Paper: arXiv 2501.15106
Authors: Tingwei Meng, Moritz Voß, Nils Detering, Giulio Farolfi, Stanley Osher, Georg Menz
Abstract
We study operator learning in the context of linear propagator models for optimal order execution problems with transient price impact à la Bouchaud et al. (2004) and Gatheral (2010). Transient price impact persists and decays over time according to some propagator kernel. Specifically, we propose to use In-Context Operator Networks (ICON), a novel transformer-based neural network architecture introduced by Yang et al. (2023), which facilitates data-driven learning of operators by merging offline pre-training with an online few-shot prompting inference. First, we train ICON to learn the operator from various propagator models that maps the trading rate to the induced transient price impact. The inference step is then based on in-context prediction, where ICON is presented only with a few examples. We illustrate that ICON is capable of accurately inferring the underlying price impact model from the data prompts, even with propagator kernels not seen in the training data. In a second step, we employ the pre-trained ICON model provided with context as a surrogate operator in solving an optimal order execution problem via a neural network control policy, and demonstrate that the exact optimal execution strategies from Abi Jaber and Neuman (2022) for the models generating the context are correctly retrieved. Our introduced methodology is very general, offering a new approach to solving optimal stochastic control problems with unknown state dynamics, inferred data-efficiently from a limited number of examples by leveraging the few-shot and transfer learning capabilities of transformer networks.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 4.0/10
- Quadrant: Lab Rats — high math complexity, limited empirical testing
Why this score: The paper is dense with advanced stochastic control theory, PDE/ODE formulations, and transformer architecture derivations, making it highly mathematically complex. However, it relies on synthetic data and lacks real-market backtests, code, or statistical metrics, resulting in low empirical rigor.
Research Flowchart
flowchart TD
A["Research Goal:<br>Learn & Control Linear Propagator Models"] --> B{"Data & Inputs"}
B --> C["Offline Training Data:<br>Various Propagator Kernels"]
B --> D["Prompting Examples:<br>Few Context Trajectories"]
C --> E["Step 1: ICON Training"]
E --> F["Pre-trained ICON<br>(Operator Approximation)"]
D --> G["Step 2: In-Context Inference"]
F --> G
G --> H["Learned Dynamics:<br>Identified Propagator Model"]
G --> I["Step 3: Neural Control"]
H --> I
I --> J["Key Outcomes"]
subgraph J
J1["Operator Learning from Data"]
J2["Model-Agnostic<br>Few-Shot Identification"]
J3["Optimal Control Recovery"]
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 proposes a novel transformer‑based neural network architecture (ICON‑OCnet) for solving optimal order execution problems in the presence of unknown price impact.
Source sentence
We propose a novel transformer-based neural network architecture (ICON-OCnet) for solving optimal order execution problems in the presence of unknown price impact.
Assumes 3
- There exists a deterministic, non‑anticipative price impact operator Iθ mapping trading rates to price impact, such that Yt = (Iθ(u))t. model
Source sentence
We assume that there exists a determinstic, non-anticipative price impact operator Iθ : L2([0, T], R) →L2([0, T], R) satisfying (Iθ(u))t = (Iθ(u|[0,t]))t (0 ≤t ≤T) (4) for all u ∈U, such that the trader’s price impact process Y in (3) is given by Yt = (Iθ(u))t (0 ≤t ≤T).
- The price impact process starts at zero, i.e., Y0 = 0 so that the initial execution price equals the unaffected price. model
Source sentence
We assume Y0 = 0 so that P0 = S0 in (3).1 More precisely, we make the following general assumption. Assumption 2.1.
- Price impact persists and decays over time according to a propagator kernel. model
Source sentence
In this model class, price impact persists and decays over time according to some propagator kernel.
Data
- Synthetic propagator model examples synthetic unknown
Source sentence
To this end, we generate five synthetic examples as context from a specific propagator model θ using the same method as described in Section 4.2. However, for the in-context prediction task, we use as question condition the actual (out-of-distribution) ground truth optimal execution strategy u⋆of model θ which is computed via the implementation provided by [AJN25].
Costs
Modeled
An instantaneous cost term with constant ε > 0 is included to represent costs such as the bid‑ask spread.
Source sentence
The second integral with constant ε > 0 describes in reduced form all additional instantaneous costs, which are not directly linked to persistent price impact but incurred, for instance, by the bid-ask spread.
Finds 3
- ICON accurately infers the underlying price impact model from few‑shot prompts, even for propagator kernels not seen during training. predictive-accuracy
Source sentence
We illustrate that ICON is capable of accurately inferring the underlying price impact model from the data prompts, even for propagator kernels not seen in the training data.
- ICON‑OCnet correctly retrieves the exact optimal order execution strategy for the model that generated the in‑context examples. theoretical
Source sentence
Moreover, we demonstrate that ICON-OCnet correctly retrieves the exact optimal order execution strategy for the model generating the in-context examples.
- The errors of the ICON‑OCnet approach are very small, confirming its effectiveness for optimal execution via few‑shot in‑context learning. performance
Source sentence
We observe that the error values are impressively small. This confirms the effectiveness of our proposed ICON-OCnet approach in finding the optimal execution strategy via few-shot in-context learning with ICON.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 14 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
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.
Cited References (30)
- Fine-tune language models as multi-modal differential equation solvers (2025) — link
- The Illustrated Transformer (2018)
- Attention is All you Need (2017) — link
- Equilibrium in F unctional S tochastic G ames with M ean- F ield I nteraction — → 2306.05433
- Optimal Signal-Adaptive Trading with Temporary and Transient Price Impact (2022) — link
- Trading with the crowd (2023) — link
- Optimal Liquidation With Signals: The General Propagator Case (2025) — link
- Fredholm A pproach to N onlinear P ropagator M odels — → 2503.04323
and 22 more
- In-context operator learning with data prompts for differential equation problems (2023) — link
- Neural networks in Fr 'e chet spaces (2023) — link
- Pricing options on flow forwards by neural networks in a Hilbert space (2024) — link
- Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators (2021) — link
- Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators (2021) — link
- Fourier Neural Operator for Parametric Partial Differential Equations (2021) — link
- Structure-informed operator learning for parabolic P artial D ifferential E quations (2024) — link
- Neural Operator: Graph Kernel Network for Partial Differential Equations (2019) — link
- Neural Operator: Learning Maps Between Function Spaces With Applications to PDEs (2022)
- Fluctuations and response in financial markets: the subtle nature of ‘random’ price changes (2004) — link
- How Markets Slowly Digest Changes in Supply and Demand (2009) — link
- No-dynamic-arbitrage and market impact (2010) — link
- Trades, quotes and prices: financial markets under the microscope (2018)
- Unwinding S tochastic O rder F low: W hen to W arehouse T rades — → 2310.14144
- An Offline Learning Approach to Propagator Models — → 2309.02994
- An adaptive dual-level reinforcement learning approach for optimal trade execution — → 2307.10649
- Statistical Learning with Sublinear Regret of Propagator Models — → 2301.05157
- Handbook of Price Impact Modeling (2023)
- Deep Learning Approximation for Stochastic Control Problems (2016) — link
- Foundations of modern probability (2002) — link
- Monte Carlo on a single sample (2025) — link
- Trading with Concave Price Impact and Impact Decay—Theory and Evidence (2025) — link
6 of 30 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.