidence that high vol predicts a future widening.
- 20-day tracking error and 20-day realized vol are overlapping windows, and the quintile cutoffs are rolling quantiles of overlapping data, so regime labels are stickier than an iid sample would produce.