Cross-Impact Kernels

• Cross-impact kernels are mathematical models that capture the time-dependent price responses in asset markets due to interdependent trading activity.
• They incorporate matrix-valued formulations and impose symmetries and positive definiteness to ensure market stability and preclude arbitrage.
• These kernels are estimated using both parametric and nonparametric techniques, providing actionable insights for optimal execution in equities, bonds, and derivatives.

Cross-impact kernels encode the dynamic relationship between trades in one financial instrument and price changes in others, generalizing the concept of self-impact in multi-asset markets. Functionally, a cross-impact kernel specifies the time-dependent price response in asset $i$ to the order flow in asset $j$, capturing both the strength and shape of this interdependence across various asset classes—equities, futures, bonds, and derivatives. The study of cross-impact kernels underpins the mathematical modeling, empirical measurement, and execution optimization of multi-asset trading strategies, with broad ramifications for optimal execution, market stability, and no-arbitrage pricing.

1. Formal Models and Mathematical Structure

In multi-asset markets, the impact of trading is described by a multivariate propagator or kernel model, typically expressed as

$$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$

where $S^i_t$ is the mid-price of asset $i$, $\dot{x}^j_s$ is the trading intensity in asset $j$, $f^{ij}(\cdot)$ is the instantaneous impact function (odd and typically linear in the absence of arbitrage), and $G^{ij}(\tau)$ is the (possibly matrix-valued) decay kernel encoding both self- ($i=j$) and cross- ($j$0) asset effects (Schneider et al., 2016, Alfonsi et al., 2013, Rosenbaum et al., 2021). In discrete time, the return formulation simplifies to

$j$1

where $j$2 is the cross-impact kernel, and $j$3 represents signed traded volume. For derivatives, the cross-impact kernel acquires a block structure reflecting the sensitivities (Greeks) of derivatives to underlying factors (Tomas et al., 2021).

A key property is the positive definiteness of the symmetrized kernel, which ensures convex, well-behaved cost functionals and rules out price manipulation (Alfonsi et al., 2013, Rosenbaum et al., 2021). The dynamics can be generalized to a Volterra propagator $j$4 for transient impact in continuous time (Jaber et al., 2024).

2. Theoretical Constraints and No-Arbitrage Admissibility

Precise restrictions on cross-impact kernels are dictated by arbitrage considerations:

• Oddness and Linearity: The instantaneous impact $j$5 must be odd and, for bounded kernels, linear in $j$6. Any nonlinearity with a bounded, non-increasing kernel allows price manipulation (Schneider et al., 2016).
• Symmetry: In the linear transient impact regime, cross-impact must satisfy $j$7. Violation leads to statistically arbitrageable strategies in the absence of transaction costs (Schneider et al., 2016, Rosenbaum et al., 2021).
• Positive Semidefiniteness: The cross-impact matrix $j$8 must be positive semidefinite to preclude arbitrage by round-trips (Alfonsi et al., 2013). Explicitly, $j$9 for any pair $$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$0 (Schneider et al., 2016).
• Kernel Decay: Empirically plausible models use exponential or power-law decay $$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$1, with further constraints on decay rates to ensure admissibility. For power-law kernels, exponent universality ($$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$2 for all $$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$3) is required to avoid manipulation. Gatheral's condition $$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$4 applies (Schneider et al., 2016, Hey et al., 8 Oct 2025, Wang et al., 2016).

In the context of martingale admissibility and statistical no-arbitrage, kernel admissibility is characterized both in the time and frequency domains. The admissible set is the intersection of martingale-admissible and no-statistical-arbitrage kernels, with explicit boundary matching at $$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$5 (immediate impact) and $$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$6 (permanent impact) (Rosenbaum et al., 2021).

3. Estimation Methodologies

Estimating cross-impact kernels involves both parametric and nonparametric approaches:

• Parametric Estimation: Exponential or power-law parameterizations are common, often using block Toeplitz or block design matrices and ordinary (ridge) least squares. Covariances of returns and order flows are used as sufficient statistics. Calibration is typically performed by maximizing in-sample $$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$7 or a Gaussian likelihood (Tomas et al., 2021).
• Nonparametric Estimation: Shape-constrained least squares, subject to nonnegativity, monotonicity, and convex decay, yield fully nonparametric estimates of the entire kernel tensor $$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$8. Projection onto the admissible cone can be efficiently implemented as a convex quadratic program (Hey et al., 8 Oct 2025).
• Block Structure in Derivatives: The estimation framework captures factor-derivative and derivative-derivative cross-impact via the instrument's sensitivity matrix, such as Greeks calculated via Black–Scholes implied volatilities (Tomas et al., 2021).
• Empirical Observables: Core observables are return covariance $$dS^i_t = \sum_{j=1}^N \int_0^t f^{ij}(\dot{x}^j_s) G^{ij}(t-s) ds + \text{noise},$$9, order-flow covariance $S^i_t$0, and order-sign correlation matrices. In high-frequency data, these are sampled over intervals and used directly in calibration (Tomas et al., 2021, Rosenbaum et al., 2021).

Using trade data (metaorders, order-flow imbalances), both proprietary and public, researchers have validated these methodologies on diverse asset universes including S&P futures, options, VIX derivatives, Italian sovereign bonds, and aggregated equities (Hey et al., 8 Oct 2025, Tomas et al., 2021, Schneider et al., 2016).

4. Empirical Findings and Kernel Shapes

Empirical measurement of cross-impact kernels leads to several robust conclusions:

• Concave Impact and Square-root Law: Both self- and cross-impact kernels are found to exhibit concave dependence on traded volume (i.e., $S^i_t$1), with $S^i_t$2 (the square-root law) holding for off-diagonal elements as well (Hey et al., 8 Oct 2025, Wang et al., 2016).
• Slow Decay and Power-law Shapes: Empirically, impact decays as a shifted power-law in time, $S^i_t$3, with $S^i_t$4. Self-impact generally decays more rapidly than cross-impact, which can persist over long lags (Hey et al., 8 Oct 2025, Wang et al., 2016).
• Asymmetry related to Liquidity: Cross-impact kernels exhibit asymmetry, where the more-liquid asset exerts stronger influence on the less liquid, but not vice versa. Statistical tests reveal significant but not exploitable violations of formal symmetry due to transaction costs (Hey et al., 8 Oct 2025, Schneider et al., 2016).
• Magnitude and Directionality: In derivative markets, the cross-impact matrix displays strong self-terms and significant cross-terms between spot and volatility factors, consistent with known leverage effects (Tomas et al., 2021).

Out-of-sample forecasting accuracy and $S^i_t$5 are increased by incorporating cross-impact, with nonparametric, shape-constrained models outperforming standard parametric forms (Hey et al., 8 Oct 2025).

5. Impact on Execution, Hedging, and Liquidity Metrics

Cross-impact kernels are central to the construction of execution cost models, hedging cost estimation, and liquidity risk metrics:

• Expected Trading Cost: The quadratic form $S^i_t$6 (with $S^i_t$7 a vector of trade schedules) directly delivers cost estimates for portfolio execution, with $S^i_t$8 expressing the full cross-impact structure (Tomas et al., 2021).
• Optimal Execution and Portfolio Choice: The Euler-Lagrange condition for optimal control with cross-impact reduces to a coupled system of stochastic Fredholm equations, solvable in closed form under admissible kernels (Jaber et al., 2024, Alfonsi et al., 2013). The optimal strategy is structurally altered by cross-impact, potentially inducing “transaction-triggered” round-trips in other assets.
• Liquidity Metrics: Portfolio “depth” is defined as $S^i_t$9 for $i$0, with the minimal eigenvalues of $i$1 quantifying the hardest-to-move or least liquid modes of the asset pool (Tomas et al., 2021).
• Risk Management: Kernel-based execution models support robust, no-arbitrage-consistent prescriptions for trading in multi-asset environments and facilitate analysis of alpha decay and transient alpha signals under dynamic portfolio adjustment (Jaber et al., 2024).

These methodologies are validated on cross-instrumental datasets, providing a unified approach for trading physicals, derivatives, and portfolios of multiple correlated assets (Tomas et al., 2021, Rosenbaum et al., 2021).

6. Construction and Admissibility of Matrix-valued Kernels

Rigorous construction of admissible cross-impact kernels leverages matrix-function theory and positive definiteness:

• Matrix-valued Kernels: Kernels $i$2 must satisfy nonnegativity, nonincreasingness, convexity, symmetry, and, ideally, commutativity ($i$3). Under these, diagonalization reduces admissibility to scalar-valued kernel conditions (Alfonsi et al., 2013).
• Matrix Functions: Flexible parameterizations via matrix exponential decay $i$4 with $i$5 positive semi-definite ensure strict positive definiteness and tractability. Decomposable, block, and scalar-kernel-multiplied kernels also provide admissible forms (Alfonsi et al., 2013).
• Spectral Criteria: Admissibility in the frequency domain requires the two-sided kernel extension to be the Fourier transform of a nonnegative-definite, Hermitian matrix-valued measure, providing a direct test for arbitrage (Rosenbaum et al., 2021).
• Projection to Admissible Set: Empirically estimated kernels may violate strict admissibility; projection onto the admissible cone (via eigenvalue clipping or convex optimization) rectifies these violations in practical implementations (Rosenbaum et al., 2021, Hey et al., 8 Oct 2025).

Explicit construction recipes and closed-form calibration steps are developed for both physical and derivative universes (Tomas et al., 2021, Rosenbaum et al., 2021).

7. Applications and Empirical Markets

Cross-impact kernels have been estimated and applied in a spectrum of markets:

Market	Data/Instrument Type	Key Kernel Features
S&P futures/options	High-frequency trades, Greeks	Instantaneous block kernels, significant spot-vol cross-impact (Tomas et al., 2021)
Italian sovereigns	Tick-level bonds	Slow-decay, asymmetric cross-impact (Schneider et al., 2016)
US equities	Order-flow imbalance, metaorders	Shifted power-law decay, asymmetric cross-impact, shape-constrained fit (Hey et al., 8 Oct 2025, Wang et al., 2016)

These analyses confirm the universality of transient, concave, and asymmetric cross-impact, as well as the critical role of exponent and decay-rate constraints for arbitrage-free modeling (Wang et al., 2016, Alfonsi et al., 2013, Rosenbaum et al., 2021).

---

In summary, cross-impact kernels are foundational objects in multi-asset market microstructure, encoding the entire interdependence of asset prices via order flow. Their admissibility, empirical estimation, and practical application are governed by an overview of stochastic control, convex optimization, and high-frequency econometrics, as detailed in the referenced arXiv literature (Alfonsi et al., 2013, Schneider et al., 2016, Tomas et al., 2021, Rosenbaum et al., 2021, Jaber et al., 2024, Hey et al., 8 Oct 2025, Wang et al., 2016).