Paper: arXiv 2407.04510
Authors: Alexander Barzykin, Robert Boyce, Eyal Neuman
Abstract
We consider a central trading desk which aggregates the inflow of clients’ orders with unobserved toxicity, i.e. persistent adverse directionality. The desk chooses either to internalise the inflow or externalise it to the market in a cost effective manner. In this model, externalising the order flow creates both price impact costs and an additional market feedback reaction for the inflow of trades. The desk’s objective is to maximise the daily trading P&L subject to end of the day inventory penalization. We formulate this setting as a partially observable stochastic control problem and solve it in two steps. First, we derive the filtered dynamics of the inventory and toxicity, projected to the observed filtration, which turns the stochastic control problem into a fully observed problem. Then we use a variational approach in order to derive the unique optimal trading strategy. We illustrate our results for various scenarios in which the desk is facing momentum and mean-reverting toxicity. Our implementation shows that the P&L performance gap between the partially observable problem and the full information case are of order $0.01%$ in all tested scenarios.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 3.0/10
- Quadrant: Lab Rats — high math complexity, limited empirical testing
Why this score: The paper presents a complex stochastic control problem with partial observations, filtering theory, and variational methods, requiring advanced mathematics. While it includes a numerical simulation and references to FX flow data, it lacks extensive backtesting, code, or empirical validation typical of high-rigor implementations, focusing more on theoretical derivations.
Research Flowchart
flowchart TD
A["Research Goal"] --> B["Formulate Stochastic Control Problem"]
B --> C["Apply Kalman Filtering for Partial Observability"]
C --> D["Derive Fully Observed Equivalent Problem"]
D --> E["Compute Optimal Strategy via Variational Approach"]
E --> F["Simulate Scenarios<br>Momentum vs Mean-Revert Toxicity"]
F --> G{"Outcomes"}
G --> H["P&L Performance Gap ~ 0.01%"]
G --> I["Robust Optimal Strategy"]
subgraph Inputs
direction LR
J["Client Order Inflow"] --> B
K["Unobserved Toxicity"] --> B
L["Inventory Constraints"] --> B
end
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 formulates the unwinding of toxic order flow as a partially observable stochastic control problem and derives a unique optimal trading strategy using filtering and a variational approach.
Source sentence
We formulate this setting as a partially observable stochastic control problem and solve it in two steps. Then we use a variational approach in order to derive the unique optimal trading strategy.
Assumes 5
- There exists a unique, continuous, invertible solution Φ to the matrix ODE (3.14) on [0,T] with bounded inverse. model
Source sentence
There exists a unique, continuous, invertible solution Φ to (3.14) on the interval [0, T], such that sup 0≤t≤T |Φ−1 t | < ∞. 11 Remark 3.5.
- Parameters at, bt, ε, β, λ and α are chosen such that the functions G, H, I, J satisfy positivity and non‑degeneracy conditions (Assumption 3.6). model
Source sentence
We assume that at, bt, ϵ, β, λ and α are chosen such that: (i) inf 0≤t≤TG4(t) > 0, inf 0≤t≤TH5(t) > 0, inf 0≤t≤TI6(t) > 0, inf 0≤t≤TJ7(t) > 0, (ii) inf t∈[0,T] |I5(t) −I6(t)| > 0, We will also need the following assumption on the functions (hq, iq, ˜hR,˜iR) which contain products, devisions and additions of the fundamental functions (G, H, I, J).
- Additional technical condition on functions (hq, iq, ˜hR, ˜iR) ensures a positive denominator (Assumption 3.7). model
Source sentence
We assume that inf t∈[0,T] 1 + G5(t) G4(t)hq(t) + G6(t) G4(t)iq(t) ∧ 1 + J5(t) J7(t) ˜hR(t) + J6(t) J7(t) ˜iR(t) > 0, Note that Assumptions 3.6 and 3.7 can be easily verified numerically, indeed they were satisfied in all of our numerical experiments in Section 4.
- The instantaneous price impact term ½ ε q_t represents trading costs associated with the bid‑ask spread. execution
Source sentence
The term 1 2ϵqt represents the instantaneous price impact as in [2] and it is also used in [31] as a proxy for the trading costs associated with the bid-ask spread.
Show 1 more
- A quadratic penalty α (X_T)^2 is imposed on end‑of‑day inventory discrepancies. market
Source sentence
The term α(Xq)2 T with a constant α > 0 imposes a costs for any discrepancies between the inflow and the outflow at the end of the day, hence it provides an incentive for the desk to unwind the order flow.
Data
- Institutional FX flow market-prices proprietary
Source sentence
Our study provides the following new insights: (i) Using institutional FX flow as an example we show that the probabilistic prop-erties of the order flow (see (2.2)) can randomly change from mean-reverting to momentum-driven every day.
Costs
Modeled
Price impact is modeled via a transient impact Y_q (Obizhaeva‑Wang) and an instantaneous impact ½ ε q_t as a proxy for spread costs.
Source sentence
(2.4) 6 We define the desk’s cumulative outflow trades as follows, Qq t = Z t 0 qsds, 0 ≤t ≤T. (2.5) We assume that the desk’s trading activity causes price impact on the risky asset’s execution price in the sense that their orders are filled at prices Sq t = Pt + Y q t + 1 2ϵqt, 0 ≤t ≤T, (2.6) where the transient price impact Y q is formulated according to the Obizhaeva and Wang model [32], Y q t = ye−βt + Z t 0
Finds 5
- The P&L performance gap between the partially observable problem and the full‑information case is of order 0.01% in all tested scenarios. performance
Source sentence
Our implementation shows that the P&L performance gap between the partially observable problem and the full information case are of order $0.01\%$ in all tested scenarios. We then studied the P&L performance gap between the partially observable problem and the full information case and found the regret to be of order 0.01% for all tested market scenarios.
- Feedback of externalisation has a prominent effect on P&L, increasing it by over 45% even for small feedback parameter values. performance
Source sentence
(ii) We show that the feedback of externalization has a prominent effect on P&L (of over 45%), even for small values of the feedback parameter (see table 4).
- A unique optimal unwind strategy exists and is given in feedback form (Theorem 3.8). theoretical figure from elsewhere in the paper
Source sentence
Then, there exists a unique minimiser q∗∈A to the cost functional (2.9). Moreover q∗satisfies, q∗ t = gX(t) ˆXq∗ t + gθ(t)ˆθq∗ t + gY (t)Y q∗ t + gP(t)Pt + E Z T t gA(s, t)dAs Yq t , (3.20) for 0 ≤t ≤T where gX, gθ, gY , gP and gA are deterministic functions defined in (C.6).
- The optimal agent consistently achieves a lower cost functional than a naive agent who mis‑specifies the feedback parameter. performance
Source sentence
We also observe, as expected, that the optimal agent consistently achieves a lower cost functional.
- Externalisation reduces inventory risk, leading to a total P&L distribution with significantly lower variance compared to no‑trading. performance
Source sentence
The total P&L can be seen to have a significantly lower variance than P&L without trading demonstrating the suitability of externalisation for managing the overall risk for the book.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 25 related papers). Coverage is partial, so this is neither replication nor refutation.
8 papers in the archive : Nash Equilibrium between Brokers and Traders ; Broker-Trader Partial Information Nash-Equilibria
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.