Paper: arXiv 2310.14144

Authors: Marcel Nutz, Kevin Webster, Long Zhao

Abstract

We study how to unwind stochastic order flow with minimal transaction costs. Stochastic order flow arises, e.g., in the central risk book (CRB), a centralized trading desk that aggregates order flows within a financial institution. The desk can warehouse in-flow orders, ideally netting them against subsequent opposite orders (internalization), or route them to the market (externalization) and incur costs related to price impact and bid-ask spread. We model and solve this problem for a general class of in-flow processes, enabling us to study in detail how in-flow characteristics affect optimal strategy and core trading metrics. Our model allows for an analytic solution in semi-closed form and is readily implementable numerically. Compared with a standard execution problem where the order size is known upfront, the unwind strategy exhibits an additive adjustment for projected future in-flows. Its sign depends on the autocorrelation of orders; only truth-telling (martingale) flow is unwound myopically. In addition to analytic results, we present extensive simulations for different use cases and regimes, and introduce new metrics of practical interest.

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 sophisticated stochastic control model with analytic solutions and heavy use of advanced probability theory (martingales, dynamic programming, stochastic maximum principle), indicating high mathematical complexity. However, it focuses on theoretical modeling and simulation studies without presenting backtest-ready implementation, code, or real empirical data, placing it in the ‘Lab Rats’ quadrant.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Optimal Unwind of Stochastic Order Flow<br>with Minimal Transaction Costs"] --> B["Methodology:<br>Modeling Stochastic In-Flow Process"]
  B --> C["Key Inputs:<br>Order Flow Characteristics<br>Market Microstructure Parameters"]
  C --> D["Computation:<br>Stochastic Optimal Control<br>Semi-Closed Form Solution"]
  D --> E["Core Analytic Finding:<br>Unwind Strategy = Base Execution<br>+ Additive Adjustment for Future Flows"]
  E --> F["Key Outcomes & Regimes:<br>• Truth-Telling Flow: Myopic Unwind<br>• Autocorrelation Sign Determines Strategy"]
  F --> G["Validation:<br>Multi-Asset Simulations &<br>New Practical Metrics"]

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 a flexible model for unwinding stochastic order flow that yields a semi‑closed‑form analytic solution and can be implemented numerically as an order‑scheduling algorithm.

Source sentence
We model and solve this problem for a general class of in-flow processes, enabling us to study in detail how in-flow characteristics affect optimal strategy and core trading metrics. 1.1 Synopsis We propose and solve a flexible model for unwinding an exogenously given order flow, rich enough to serve as a practical order-scheduling algorithm.

Assumes 6

  • The analysis focuses on a single trading day and assumes the desk holds no overnight positions. model
    Source sentence
    For simplicity, our model focuses on a single trading day and a desk that does not hold overnight positions.
  • A liquidation constraint XT = 0 must be satisfied at the end of the horizon T. execution
    Source sentence
    Hence, we impose the liquidation constraint XT = 0 at the time horizon T corresponding to the end of the closing auction.
  • Price impact follows a generalized Obizhaeva–Wang model with time‑dependent but deterministic coefficients λt and βt. model
    Source sentence
    6 a generalized Obizhaeva–Wang (OW) model [48] with time-dependent coefficients, dYt = −βtYt dt + λt dQt. We use time-dependent but deterministic coefficients.
  • Bid‑ask spread costs are modeled as a quadratic instantaneous cost ½ εt q² on trading speed. model
    Source sentence
    For costs related to the bid-ask spread, we use a straightforward model as in [5]: intraday trading incurs a quadratic instantaneous cost 1 2εtq2 t on the trading speed qt = ˙Qt.
Show 2 more
  • The in‑flow driver can be any martingale; simulations use a discrete i.i.d. Gaussian shock process for tractability. data
    Source sentence
    We use a Brownian motion driver σt dWt to keep the model readily understandable to a broader audience but show in Section 2.4 that the results extend to a general martingale driver dMt with minor modifications, thus covering in-flows with jumps, self-excitement, etc. Instead, our experiments use a discrete driver where the inventory is shocked by a block trade approximately every 20 minutes.
  • Inventory dynamics are given by Xt = Qt – Zt, where Qt is the cumulative unwind trades (control) and Zt is the cumulative in‑flow. model
    Source sentence
    We denote by Qt the cumulative unwind trades up to time t; this is the control variable in our model. The desk’s outstanding position or inventory Xt = Qt −Zt is the difference between the unwind trades and the in-flow.

Data

  • LOBSTER limit order book events · NASDAQ order-book public
    S&P 500 stocks
    Source sentence
    Our data source is LOBSTER [41], which provides all Nasdaq limit order book events in a standard table format and is aimed at academic research. In line with [15] and the references therein, we find that the market order flow consistently exhibits momentum across the S&P 500 for time steps ranging from ten seconds to one hour.
  • Nasdaq 2022 Intern’s Guide to Trading · Nasdaq synthetic public
    Source sentence
    For example, Nasdaq’s The 2022 Intern’s Guide to Trading [42] gives numbers for large-cap stocks of 30 basis points for price impact and 5 basis points for spread costs. 18 These parameters were picked to match the overall statistics (spread and impact cost, volume on close) from Nasdaq’s 2022 Intern’s Guide to Trading [42] and align with the empirical estimates found in [27] and [44]. In-flow parameters.

Recognised sources: LOBSTER

Costs

Modeled Costs are part of the reported headline results.

Price impact is captured by a generalized Obizhaeva–Wang model (push factor λt and exponential decay βt) and spread costs are quadratic in trading speed with coefficient εt.

Source sentence
6 a generalized Obizhaeva–Wang (OW) model [48] with time-dependent coefficients, dYt = −βtYt dt + λt dQt. For costs related to the bid-ask spread, we use a straightforward model as in [5]: intraday trading incurs a quadratic instantaneous cost 1 2εtq2 t on the trading speed qt = ˙Qt.

Finds 4

  • The unwind strategy differs from a standard execution problem by an additive adjustment that depends on the autocorrelation of the order flow; only a martingale (truth‑telling) flow is unwound myopically. theoretical
    Source sentence
    Compared with a standard execution problem where the order size is known upfront, the unwind strategy exhibits an additive adjustment for projected future in-flows. Its sign depends on the autocorrelation of orders; only truth-telling (martingale) flow is unwound myopically.
  • A novel metric called internalization regret is introduced; it equals zero for the naive strategy that warehouses all in‑flow and externalizes only at the close. theoretical
    Source sentence
    We define internalization regret as10 internalization regret = 1 − |terminal in-flow| total variation of out-flow; (3.2) roughly speaking, it is one minus the ratio of minimally necessary over executed out-flow trades. It has value zero if and only if the out-flow is monotone (the desk only sells or only buys in the market), indicating that the desk captured all possible internalization.
  • Simulations show that higher initial impact states reduce the size of opening block trades and shift volume later in the day to exploit impact reversal. performance
    Source sentence
    For increasing values of y, the block trade is reduced; volume is deferred to exploit impact reversal. For high initial impact, reducing the impact state takes priority over reducing inventory and the opening trade adds to the initial inventory rather than hedging it.
  • Empirical analysis of LOBSTER data finds that market order flow exhibits momentum (positive autocorrelation) across S&P 500 stocks for time steps from ten seconds to one hour. statistical
    Source sentence
    In line with [15] and the references therein, we find that the market order flow consistently exhibits momentum across the S&P 500 for time steps ranging from ten seconds to one hour.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 16 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 12 papers in the archive (10 by independent authors): In-Context Operator Learning for Linear Propagator Models (uses its method or data); Nash Equilibrium between Brokers and Traders (uses its method or data)

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