Part (3/3) - Refiner Trade: A Second Signal and the Case for Trading Less
The first part described the idea and put a gross Sharpe of about one and a half on it. The second part set it in front of a broker’s fees and an integer number of shares, and watched most of the edge go to the cost of trading. This part adds a second signal, drawn from the same spread. It does not raise the return by much. What it does is trade far less, and on a small account that is worth more than return.
The carry of the crack
The momentum signal reads the crack’s motion. But the crack has a second readable property: its shape in time. The front three-two-one crack and the second-month crack trade side by side, and their difference — the carry — says what the market pays for margin now against margin later. When the front trades above the deferred, backwardation, product is tight and the prompt margin is bid. It sits above zero most of the time; the spells of contango are shorter, and they mark slack. This is the raw daily series, and it is what everything below is built on.
Which feature of it predicts
A new object brings a choice: what do you read off it? The two candidates are the level itself — how deep in backwardation the crack stands — and the motion of that level, its own momentum, the same feature that worked on the crack price. This is not a choice to make by taste, so I test both the same way: sort the next quarter’s refiner return into ten buckets by each candidate and see whether the buckets line up.
For the level they do: the lowest-carry tenth loses about two percent over the following quarter, the highest earns better than twelve, and the order holds almost perfectly across all ten. The lowest bucket losing money is the part to look at — a centred long-short book cannot bank a drift, so the score is timing the refiner against its sector, not collecting an average. For the carry’s own momentum the sort is much weaker, and there is a structural reason. Summing the carry’s changes over a long window roughly rebuilds the level, so the carry’s momentum is a noisy copy of the level — the two run about 0.7 correlated — with a piece of the crack momentum mixed in, which we already hold. Its information content is about a third of the level’s at every horizon, and adding it to a book that already holds the level and the crack momentum makes that book slightly worse, not better. So the two signals split the crack cleanly between them: the price carries its information in its motion, the term structure carries it in its level.
Only the whole crack
It has to be the whole crack, and not for a reason of sign. Each leg has a term structure of its own. The crude leg reads backwards, because crude sits in the margin with a minus in front of it — but turn its sign around, so it points the right way, and it still barely sorts anything: a spread of about one and a half percent across the buckets against the whole crack’s twelve, and an order that is half noise. The crack is the better object because it weighs crude against gasoline and distillate the way the refiner’s own margin does. The legs are not stand-ins for the spread.
From feature to score
The level is the feature, so it becomes a score, built exactly like the momentum score so the two can meet on equal terms. Take the level, measure it as a z-score against its whole history to date — not the last year, because a level’s information is where it stands in the long run — clip at two standard deviations so no single stretch dominates, and rescale to sit between minus one and plus one. The momentum score was built the same way from the crack’s changes over a few short lookbacks. Two scores, one scale, both centred on zero.
Then they combine, and the combination preserves the scale: the prediction score is one minus w times momentum plus w times carry. It is a plain mix. With w at zero the book is all momentum, at one it is all carry, and in between the score never leaves its bounds — w is simply the share of the book that listens to the carry. That share is the only free number. The prediction score goes to the volatility target, which sizes the book at two percent a year and holds it hedged against the sector, as before. The score points the book; the volatility target decides how large.
A second bet
Two signals are only worth running together if they disagree often enough to matter, and these do. The carry score and the momentum score run about a quarter correlated across the whole history, and they stay that way through the crises where things that are supposed to be independent stop being independent. One reads the crack’s shape and one reads its motion; they are rarely early or late on the same day.
More carry, less trading
Everything to here is the idealized book: continuous weights, no fees, a plain cumulative-return curve. On that basis the blend is unremarkable. The combined Sharpe tops out near a carry share of one-third — about what the two strengths and their correlation call for — and slips past that as the weaker score crowds the stronger. But there is a second line on the same chart, and it does not top out. Turnover falls the whole way, from eleven trades a year to under three, because the carry score sits still while momentum keeps moving. Every step toward carry takes trades out of the book. The two lines do not agree on a share, and that disagreement is the whole of this part.
The share that is not on the paper
Here is the thing I most wanted to write down. Everything to this point — the Sharpe of each score, the blend curve — is the idealized strategy: continuous, frictionless, a cumulative-return line. That is not a backtest, and the two get confused all the time. A backtest is run forward one day at a time: whole shares, a fee on every order, the profit earned on the position actually held, not the one the model wished for. The idealized line tells you a signal is there. Only the backtest tells you what the account keeps, and the two can disagree about something as basic as how much carry to hold.
There are two numbers to set, and they pull on each other: the carry share, and the no-trade band from the second part, now read off the blended position. I run it on fifty thousand dollars — a small account, because the fixed fee bites hardest there and the hard case is the one worth showing — and sweep both.
Trade to target every day and the account keeps a Sharpe of about 0.6, the fee taking most of it. Add carry and open a band, and the surface climbs to a ridge a little above one, the carry share between a third and a half, the band around half to one percent. Past that the band turns against you: hold a position too long and you are trading a stale signal, and by three percent the edge is mostly gone. So there is an optimum, and it is not at the edge — it sits in the middle, where the fee you save and the signal you lose to staleness come into balance. That the top is a ridge and not a spike is what makes it usable: the numbers follow from the known cost of trading, not from fitting the returns.
Every account its own settings
Part 2 ended with a fan: what each account size keeps, from ten thousand dollars to a million. Here is the same fan for the blend — and this time each account gets its own pair of numbers, because the right band and the right carry share both move with size. The ten-thousand-dollar account wants a wide band and a heavy carry share, 1.6 percent and 0.55, because avoided trades are worth more to it than anything else; it climbs from minus 1.3 traded naively to plus 0.68. The mid-sized accounts settle around a one-percent band and a share near 0.43. At the top, the million-dollar account runs no band at all and holds the carry share near the gross optimum of a third, because for it the fee is noise and the paper answer is the right answer. Fifty thousand keeps 1.11, a hundred 1.20, a quarter million 1.28, a million 1.30 against a gross ceiling near one and a half. The edge is the same for everyone; the settings that keep it are not.
The whole trade, one account
To close, the whole thing in one picture, on one account: a hundred thousand dollars, carry share 0.55, a 0.4-percent band. The top panel is what it earns, net of every fee: a Sharpe of 1.20, which at this two-percent-vol setting means 2.3 percent a year at 1.9 percent volatility — scale the vol and everything scales with it. The middle panel is what it costs to sit through: a maximum drawdown of 3.2 percent in eighteen years. And the bottom panel is the machine itself, running: the refiner basket on one side of zero and the XLE hedge on the other, mirror images by construction, the pair flipping together as the crack changes its mind — long the margin, then short it, about seven trades a year. At the right edge the basket sits below zero: the margin signal reads short today. That is the point of the construction. It is not a story about refiners going up. It is a dial that reads the crack and points the book either way.
What the backtest leaves out
The backtest charges the fee on every order. It does not charge the cost of borrowing stock to sell short, and that gap is worth a word. The book is market-neutral and runs about nine percent gross on average — so the short side is a few percent of the account at any time, split between a large energy ETF and a handful of refiners. My guess is that borrow is cheap here. The ETF and the large refiners — Valero, Marathon, Phillips 66 — lend for a few basis points, and the smaller ones — Par Pacific, CVR, PBF — are still liquid enough to borrow, if now and then at a higher fee when short interest runs up. On a short side that small, even a fifty-to-a-hundred-basis-point fee is a few basis points on the account a year: real, small, and not in the numbers above. It would take a stock-loan feed to pin down, and my expectation is that it trims the net figure a little and moves none of the conclusions, with the smaller refiners, not the ETF, holding whatever cost there is.
Why it stays small
It is a low-capacity trade. It lives in a spread of a few tickers that are driven mostly by other things — oil, the broad market, the energy sector — with the margin signal a thin residual on top. That is the whole shape of it: a small, specific edge in a handful of names, not something that scales.
The crack tells you the margin, the stock is slow to follow, and the cheapest way to be early is, more often than the books admit, to trade less.