Treating a 0.96-correlated calendar spread as independent legs prices the WTI roll trade out of the book. The difference is large. Around the S&P GSCI roll, the standard multi-period framework's separable 3/2-power cost captures 0.6bp of gross alpha a month from WTI nearby-versus-deferred positioning. That monthly average covers the 96 months from 2004 to 2011 and assumes a $1B gross notional cap. On the same basis, the model Yinjun-Wang and Udell propose captures 12.9bp of monthly alpha. Whether that much lower charge is warranted drives the rest of the paper.
How trades incur cost
Take one contract during one period. Self-impact is gamma_{t,i} = l_i * sigma_{t,i} / v_{t,i}: forecast return volatility divided by forecast dollar volume, then scaled by a per-contract parameter. For the cross-section, the period impact matrix becomes Lambda_t = Gamma_t^{1/2} C_t Gamma_t^{1/2}. Here, Gamma_t is diagonal, while C_t is a correlation matrix with unit diagonal. The diagonal of Lambda_t therefore stays unchanged, adding cross-impact without altering self-impact. Return correlation serves as the proxy for C_t. It comes from an EWMA of daily settlement returns with decay 0.97 and remains fixed within each month.
A power-law kernel couples time: k(h) = (1 + h/tau)^{-beta}, with tau = 1 day and beta = 1/2. After h days, a trade's impact has fallen to k(h) of its initial size. Another trade in the same direction pays the residual displacement; an unwind recovers some entry cost. The kernel matrix begins lower triangular, preserving causality, before being symmetrized as G = (K + K')/2. Total cost is u'Au, where block A_ts = G_ts * Lambda_t^{1/2} Lambda_s^{1/2}. Since the diagonal blocks remain exactly Lambda_t, this construction contains the single-period model.
The index roll supplies the return pattern. S&P GSCI moves one fifth of its long position on each of the fifth through ninth business days, selling the nearby contract and buying the deferred according to a published calendar. Across the 96 months from 2004 to 2011, the nearby-minus-deferred spread declines 9bp during business days two to four. It loses another 32bp through the roll window, then rebounds 65bp by day 15. Irwin, Sanders and Yan report a consistent 30 to 40 basis point compression and roughly two-week recovery over the same 2004 to 2011 window. The trade shorts the spread ahead of the flow, turns long for the recovery, and exits before liquidity in the nearby dries up. Timing and position size are left to the optimizer.
Convexity comes with a mis-specification
The convexity proof is compact. Express the power-law kernel as a Gamma-weighted mixture of exponentials. Each exponential kernel e^{-c|h|} is the characteristic function of a Cauchy density and is therefore positive semidefinite. Integration against nonnegative weights yields G >= (1/2) I_T, which gives A >= (1/2) diag(Lambda_1, ..., Lambda_T). Under any liquidity forecast over the horizon, every path costs at least half the sum of its single-period costs. Round trips cannot carry negative cost, preventing the optimizer from recommending price manipulation.
This bound matters more than the cheaper cost estimate because it survives any liquidity path generated by the forecast. With a nonconvex impact cost, an optimizer can find and accept a round trip whose expected cost is negative. Exploiting that price manipulation is illegal.
The authors openly describe the guarantee as the product of a mis-specification. Under the causal cost, each trade's displacement is scaled by the liquidity in force when the trade occurred, giving blocks G_ts Lambda_s. Their two-period counterexample defines rho = k(a_2 - a_1). The determinant gamma_1 gamma_2 - rho^2 gamma_1^2 / 4 becomes negative whenever gamma_2 < rho^2 gamma_1 / 4, as can happen when an illiquid day is followed closely by a liquid one. Using the symmetric coupling Lambda_t^{1/2} Lambda_s^{1/2} removes that problem, while departing from the causal form the paper itself calls true. Fruth, Schöneborn and Urusov, whom they cite, reach the same issue for time-varying liquidity in propagator models.
Replication was unavailable
We produced no figures of our own, so this review does not test the paper's claim. Three material substitutions would be required. First, the 2004 to 2011 sample lies beyond our CL price coverage, which begins around 2010 and becomes shorter still by root. Second, we lack both the licensed exchange volume and the CFTC volume study used for calibration. Contract-level liquidity would therefore need to come from the daily CL dollar-volume fields available to us. Third, daily bars cannot verify realized impact or bid/ask cost. Any impact scale or spread charge would remain an assumption requiring stress tests.
Three prices for the roll trade
These are monthly averages in basis points of a $1B gross cap, with bid/ask charged at 1.0bp for the nearby leg and 1.5bp for the deferred. Transient cross-impact: 12.9 alpha, 5.9 predicted impact, 5.9 simulated impact, 3.9 spread, 3.1 net. Self-impact only: 7.7 alpha, 6.9 predicted, 2.7 simulated, 2.7 spread, 2.3 net. The 3/2-power baseline: 0.6 alpha, 1.0 predicted, 0.0 simulated, 0.2 spread, 0.4 net.
The baseline barely trades.
In the transient cross-impact row, predicted impact is 5.9 and simulated impact is also 5.9. Agreement follows because the simulator uses the transient cross-impact model. The paper states that it "serves two roles: it is the simulator's cost model and one of the tested models," and describes this limitation as standard for impact backtests. It is standard. The claimed 2.6x overpricing by the self-impact model, 6.9 predicted versus 2.7 simulated, consequently measures its distance from the assumed truth.
Invented volumes, assumed impact
Licensed exchange data would be needed for dollar volumes, so the paper simulates them. It starts with an annual average contract count from a CFTC study, multiplies by 1,000 barrels and the month's first settlement price, then allocates volume between the contracts through a logistic function with w_0 = 0.8, c = 7, s = 2. For the EIA report, Wednesdays receive a 1.08 uplift to volume and a 1.02 uplift to volatility. Lacking data to estimate the impact scale, the authors set l_i = 1. They also chose (tau, beta) = (1, 1/2) by hand, setting tau to match the one-decision-per-day interval. Every reported impact figure follows from those choices.
With n = 2 and estimated correlation 0.96, the cross-impact matrix contains one off-diagonal estimate. The paper reports no Sharpe, t-statistic, or dispersion across the 96 monthly observations, leaving no way to judge whether 3.1bp differs from 2.3bp by anything. Its sample ends in 2011 because that was "when the roll effect was economically material," in the paper's description. No post-2011 test appears.
The convexity result is worth retaining. A PSD per-period matrix combined with a power-law kernel gives G >= I_T/2 for any liquidity path, making it a useful component in a multi-period optimizer facing a scheduled liquidity calendar. The empirical case would become more persuasive with l_i and (tau, beta) estimated from execution records, alongside a cost comparison using a simulator separate from the contestants. The paper acknowledges both gaps. It chooses l_i = 1 for lack of data and says the transient cross-impact model "serves two roles" as simulator and tested model.