betaFeedback

WorksMathematics

Convex Modeling of Price Cross-Impact over Time

Abstract

Transaction costs can make or break a trading strategy, particularly in relative-value trading of commodity and macro markets, where edges are a few basis points. Price impact is a central component of transaction cost. Price impact models usually include self-impact (a trade in a contract moves that contract's price) but omit two well-documented effects: cross-impact (a trade in one contract also moves the prices of related contracts) and transient impact (price impact decays over time, so an unwind recovers part of the entry cost). A model without these effects overprices the impact of relative-value trades, whose correlated legs are built and unwound over days, and so forgoes potentially profitable trades. This paper models both effects with a convex quadratic cost. In each period, a positive semidefinite matrix built from volatility, volume, and correlation forecasts couples trades across contracts. A power-law decay kernel then couples trades across periods. The resulting cost admits no price manipulation even when liquidity varies over the planning horizon. The model is demonstrated empirically on calendar spread trading of crude oil futures around the commodity index roll.

Introduction

Multi-period portfolio construction chooses both what to trade and when to trade it. We take two empirical regularities as requirements for a useful transaction cost model. First, trades in related contracts3 interact: a relative-value trade is cheaper than the sum of its independently executed legs. Second, price impact persists: a later trade in the same direction pays the residual impact of earlier trades as an additional cost, while an unwind recovers part of the entry cost. These effects are usually called cross-impact and transient impact. Our model captures both simultaneously.

Modeling these two impacts is especially useful when liquidity varies predictably over the planning horizon. Published roll calendars, scheduled reports, settlement windows, and expiry cycles all generate recurring changes in volume and volatility. Relative-value trading in macro and commodity markets is the core application. Calendar spread trading of crude oil futures provides a running example, which we study empirically in Section 4.

The index roll. Commodity index exposure (e.g., to the S&P GSCI) is typically implemented through long positions in commodity futures. Because futures contracts expire, the index periodically rolls its positions by selling the expiring nearby contract and buying the next deferred contract before physical delivery. This scheduled, passive order flow creates temporary selling pressure in the nearby contract and buying pressure in the deferred contract, narrowing the nearby-minus-deferred calendar spread. Once the roll is complete, the spread tends to recover as the price pressure dissipates. From 2004 to 2011, the WTI nearby-minus-deferred spread fell by roughly 30–40 basis points around the roll window and recovered within about two weeks [19]. The index roll schedule is public and fixed in advance. For example, S&P GSCI rolls one fifth of its position on each of the fifth through ninth business days of the month [26]. The approximate size of the flow can also be inferred from public position reports and past roll activity.

This predictable spread compression creates a speculative trade: build a short spread position (short the nearby contract and long the deferred contract) before the roll, then reverse it (long the nearby, short the deferred) into the recovery [24]. Recognizing this price dislocation, however, does not determine how early to enter, how much to trade each day, or when to unwind. Intuitively, greater liquidity means lower transaction costs, so traders have an incentive to shift their trades toward liquid periods. Making these choices by discretion is further complicated by event-driven liquidity changes: for example, the EIA’s Weekly Petroleum Status Report, released Wednesdays at 10:30 a.m. ET, concentrates liquidity on the release day [30, 5]. Existing multi-period portfolio construction frameworks [10] model the price impact cost as separable, contract by contract and period by period, which overprices exactly relative-value trades: correlated legs built and unwound over days, where the legs’ impacts partially cancel and the entry’s impact is still alive at the exit.

Contributions. We propose a convex quadratic model of transient cross-impact cost. At time period tt, a positive semidefinite matrix Λt\Lambda_t describes self- and cross-impact under that period’s liquidity forecast. A power-law decay kernel then couples trades across time periods. We show that the resulting impact cost is convex. Convexity simplifies computation, and encodes a no-arbitrage condition. A nonconvex impact cost can assign negative expected cost to some round trip (a trade that buys and then sells back to zero inventory), and exploiting such price manipulation is illegal [18, 17]. Requiring convexity rules it out of the model. We empirically show the overpricing of existing models, and the usefulness of ours, on public data, trading calendar spreads of crude oil futures.

Portfolio construction and execution. Multi-period portfolio construction models trade expected return against risk and transaction cost over a planning horizon [10, 9]. Their price impact cost is modeled as separable, contract by contract and period by period. In Section 4, we empirically study how such separable modeling overprices the impact and yields a highly conservative portfolio. Limit-order-book and optimal execution models [25, 8, 20] address the complementary problem of executing each period-level trade. Our model operates at portfolio construction timescale.

Cross impact. Price impact across related contracts is often aligned with the factor structure of contracts’ returns [4, 11, 22]. Following this observation, we use return correlation as a proxy for impact correlation. Many models assume cross-impact proportional to return covariance (the risk matrix), Λt=λΣt\Lambda_t = \lambda\Sigma_t [16, 15]. Our construction (Section 3.1) keeps the shared correlation matrix but allows the cross-impact and risk matrices different scales.

Transient impact. Empirical studies find that price impact decays according to a power law, more slowly than an exponential [7, 6, 12, 8]. Exponential decay is nonetheless common despite its empirical defects, because the impact state is then Markovian and closed-form decision policy can be derived [25, 16]. Power-law decay is not Markovian, so we optimize the sequence of all trades jointly under the predicted liquidity over the planning horizon.

Linear and square-root price impact. Empirically, the single-contract price impact of a metaorder of size QQ (a succession of trades, all executed in the same direction and originating from the same market participant) scales as the square root of QQ [21, 3, 28, 23]. In this paper, we do not reproduce the square-root law, but model single-contract, single-period impact as linear with power-law decay. Nonlinear impact with decay generally admits illegal price manipulation [17, 27], while linear impact with a positive definite decay kernel rules it out [2, 1].

Convexity. For linear price impact and resulting quadratic cost, positive semidefiniteness of the full matrix over contracts and times is a sufficient condition for the absence of negative-cost round trips, i.e., price manipulation is never beneficial [18, 17]. A collection of positive semidefinite single-period matrices does not by itself ensure this property when liquidity changes over time [14].

A multi-period cross-impact model

In this section, we build the cost of transient cross-impact ϕimp\phi^{\mathrm{imp}} in two steps. First, inspired by risk matrix, we introduce a cross-impact matrix for each time period, and couple the periods through a temporal decay kernel.

We trade nn contracts at dates a1<⋯<aTa_1 < \cdots< a_T. The trade at date ata_t executes over period tt. The vector ut∈Rn\mathbf{u}_t \in\mathbb{R}^n is the signed dollar value traded at date tt, and dollar holdings satisfy

xt=xt−1+ut,t=1,…,T.\mathbf{x}_t = \mathbf{x}_{t-1} + \mathbf{u}_t,\qquad t = 1,\ldots,T.

The dates need not be equally spaced. They can be days, or intraday windows around scheduled events such as an index roll or a storage report. Write u=(u1,…,uT)∈RnT\mathbf{u} = (\mathbf{u}_1,\ldots,\mathbf{u}_T) \in\mathbb{R}^{nT}.

Single-period cross-impact

We first model price impact of trading a single contract ii at a single period tt, under that period’s liquidity forecast. Following [10], let vt,i>0v_{t,i} > 0 and σt,i>0\sigma_{t,i} > 0 denote forecasted dollar volume and return volatility for contract ii over period tt. A trade moves the price more when the contract is volatile, and less when it is heavily traded. We model the self-impact of contract ii at date tt as

γt,iut,i,where γt,i=ℓiσt,ivt,i,(1)\gamma_{t,i}u_{t,i},\qquad\text{where }\gamma_{t,i} = \ell_i\frac{\sigma_{t,i}}{v_{t,i}}, \tag*{(1)}

where ℓi∈R\ell_i \in\mathbb{R} is a learnable parameter.

Trading one contract also moves the prices of related contracts, so we combine the single-contract price impact γt,i\gamma_{t,i} with a correlation across contracts. Define Γt=diag⁡(γt,1,…,γt,n)\Gamma_t = \operatorname{diag}(\gamma_{t,1},\ldots,\gamma_{t,n}), and let Ct∈Rn×nC_t \in\mathbb{R}^{n \times n} be a positive semidefinite matrix with unit diagonal. We model the cross-impact over period tt as

Λt=Γt1/2CtΓt1/2,\Lambda_t = \Gamma_t^{1/2} C_t \Gamma_t^{1/2},

so that the cost of such cross-impact is utTΛtut\mathbf{u}_t^\mathsf{T}\Lambda_t\mathbf{u}_t. Because Γt\Gamma_t is diagonal and CtC_t has unit diagonal, the construction of Λt\Lambda_t adds cross-impact without disturbing the self-impact (1): Λt\Lambda_t has diagonal γt,1,…,γt,n\gamma_{t,1},\ldots,\gamma_{t,n}. We refer to CtC_t as impact correlation, and use return correlation as its proxy [16, 15].

The cross-impact matrix and the risk matrix share the same cross-sectional structure, as the risk matrix can be decomposed as

Σt1/2CtΣt1/2,\Sigma_t^{1/2} C_t \Sigma_t^{1/2},

where Σt1/2=diag⁡(σt,1,…,σt,n)\Sigma_t^{1/2} = \operatorname{diag}(\sigma_{t,1}, \ldots, \sigma_{t,n}). In practice, CtC_t can be further calibrated using limit order book data [4, 11, 22].

Transient impact

Cross-impact does not vanish when the period ends. We now couple cross-impact over time. We introduce a temporal decay kernel

k(h)=(1+h/τ)−β,τ,β>0,k(h) = (1 + h/\tau)^{-\beta}, \qquad\tau, \beta> 0,

consistent with the power-law decay found empirically [7, 12]. The kernel k(⋅):[0,∞)→[0,1]k(\cdot): [0, \infty) \to[0, 1] is nonincreasing, with k(0)=1k(0) = 1. At date ata_t, price impact of a trade at date asa_s has decayed to a fraction k(at−as)k(a_t-a_s) of its initial size. The width τ\tau sets the timescale of the decay: larger τ\tau, slower decay.

To charge each trade against the decayed impact of earlier trades, we model the impact cost of trades u\mathbf{u} as

ϕimp⁡(u)=∑t=1T∑s≤tk(at−as)utTΛt1/2Λs1/2us,(2)\phi^{\operatorname{imp}}(\mathbf{u}) = \sum_{t=1}^{T} \sum_{s \le t} k(a_t-a_s)\mathbf{u}_t^{\mathsf{T}}\Lambda_t^{1/2}\Lambda_s^{1/2}\mathbf{u}_s, \tag*{(2)}

with k(0)=1k(0)=1. This cost is a convex quadratic. To write the cost more compactly, we collect the kernel values in K∈RT×T\mathbf{K} \in\mathbb{R}^{T \times T}:

Kts={k(at−as),s≤t,0,s>t.K_{ts} = \begin{cases} k(a_t-a_s), & s \le t, \\ 0, & s > t. \end{cases}

K\mathbf{K} is lower triangular to enforce causality. The quadratic form is unchanged if we replace K\mathbf{K} by symmetric matrix

G=K+KT2.\mathbf{G} = \frac{\mathbf{K}+\mathbf{K}^{\mathsf{T}}}{2}.

The impact cost of trades u\mathbf{u} is then

ϕimp⁡(u)=uTAu,\phi^{\operatorname{imp}}(\mathbf{u}) = \mathbf{u}^{\mathsf{T}}\mathbf{A}\mathbf{u},

where

B=diag⁡(Λ11/2,…,ΛT1/2),A=B(G⊗In)B.(3)\mathbf{B} = \operatorname{diag}(\Lambda_1^{1/2}, \ldots, \Lambda_T^{1/2}), \qquad\mathbf{A} = \mathbf{B}(\mathbf{G} \otimes I_n)\mathbf{B}. \tag*{(3)}

In block form,

Ats=GtsΛt1/2Λs1/2.(4)A_{ts} = G_{ts}\Lambda_t^{1/2}\Lambda_s^{1/2}. \tag*{(4)}

And in particular, Att=ΛtA_{tt} = \Lambda_t, recovering the single-period model of Section 3.1.

Our model uses the symmetric coupling Λt1/2Λs1/2\Lambda_t^{1/2}\Lambda_s^{1/2}, which guarantees convexity. In contrast, the true, causal cost with time-varying liquidity, GtsΛsG_{ts}\Lambda_s, can make A\mathbf{A} indefinite and admit negative-cost round trips, i.e., illegal price manipulation. Appendix A gives explicit examples. The symmetric coupling avoids this failure.

Convexity and no price manipulation

With linear price impact, positive semidefiniteness of A\mathbf{A} is sufficient for the absence of negative-cost round trips: it gives

ϕimp(u)=uTAu≥0for every u,\phi^{\mathrm{imp}}(\mathbf{u}) = \mathbf{u}^{\mathsf{T}}\mathbf{A}\mathbf{u} \ge0 \quad\text{for every } \mathbf{u},

so the model admits no price manipulation.

We now show that A⪰0\mathbf{A} \succeq0 holds for any Λt⪰0\Lambda_t \succeq0 constructed in Section 3.1, so the cost (2) is convex no matter how liquidity varies over the horizon. The argument has two steps: we write the power-law kernel as a mixture of exponential kernels with nonnegative weights, and each exponential kernel is positive semidefinite because it is the characteristic function of a Cauchy distribution.

Define H∈RT×T\mathbf{H} \in\mathbb{R}^{T \times T} by

Hts=k(∣at−as∣),t,s=1,…,T.H_{ts} = k(\lvert a_t-a_s\rvert), \qquad t,s=1,\ldots,T.

Step 1: the kernel is a nonnegative mixture of exponentials. The substitution w=r(1+h/τ)w=r(1+h/\tau) in Γ(β)=∫0∞wβ−1e−w dw\Gamma(\beta)=\int_0^\infty w^{\beta-1}e^{-w}\,\mathrm{d}w gives ∫0∞rβ−1e−re−r(1+h/τ) dr=Γ(β)k(h)\int_0^\infty r^{\beta-1}e^{-r}e^{-r(1+h/\tau)}\,\mathrm{d}r=\Gamma(\beta)k(h), so that

Hts=1Γ(β)∫0∞rβ−1e−re−(r/τ)∣at−as∣ dr.H_{ts}=\frac{1}{\Gamma(\beta)}\int_0^\infty r^{\beta-1}e^{-r}e^{-(r/\tau)\lvert a_t-a_s\rvert}\,\mathrm{d}r.

Step 2: each exponential kernel is positive semidefinite. For c>0c>0, the Cauchy density with scale cc has characteristic function e−c∣h∣e^{-c\lvert h\rvert},

e−c∣h∣=1π∫−∞∞cc2+ω2eiωh dω,e^{-c\lvert h\rvert}=\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{c}{c^2+\omega^2}e^{i\omega h}\,\mathrm{d}\omega,

so for every z∈RT\mathbf{z}\in\mathbb{R}^T,

∑t,sztzse−c∣at−as∣=1π∫−∞∞cc2+ω2∣∑t=1Tzteiωat∣2 dω≥0.\sum_{t,s}z_tz_se^{-c\lvert a_t-a_s\rvert} =\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{c}{c^2+\omega^2}\left\lvert\sum_{t=1}^{T}z_te^{i\omega a_t}\right\rvert^2\,\mathrm{d}\omega\ge0.

Combining the two steps, by taking c=r/τc=r/\tau in Step 2 and integrating against the nonnegative weight rβ−1e−r/Γ(β)r^{\beta-1}e^{-r}/\Gamma(\beta) from Step 1, gives H⪰0\mathbf{H}\succeq0. Therefore G\mathbf{G} is positive semidefinite,

G=K+KT2=12(IT+H)⪰12IT.\mathbf{G}=\frac{\mathbf{K}+\mathbf{K}^{\mathsf{T}}}{2} =\frac{1}{2}(\mathbf{I}_T+\mathbf{H})\succeq\frac{1}{2}\mathbf{I}_T.

and by (3) so is A\mathbf{A},

A⪰12diag⁡(Λ1,…,ΛT)⪰0.\mathbf{A} \succeq\frac{1}{2}\operatorname{diag}(\Lambda_1,\ldots,\Lambda_T) \succeq0.

Every u\mathbf{u} thus costs at least half the sum of its single-period costs, and no round trip has negative cost.

Numerical study

We evaluate the model by relative-value trading of WTI calendar spreads around the S&P GSCI index roll. The study runs on public data. While a better practice of speculative traders would certainly require licensed exchange data, the study shows how our careful model of cross-impact and transient impact changes the constructed portfolio, compared with existing baselines.

WTI index-roll positioning

Setup. In each month we trade n=2n = 2 contracts: the nearby contract the index is selling and the deferred contract it is buying. The S&P GSCI roll window is business days five through nine. Trading takes place on the business days of the month, and the position is closed by the business day before the nearby contract’s last trading day.

Let rt∈R2\mathbf{r}_t \in\mathbb{R}^2 denote the forecast returns of the two contracts over period tt. The trader chooses, over the month, how large a spread position to hold and when to build and unwind it, by solving

maximize⁡∑t=1TrtTxt−ϕimp(u)−∑t=1Tψt(ut)subject toxt=xt−1+ut,x0=xT=0,1Txt=0,∥xt∥1≤xmax⁡,CVaR⁡α(−r~Txt)≤rmax⁡,t=1,…,T,(5)\begin{aligned} \operatorname{maximize}\quad& \sum_{t=1}^{T} \mathbf{r}_t^{\mathsf{T}}\mathbf{x}_t-\phi^{\mathrm{imp}}(\mathbf{u})-\sum_{t=1}^{T}\psi_t(\mathbf{u}_t)\\ \text{subject to}\quad& \mathbf{x}_t=\mathbf{x}_{t-1}+\mathbf{u}_t,\quad\mathbf{x}_0=\mathbf{x}_T=0,\\ & \mathbf{1}^{\mathsf{T}}\mathbf{x}_t=0,\quad\lVert\mathbf{x}_t\rVert_1\leq x^{\max},\\ & \operatorname{CVaR}_{\alpha}\left(-\tilde{\mathbf{r}}^{\mathsf{T}}\mathbf{x}_t\right)\leq r^{\max},\quad t=1,\ldots,T, \tag*{(5)} \end{aligned}

with variables u1,…,uT\mathbf{u}_1,\ldots,\mathbf{u}_T (dollar trades) and x1,…,xT\mathbf{x}_1,\ldots,\mathbf{x}_T (dollar holdings). The objective is forecast return, minus the impact cost ϕimp\phi^{\mathrm{imp}} of Section 3, minus a bid/ask transaction cost ψt(ut)=aT∣ut∣\psi_t(\mathbf{u}_t)=\mathbf{a}^{\mathsf{T}}|\mathbf{u}_t|, with a=(1.0,1.5)\mathbf{a}=(1.0,1.5) basis points for the nearby and deferred legs. The constraint 1Txt=0\mathbf{1}^{\mathsf{T}}\mathbf{x}_t=0 makes the position a calendar spread with equal dollars long and short, and xmax⁡x^{\max} caps its gross notional. The last constraint is a daily risk limit, imposed for each t=1,…,Tt=1,\ldots,T: r~\tilde{\mathbf{r}} is a one-day return vector drawn from the empirical distribution of the NN trading days before the month, and CVaR⁡α\operatorname{CVaR}_{\alpha} is the expected loss over the worst 1−α1-\alpha fraction of those scenarios, so each day’s tail loss is capped at rmax⁡r^{\max}.

We compare three specifications of impact cost:

  1. baseline: the separable 3/2-power cost ∑t,iℓiσt,i∣ut,i∣3/2/vt,i1/2\sum_{t,i}\ell_i\sigma_{t,i}|u_{t,i}|^{3/2}/v_{t,i}^{1/2} of [10], separable across both contracts and times, in place of ϕimp\phi^{\mathrm{imp}};

  1. self-impact: Ct=I\mathrm{C}_t = \mathrm{I} (no cross-impact) and G=I\mathrm{G} = \mathrm{I} (no transient impact);

  1. transient cross-impact: the estimated Ct\mathrm{C}_t and the kernel G\mathrm{G}.

Data. Prices are the daily settlements of the WTI futures contracts published by the U.S. EIA [29]. The data covers 2004–2011, when the roll effect was economically material.

The return forecast rtr_t for business day tt of a month is the mean realized return on business day tt over all prior months of the sample.

The dollar volumes vt,iv_{t,i} of (1) are licensed exchange data and not public, so we simulate them instead. Let vv denote the total daily dollar volume of both the nearby and deferred contracts in a given month, held constant within the month. We simulate vv with public data as

v=q⏟contracts/day×1,000⏟bbl/contract×p0⏟$/bbl.v = \underbrace{q}_{\text{contracts/day}} \times\underbrace{1{,}000}_{\text{bbl/contract}} \times\underbrace{p_0}_{\$/\text{bbl}} .

where qq is the average daily volume of all WTI futures over the prior calendar year and p0p_0 is the nearby contract’s settlement price on the month’s first trading day. We read qq from a CFTC study [13]. The nearby contract’s share of vv is a logistic function of the number of business days ete_t remaining until its last trading day,

wt=w01+exp⁡((c−et)/s).w_t = \frac{w_0}{1 + \exp\left((c-e_t)/s\right)} .

and the deferred contract takes the rest, i.e. vt,1=wtvv_{t,1} = w_t v and vt,2=(1−wt)vv_{t,2} = (1-w_t)v. The share starts near w0=0.8w_0 = 0.8 early in the month, and falls to half c=7c = 7 business days before the last trading day; with s=2s = 2, about half of the wtw_t-migration falls inside the five-day roll window, where 55–62% of index-roll activity is documented [19]. The Wednesday inventory report also concentrates liquidity on release days. We simulate this by scaling vv and σt,i\sigma_{t,i} by 1.08 and 1.02 on Wednesdays, estimated from past daily volume and return data.

The volatilities and the impact correlation are estimated by an exponentially weighted moving average (EWMA) of daily settlement returns, with decay rate λ=0.97\lambda= 0.97. Writing dmd_m for the last trading day of month m−1m-1, we estimate the contract covariance at dmd_m as Σdm\Sigma_{d_m}, and the volatility of every time period tt in month mm as σt,i=(Σdm)ii1/2\sigma_{t,i} = (\Sigma_{d_m})_{ii}^{1/2}. For simplicity, we freeze the estimate within each month: (5) is solved once, before the month begins, so the decision making is based on the information available at dmd_m.4 For the risk limit, we set the confidence level α=0.95\alpha= 0.95 and use each of the N=250N = 250 trading days before the month as a scenario to compute CVaR⁡α\operatorname{CVaR}_{\alpha}. We set the tail-loss cap rmax⁡r^{\max} to 50 basis points of xmax⁡x^{\max}: the trader accepts a one-day loss of at most 0.5% of gross capacity on the worst 5% of days.

We set the parameters of temporal decay kernel to (τ,β)=(1,1/2)(\tau,\beta) = (1,1/2). The normalizer τ=1\tau= 1 matches the decision interval, as we consider one decision per day. As we lack available data to learn the impact scaling ℓi\ell_i, we follow the empirics of [10] to set ℓi=1\ell_i = 1, so that trading one day’s volume moves the price by about one day’s volatility.

Metrics. We report the average of monthly profit and loss, in basis points of xmax⁡=$1Bx^{\max} = \$1\mathrm{B},

net⁡=∑t=1TrtrealizedTxt−ϕimp(u)−∑t=1TaT∣ut∣(6)\operatorname{net} = \sum_{t=1}^{T} \mathbf{r}_{t}^{\mathrm{realized}\mathsf{T}}\mathbf{x}_{t} - \phi^{\mathrm{imp}}(\mathbf{u}) - \sum_{t=1}^{T} \mathbf{a}^{\mathsf{T}}|\mathbf{u}_{t}| \tag*{(6)}

We report impact cost in two ways in Table 1, as it cannot be measured directly. Simulated impact cost is calculated by (2) with the estimated Ct\mathbf{C}_{t} and the kernel G\mathbf{G}. Predicted impact cost reports each model’s own forecast of ϕimp(u)\phi^{\mathrm{imp}}(\mathbf{u}) under its cost specification from the Setup paragraph. The transient cross-impact model therefore serves two roles: it is the simulator’s cost model and one of the tested models. This limitation is standard in impact backtests.

policyalphapredictedsimulated (2)spreadnet
baseline [10]0.61.00.00.20.4
self-impact7.76.92.72.72.3
transient cross-impact12.95.95.93.93.1

Table 1. Index-roll positioning, monthly averages over 2004–2011, in basis points of the gross limit xmax⁡=$1Bx^{\max} = \$1\mathrm{B}; the alpha, impact, spread, and net columns are defined in (6).

Results. Figure 1(a) shows the price dislocation: averaged over the 96 months, the nearby-minus-deferred calendar spread falls 9 basis points over business days two through four, falls another 32 over the roll window, and rebounds 65 by day 15, consistent with the documented 30–40 basis point compression and roughly two-week recovery [19]. Figure 1(b) shows the liquidity over time: volume migrates from nearby to deferred contract, with the weekly ripple of the Wednesday inventory report. Figure 1(c) shows the trading positions: all short the spread ahead of the roll window, flip long near its end, and are flat before the nearby’s liquidity dries up. The models differ in size and timing.

Figure 1. Index-roll positioning. (a) Averaged cumulative nearby-minus-deferred spread return by business day of month, 2004–2011; the shaded band is the S&P GSCI roll window. (b) Simulated daily dollar volumes of the two contracts. (c) Averaged nearby contract holding of all three models.

Figure 1(c) illustrates that the self-impact model and the baseline [10] overprice their impact. With estimated correlation 0.96, the price impact of the long and short legs partially cancel, but neither model captures this netting. The baseline is highly conservative and barely trades, as it charges each contract independently at the 3/23/2 power, so the spread trade looks expensive and gets no netting.

Conclusion

We proposed a convex quadratic model of price impact cost in which a positive semidefinite matrix couples trades across contracts and a power-law kernel couples trades across dates. We showed that the resulting cost admits no price manipulation. We empirically demonstrated our model’s usefulness, and the overpricing of existing models, in relative-value trading of crude oil futures.

Acknowledgements

We gratefully acknowledge support from the Office of Naval Research under award N000142412306, Air Force Office of Scientific Research under award FA9550-26-1-0012, the Alfred P. Sloan Foundation, the Stanford Institute for Human-Centered Artificial Intelligence (HAI), and from IBM Research as a founding member of Stanford Institute for Human-centered Artificial Intelligence.

Model preserves diagonals but not convexity

The construction (4) couples time periods through the symmetric matrix Ats=GtsΛt1/2Λs1/2\mathbf{A}_{ts} = G_{ts}\Lambda_t^{1/2}\Lambda_s^{1/2}. Here we show that the true, causal matrix GtsΛsG_{ts}\Lambda_s fails to be positive semidefinite. The counterexample uses n=1n = 1, T=2T = 2. Write Λ1=γ1\Lambda_1 = \gamma_1, Λ2=γ2\Lambda_2 = \gamma_2, and let ρ=k(a2−a1)∈(0,1]\rho= k(a_2 - a_1) \in(0, 1], so the off-diagonal weight is G12=ρ/2G_{12} = \rho/2. A model in which the trade at ss displaces prices in proportion to the liquidity prevailing at ss gives cross terms k(at−as)xtTΛsxsk(a_t - a_s)x_t^\mathsf{T}\Lambda_s x_s for t>st > s, i.e.,

M=[γ1ργ12ργ12γ2],det⁡M=γ1γ2−ρ2γ124.\mathbf{M} = \begin{bmatrix} \gamma_1 & \dfrac{\rho\gamma_1}{2} \\ \dfrac{\rho\gamma_1}{2} & \gamma_2 \end{bmatrix}, \qquad \det\mathbf{M} = \gamma_1\gamma_2 - \dfrac{\rho^2\gamma_1^2}{4}.

which is negative whenever γ2<ρ2γ1/4\gamma_2 < \rho^2\gamma_1/4: an illiquid period followed closely by a liquid one. Also see [14] for price manipulation in propagator models with time-varying liquidity.

References

  1. [1]Aurélien Alfonsi, Florian Klöck, and Alexander Schied. “Multivariate Transient Price Impact and Matrix-Valued Positive Definite Functions”. In: Mathematics of Operations Research 41.3 (2016), pp. 914–934.
  2. [2]Aurélien Alfonsi, Alexander Schied, and Alla Slynko. “Order book resilience, price manipulation, and the positive portfolio problem”. In: SIAM Journal on Financial Mathematics 3.1 (2012), pp. 511–533.
  3. [3]Robert Almgren et al. “Direct Estimation of Equity Market Impact”. In: Risk 18.7 (2005), pp. 58–62.
  4. [4]Michaël Benzaquen et al. “Dissecting cross-impact on stock markets: an empirical analysis”. In: Journal of Statistical Mechanics: Theory and Experiment 2017.2 (2017), p. 023406.DOI
  5. [5]Johan Bjursell, James E. Gentle, and George H. K. Wang. “Inventory Announcements, Jump Dynamics, Volatility and Trading Volume in U.S. Energy Futures Markets”. In: Energy Economics 48 (2015), pp. 336–349.DOI
  6. [6]Jean-Philippe Bouchaud, J Doyne Farmer, and Fabrizio Lillo. “How markets slowly digest changes in supply and demand”. In: Handbook of financial markets: dynamics and evolution. Elsevier, 2009, pp. 57–160.
  7. [7]Jean-Philippe Bouchaud et al. “Fluctuations and response in financial markets: the subtle nature of ‘random’ price changes”. In: Quantitative finance 4.2 (2003), p. 176.
  8. [8]Jean-Philippe Bouchaud et al. Trades, Quotes and Prices: Financial Markets under the Microscope. Cambridge University Press, 2018. ISBN: 9781107156050.DOI
  9. [9]Stephen Boyd et al. “Markowitz portfolio construction at seventy”. In: arXiv preprint arXiv:2401.05080 (2024).arxiv.org/abs/2401.05080
  10. [10]Stephen Boyd et al. “Multi-period trading via convex optimization”. In: Foundations and Trends in Optimization 3.1 (2017), pp. 1–76.arxiv.org/abs/1705.00109
  11. [11]Frederic Bucci et al. “Co-impact: crowding effects in institutional trading activity”. In: Quantitative Finance 20.2 (2020), pp. 193–205.
  12. [12]Frédéric Bucci et al. “Slow Decay of Impact in Equity Markets: Insights from the ANcerno Database”. In: Market Microstructure and Liquidity 4.03n04 (2018), p. 1950006. DOI: 10.1142/S2382626619500065. arXiv: 1901.05332 [q-fin.TR].arxiv.org/abs/1901.05332
  13. [13]Commodity Futures Trading Commission. Impact of U.S. Tight Oil on NYMEX WTI Futures. Tech. rep. https://www.cftc.gov/sites/default/files/2018-09/DMO_TightOilImpactNYMEX_WTI0818.pdf. CFTC Division of Market Oversight, 2018.
  14. [14]Antje Fruth, Torsten Schöneborn, and Mikhail Urusov. “Optimal Trade Execution and Price Manipulation in Order Books with Time-Varying Liquidity”. In: Mathematical Finance 24.4 (2014), pp. 651–695.
  15. [15]Nicolae Gârleanu and Lasse Heje Pedersen. “Dynamic portfolio choice with frictions”. In: Journal of Economic Theory 165 (2016), pp. 487–516.
  16. [16]Nicolae Gârleanu and Lasse Heje Pedersen. “Dynamic trading with predictable returns and transaction costs”. In: The Journal of Finance 68.6 (2013), pp. 2309–2340.
  17. [17]Jim Gatheral. “No-dynamic-arbitrage and market impact”. In: Quantitative Finance 10.7 (2010), pp. 749–759.DOI
  18. [18]Gur Huberman and Werner Stanzl. “Price manipulation and quasi-arbitrage”. In: Econometrica 72.4 (2004), pp. 1247–1275.
  19. [19]Scott H. Irwin, Dwight R. Sanders, and Lei Yan. “The Order Flow Cost of Index Rolling in Commodity Futures Markets”. In: Applied Economic Perspectives and Policy 45.2 (2023), pp. 1025–1050.DOI
  20. [20]Eduardo Abi Jaber, Eyal Neuman, and Sturm­ius Tuschmann. “Optimal Portfolio Choice with Cross-Impact Propagators”. In: arXiv preprint arXiv:2403.10273 (2024).arxiv.org/abs/2403.10273
  21. [21]Thomas F. Loeb. “Trading Cost: The Critical Link Between Investment Information and Results”. In: Financial Analysts Journal 39.3 (1983), pp. 39–44. doi: 10.2469/faj.v39.n3.39.DOI
  22. [22]Iacopo Mastromatteo et al. “Trading lightly: cross-impact and optimal portfolio execution”. In: Risk Magazine (July 2017). arXiv:1702.03838.arxiv.org/abs/1702.03838
  23. [23]Esteban Moro et al. “Market impact and trading profile of hidden orders in stock markets”. In: Physical Review E 80.6 (2009), p. 066102. DOI: 10.1103/PhysRevE.80.066102.DOI
  24. [24]Yiqun Mou. “Limits to Arbitrage and Commodity Index Investment: Front-Running the Goldman Roll”. Working paper, Columbia University. SSRN 1716841. 2011.DOI
  25. [25]Anna A. Obizhaeva and Jiang Wang. “Optimal trading strategy and supply/demand dynamics”. In: Journal of Financial Markets 16.1 (2013), pp. 1–32.
  26. [26]S&P Dow Jones Indices. S&P GSCI Methodology. https://www.spglobal.com/spdji/en/documents/methodologies/methodology-sp-gsci.pdf. 2026.
  27. [27]Michael Schneider and Fabrizio Lillo. “Cross-Impact and No-Dynamic-Arbitrage”. In: Quantitative Finance 19.1 (2019), pp. 137–154.
  28. [28]Bence Tóth et al. “Anomalous Price Impact and the Critical Nature of Liquidity in Financial Markets”. In: Physical Review X 1.2 (2011), p. 021006. DOI: 10.1103/PhysRevX.1.021006. arXiv: 1105.1694 [q-fin.TR].DOI
  29. [29]U.S. Energy Information Administration. Petroleum and natural gas futures prices: NYMEX contracts 1–4, daily settlements. https://www.eia.gov/dnav/pet/pet_pri_fut_sl_d.htm; https://www.eia.gov/dnav/ng/ng_pri_fut_sl_d.htm. Accessed August 2026. 2026.
  30. [30]U.S. Energy Information Administration. Weekly Petroleum Status Report: Release Schedule. https://www.eia.gov/petroleum/supply/weekly/. 2026.

Paper details

vlnc:W_ymcwsgsh
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents