Paper: arXiv 2503.02496

Authors: Philippe Bergault, Olivier Guéant, Hamza Bodor

Abstract

This paper addresses the trade-off between internalisation and externalisation in the management of stochastic trade flows. We consider agents who must absorb flows and manage risk by deciding whether to warehouse it or hedge in the market, thereby incurring transaction costs and market impact. Unlike market makers, these agents cannot skew their quotes to attract offsetting flows and deter risk-increasing ones, leading to a fundamentally different problem. Within the Almgren-Chriss framework, we derive almost-closed-form solutions in the case of quadratic execution costs, while more general cases require numerical methods. In particular, we discuss the challenges posed by artificial boundary conditions when using classical grid-based numerical PDE techniques and propose reinforcement learning methods as an alternative.

Complexity vs Empirical Score

  • Math Complexity: 9.0/10
  • Empirical Rigor: 3.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper involves advanced stochastic optimal control, HJB equations, and Riccati differential equations, placing it in the high math complexity range. However, it lacks empirical backtests or real-world data implementation, focusing instead on theoretical models and proposing reinforcement learning as an alternative to numerical methods, resulting in low empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Optimal Stochastic<br>Trade Flow Management"] --> B["Methodology:<br>Almgren-Chriss Framework<br>+ Reinforcement Learning"]
  B --> C["Inputs & Constraints:<br>Stochastic Flows, Quadratic<br>Costs, No Quote Skewing"]
  C --> D{"Computational Process"}
  D --> E["Analytical Solution:<br>Quadratic Costs"]
  D --> F["Numerical PDE:<br>Artificial Boundary Challenges"]
  D --> G["RL Approach:<br>Bypasses Grid Limitations"]
  E --> H["Key Findings & Outcomes"]
  F --> H
  G --> H
  H --> I["Pareto-optimal<br>Risk-Cost Trade-offs"]
  H --> J["RL Effectiveness for<br>Complex/Multidimensional Cases"]
  H --> K["Generalized Framework<br>beyond Market Making"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes using reinforcement learning as an alternative to classical PDE‑based methods for determining optimal trading policies in stochastic trade‑flow management, and derives almost‑closed‑form solutions for quadratic execution costs.

Source sentence
Within the Almgren-Chriss framework, we derive almost-closed-form solutions in the case of quadratic execution costs, while more general cases require numerical methods. In particular, we discuss the challenges posed by artificial boundary conditions when using classical grid-based numerical PDE techniques and propose reinforcement learning methods as an alternative.

Assumes 6

  • Trade flows follow Brownian dynamics. model
    Source sentence
    Assuming Brownian dynamics for the trade flows,2 and absolutely continuous trading strategies, the inven-tory process (qt)t∈[0,T ] evolves as: dqt = vtdt + νdBt, (1) where q0 ∈R and ν ∈R+ are given, (vt)t∈[0,T ] represents the trading velocity of the trader, and (Bt)t∈[0,T ] is a standard Brownian motion.
  • The asset price is a Brownian motion impacted linearly by the trading velocity. model
    Source sentence
    The asset price process (St)t∈[0,T ] is modelled as a Brownian motion impacted by the trader’s strategy: dSt = kvtdt + σdWt, (2) where S0 ∈R is given, σ ∈R∗ + is the volatility parameter, k ∈R+ represents the permanent market impact parameter, and (Wt)t∈[0,T ] is a standard Brownian motion.
  • Execution cost function L(v) is strictly convex and asymptotically super‑linear (e.g., L(v)=ψ²|v|+ηv^{1+ϕ}). execution
    Source sentence
    In this paper, we primarily focus on cases where L is a strictly convex and asymptotically super-linear function, such as L(v) = ψ 2 |v| + ηv1+ϕ, with ψ ∈R+, η ∈R∗ +, and ϕ ∈R∗ +.
  • The trader is risk‑averse and maximises expected CARA utility or a risk‑adjusted expectation. other
    Source sentence
    More precisely, we consider a risk-averse trader whose objective is either to maximise the expected CARA utility or a risk-adjusted expectation (à la Cartea-Jaimungal; see [15]) of their profit at a terminal time T. We show that, in both cases, the problem with quadratic execution costs can be tackled almost explic-itly, as it reduces to solving a matrix Riccati differential equation.
Show 2 more
  • The Brownian motions driving trade flow and price have constant correlation ρ. model
    Source sentence
    We assume that, together, (Bt)t∈[0,T ] and (Wt)t∈[0,T ] form a two-dimensional Brownian motion with a quadratic covariation coefficient ρ ∈[−1, 1].
  • Inventory and cash processes are absolutely continuous and start from given deterministic values. model
    Source sentence
    Assuming Brownian dynamics for the trade flows,2 and absolutely continuous trading strategies, the inven-tory process (qt)t∈[0,T ] evolves as: dqt = vtdt + νdBt, (1) where q0 ∈R and ν ∈R+ are given, (vt)t∈[0,T ] represents the trading velocity of the trader, and (Bt)t∈[0,T ] is a standard Brownian motion. The trader’s cash account (Xt)t∈[0,T ] evolves according to: dXt = −vtStdt −L(vt)dt, (3) where X0 ∈R is given,

Data

  • Synthetic market environment synthetic unknown
    Source sentence
    Unless specified otherwise, inventory is measured in lots of 5000 shares.4 Environment parameters: • Trading horizon: T = 1 day • Discrete time step: ∆t = 1 100 day • Initial lot price: S0 = 500, 000 $ · lot−1 • Arithmetic volatility: σ = 10,

Costs

Modeled Costs are part of the reported headline results.

Execution costs are modelled as L(v)=ψ²|v|+ηv² (linear spread term ψ and quadratic penalty η).

Source sentence
In this paper, we primarily focus on cases where L is a strictly convex and asymptotically super-linear function, such as L(v) = ψ 2 |v| + ηv1+ϕ, with ψ ∈R+, η ∈R∗ +, and ϕ ∈R∗ +. Unless specified otherwise, inventory is measured in lots of 5000 shares.4 Environment parameters: • Trading horizon: T = 1 day • Discrete time step: ∆t = 1 100 day • Initial lot price: S0 = 500, 000 $ · lot−1 • Arithmetic volatility: σ =

Finds 5

  • The RL method converges to near‑optimal solutions, achieving performance almost similar to the PDE‑based solution. performance
    Source sentence
    We proposed the use of an RL-based approach as an alternative to a more classical PDE-based approach in order to determine optimal trading policies. According to our results, our RL method converges to near-optimal solutions successfully, resulting in per-formance almost similar to the PDE-based solution, or exact solutions in contexts where they are known.
  • The PDE‑based solution slightly outperforms the RL policy on average. performance
    Source sentence
    On average, the PDE-based solution achieves a slight but consistent outperformance compared to the RL-based policy.
  • For quadratic execution costs the problem admits an almost‑closed‑form solution via a matrix Riccati differential equation. theoretical
    Source sentence
    More precisely, we consider a risk-averse trader whose objective is either to maximise the expected CARA utility or a risk-adjusted expectation (à la Cartea-Jaimungal; see [15]) of their profit at a terminal time T. We show that, in both cases, the problem with quadratic execution costs can be tackled almost explic-itly, as it reduces to solving a matrix Riccati differential equation. In Section 3, we show that
  • When execution costs are purely linear (η=0) the RL policy captures the optimal control structure with a no‑trading zone and increasing execution speed outside it. performance
    Source sentence
    The reinforcement learning agent successfully captures the policy structure, with no execution within a given interval and an execution speed that increases with inventory outside of this interval. However, the frontier between internalisation and externalisation varies more over time and the RL agent does not bring the inventory back to the frontier in one step but adopts a less aggressive approach.
  • The density of total rewards from the PDE solution is nearly indistinguishable from the closed‑form solution. statistical
    Source sentence
    The PDE-based solution produces a density curve that is very similar to the closed-form solution, making them nearly indistinguishable in the plot.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 31 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive (1 by independent authors)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.