Energy Pairs: One Screen Adds 4.8 Points, Another Adds Nothing
The news wire gave us nothing to work with today, which is its own kind of gift. It means the only thing on the desk is the research, and this week's batch — six findings spanning September 27 to 29 — is unusually instructive, mostly because it refuses to flatter the setups it tests.
The edge is real, and it is tiny
Start with the most quotable result. When CPI runs hot and crude is sliding, XOM has historically edged out XLE over the following 20 sessions. Across 97 such days, XOM averaged +2.36% against XLE's +1.93% — a spread of 0.43 points, or 43 basis points, with XOM ahead 57% of the time. The t-stat of 2.10 and p of 0.038 clear the usual bar; this is not statistical noise dressed up as signal. It is, however, a very small thing. A 43-basis-point edge over a month in the energy sector is roughly the width of the bid-ask noise you'd expect from a single big inventory print. The finding is directionally true and practically borderline, and the research says so plainly: don't bank on it.
Same pair, two filters, opposite verdicts
The more interesting pair of results involves VLO against XLE, tested twice with different conditioning variables.
First: when retail-sales growth sits in its top trailing 252-session quintile and Brent's 20-day return is positive, VLO beat XLE handily — 49 qualifying days, VLO averaging +1.26% while XLE lost 1.81%, a 3.07-point edge, with VLO ahead on roughly 90% of those days. On its face, a screen worth framing. Then the baseline arrives: across all 421 eligible days, that same VLO-minus-XLE spread averaged 3.32%. The filter landed below the unconditional average. The setup deserves essentially none of the credit — the drift lives in the pair, not the screen. This is the quiet failure mode of relative-value research: a filter that looks selective while actually doing nothing except pointing at a spread that was already there.
Second: a low oil-beta, falling-Brent screen on the same pair. Twenty-seven qualifying days, forward 20-session VLO-minus-XLE spread averaging +7.5%, median +10.1%, VLO ahead in 81.5% of cases. Against the same +2.7% baseline on the other 421 days, that filter added roughly 4.8 points of lift. Identical tickers, identical baseline, opposite conclusion. The conditioning variable is doing all the work — which is the whole argument for testing screens rather than trusting them.
When a backtest wins and still loses
The TTE study is the cleanest cautionary tale. Buy TTE at the close when its trailing 20-day total return lags Brent crude: 26 closed trades, +35.85% on $100,000 of starting capital, a 54% win rate, best trade +9.61%, worst -3.75%. Over the same window, SPY buy-and-hold returned +76.34%. The strategy finished 40.49 points behind the benchmark. A positive absolute number that is a large relative failure is the most seductive kind of bad result, because the headline return reads fine until you put it next to the alternative.
Two more results round out the week, both of them negative in the useful sense. FANG trailed XOM after a 15bp-plus steepening in the 2s/10s curve with positive Brent momentum — average spread -0.42% across 36 event days, FANG ahead only about 42% of the time, median -4.29%, so the typical event was a loss. And the intraday range asymmetry in XOM on big Brent moves? The lean points the right way — 5.84% average range on days Brent fell more than 2% versus 5.32% when it rose more than 2% — but half a percentage point is indistinguishable from noise, and the medians say the same thing flatly.
The throughline: energy relative-value screens cluster in low-single-digit edges, and the ones that look strongest are frequently just inheriting the baseline they were supposed to beat. The screens worth trusting are the ones that publish the baseline next to the result — and this week's batch does, which is why it reads less like a sales pitch and more like a lab notebook.