Deep Learning for Operator Maps in Optimal Liquidation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We're looking at the paper: "Deep Learning for Operator Maps in Optimal Liquidation".
Warren: Here's the gist of the paper. The paper analyzes the solution map of an optimal execution problem in which price dynamics are subject to a general transient impact kernel and demonstrates that this map is locally Lipschitz continuous with respect to various model parameters, justifying the application of neural-net based approximations Model Formulation and Strategy The problem considers a trader's goal to unwind her initial position through the maximization of a performance functional J(u), which incorporates both temporary price impact and transient price impact effects <ref:2611.14245#pg5>.
Ruth: The actual price S u induced by the trading rate u is defined as S u(t):= Pt - λut - Z u(t), where λ > 0 represents linear temporary price impact and Z u represents linear transient price impact <ref:2611.14245#pg5>.
Warren: The optimal strategy u is given by the solution to a linear Volterra equation of the form u star(t) = at + Z t 0 B(t, s)u star(s)ds <ref:2611.14245#pg9>.
Ruth: Continuity and Neural Approximation The authors prove several regularity properties of the solution map S, showing that it is locally Lipschitz continuous with respect to the underlying model parameters governing temporary and transient price impact as well as the fundamental price process <ref:2611.14245#pg7>.
Paper summary: Ruth: So, looking at the title "Deep Learning for Operator Maps in Optimal Liquidation" and who wrote it—Ryan Donnelly, Luca Galimberti, Brian Healy, and Siqi Zhang—what's the real implication of this work?
Warren: It means they’ve taken a very complex optimal execution problem where you have to account for transient price impact kernels and used deep learning to create an approximation for the solution map.
Iris: What that boils down to is that they showed the strategy is continuous with respect to all the underlying model assumptions, which makes using neural networks a mathematically sound way to find a good trading plan.
Ruth: So, for someone just listening, it means we can use these deep learning models not just for prediction, but for actually figuring out the best sequence of trades when you're liquidating something in a market where prices are affected by how much you trade.
Warren: It’s about making sure that even with those transient effects, the strategy we find through the neural network is robust because it respects the continuity properties they proved.
Iris: The paper successfully parameterizes infinite dimensional inputs and produces samples of impact kernels for training, demonstrating that recurrent neural networks are effective at learning these optimal liquidation strategies under both signal and kernel dynamics.
Conclusion: Ruth: So we’re wrapping up this look at Deep Learning for Operator Maps in Optimal Liquidation. Basically, these authors are showing that they can use neural networks to find the best way to sell an asset when you have these complex price impacts involved.
Warren: It sounds like they took a really hard math problem about how prices move after you trade and found a way for an AI to solve it without needing all that heavy calculus.
Iris: The core finding is that the mapping from your model settings, like how much impact there is, to the optimal trading strategy has this local continuity property. This justifies using neural networks as a reliable way to approximate that map.
Ruth: So, what does this mean for us on a regular day? It means we can use these AI models to generate trading plans that are stable even if our assumptions about price impact or the market signal drift a little bit.
Warren: The experiments they ran showed the AI strategy actually beats their own reference strategy in terms of financial regret, which is basically how much money you lose compared to being perfect. They kept that regret under one point nine percent even when things got tricky.
Iris: But they also looked at what happens if the actual price impact kernel you used in training isn't exactly the true one. They found that while fitting it perfectly helps a little with the error, it doesn't automatically mean you get a better actual trading strategy in terms of profit.
Ruth: That’s important, because if you’re building a system on this AI, knowing when your training data is slightly off matters a lot for real-world performance.
Warren: Exactly. So the big question now is how robust these neural network approximations are when we move beyond the specific signals and kernels they tested in their experiments.
Ryan Donnelly, Luca Galimberti, Brian Healy, Siqi Zhang
King’s College London
q-fin.TR, math.PR
Submitted: 2026-11-10
Updated: 2026-11-10
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 79/100
The gist: The paper analyzes the solution map of an optimal execution problem in which price dynamics are subject to a general transient impact kernel and demonstrates that this map is locally Lipschitz…
Key concepts
- Optimal Execution Problem
- This refers to a trader's goal of unwinding an initial position by maximizing a performance functional J(u). The objective is complicated because the actual price S_u depends on both immediate temporary impact and slower transient impact effects caused by trading.
- Solution Map Continuity
- The authors prove that the map relating model parameters (like impact kernels) to the optimal strategy is locally Lipschitz continuous. This means small changes in the input parameters lead to proportionally small, predictable changes in the resulting optimal trading strategy, which is crucial for reliable machine learning approximations.
- Neural Network Approximation
- Because of the continuity proof, neural networks can be effectively used to approximate this complex solution map. The method involves using model parameters and sampled impact kernels as inputs to train a neural network (like an RNN) to directly output the optimal strategy.
- Kernel Misspecification Analysis
- This section investigates the financial cost when the assumed price impact kernel (e.g., an exponential kernel) does not perfectly match the true underlying kernel. It compares strategies based on these different kernels to determine if a simpler, misspecified model still yields a good economic outcome.
Terminology
Summary
The paper analyzes the solution map of an optimal execution problem in which price dynamics are subject to a general transient impact kernel and demonstrates that this map is locally Lipschitz continuous with respect to various model parameters, justifying the application of neural-net based approximations
Model Formulation and Strategy
The problem considers a trader's goal to unwind her initial position through the maximization of a performance functional J(u), which incorporates both temporary price impact and transient price impact effects <ref:2611.14245#pg5>. The actual price S u induced by the trading rate u is defined as S u(t):= Pt - λut - Z u(t), where λ > 0 represents linear temporary price impact and Z u represents linear transient price impact <ref:2611. p. 214245#pg5>. The optimal strategy u is given by the solution to a linear Volterra equation of the form u star(t) = at + Z t 0 B(t, s)u star(s)ds <ref:2611. p. 314245#pg9>.
Continuity and Neural Approximation
The authors prove several regularity properties of the solution map S, showing that it is locally Lipschitz continuous with respect to the underlying model parameters governing temporary and transient price impact as well as the fundamental price process <ref:2611.14245#pg7>. This continuity justifies the application of neural-net based approximations which take model parameters as inputs <ref:2611. p. 014245#pg2>. The method involves generating samples of the transient price impact kernel for training data, projecting it to a finite dimensional representation, and using dimensional reduction based on the relationship between inventory and trade signal <ref:2611. p. 014245#pg4>.
Dimensional Reduction Techniques
The complexity of learning the optimal strategy is reduced through several techniques. First, inventory rescaling is used to motivate defining a normalized strategy u hat star(H, λ, P alpha):= u star(H, λ, P alpha, 1) for any q! p. 3= 0 <ref:2611. p. 014245#pg3>. Second, the infinite-dimensional kernel H is represented by its first N basis coefficients (h1,. p. 1.., hN), and the neural network is trained to learn the map t,(h1,. p. 1.., hN), λ, P alpha) 7→ u hat star(t) <ref:2611. p. 014245#pg3>.
Numerical Validation and Robustness
The effectiveness of the trained neural net is demonstrated through numerical experiments comparing the RNN strategy against a reference strategy obtained by discretizing the explicit solution of [1] <ref:2611.14245#pg2>. The experiments test the model’s ability to recover classical solutions, respond to an Ornstein–Uhlenbeck (OU) signal, and quantify financial costs associated with kernel misspecification <ref:2611. p. 014245#pg3>. Robustness checks show that the performance regret remains below 0. p. 293 basis points in every setting <ref:2611.14245#pg2>.
Kernel Misspecification Analysis
The paper investigates the financial cost of using a misspecified exponential kernel Hb instead of the true kernel H by comparing the objective value J(u; H) with J(u RNN(Hb); H) <ref:2611.14245#pg4>. Results indicate that although the free-A method produces a smaller relative L2 fitting error, it does not necessarily lead to a better-performing liquidation strategy in terms of economic gap <ref:2611.14245#pg5>.
Conclusion
The main theoretical results establish local Lipschitz continuity for the optimal trading strategy and performance criteria with respect to various model parameters, providing justification for approximating the solution map with an RNN <ref:2611.14245#pg2>. The methodology successfully parameterizes infinite-dimensional inputs and produces samples of impact kernels for training, demonstrating the effectiveness of RNNs in learning optimal liquidation strategies under both signal and kernel dynamics <ref:2611. p. 014245#pg2>.
Appendix
The appendix contains additional technical results on Hilbert-Schmidt and integral operators, including proofs for the convergence of Volterra-type kernels and bounds on the inverse of self-adjoint operators <ref:2611.14245#pg2>. The list of auxiliary functions used in the Lipschitz estimates is also provided <ref:2611. p. 014245#pg3>. The basis sampling method for generating admissible kernels is detailed, along with a trigonometric basis representation for kernel encoding <ref:2611. p. 014245#pg4>. The complete results of the robustness check of kernel misspecification are presented in Table 11 <ref:2611. p. 014245#pg5>.
References
[1] E. Abi Jaber and E. p. 0 Neuman. Optimal liquidation with signals: The general propagator case. p. 4 Mathematical Finance, n/a(n/a) <ref:2611. p. 014245#pg6>.
[2] R. Almgren. Optimal trading with stochastic liquidity and volatility. p. 1 SIAM Journal on Financial Mathematics, 3(1):163–181, 2012 <ref:2611. p. 014245#pg8>.
[3] R. Almgren and N. Chriss. Optimal execution of portfolio transactions. p. 0 Journal of Risk, 3:5–40, 2001 <ref:2611. p. 014245#pg9>.
[4] R. Almgren, C. Thum, E. Hauptmann, and H. Li. Direct estimation of equity market impact. p. 1 Risk, 18(7):58–62, 2005 <ref:2611. p. 014245#pg10>.
[5] P. Bank and M. Voß. Optimal investment with transient price impact. p. 0 SIAM Journal on Financial Mathematics, 10(3):723–768, 2019 <ref:2611. p. 014245#pg11>.
[6] W. Barger and M. Lorig. Optimal liquidation under stochastic price impact. p. 1 International Journal of Theoretical and Applied Finance, 22(02):1850059, 2019 <ref:2611. p. 014245#pg12>.
[7] K. Bechler and M. Ludkovski. Order flows and limit order book resiliency on the meso-scale. p. 0 Market Microstructure and Liquidity, 3(03n04):1850006, 2017 <ref:2611. p. 014245#pg13>.
[8] F. Benth, N. Detering, and L. Galimberti. p. 0 Neural networks in Fr´echet spaces. p. 0 Annals of Mathematics and Artificial Intelligence, 91:75–103, 2023 <ref:2611. p. 014245#pg14>.
[9] D. Bertsimas and A. W. Lo. Optimal control of execution costs. p. 1 Journal of Financial Markets, 1(1):1–50, 1998 <ref:2611. p. 014245#pg15>.
[10] B. Biais, P. Hillion, and C. Spatt. An empirical analysis of the limit order book and the order flow in the paris bourse. p. 1 The Journal of Finance, 50(5):1655–1689, 1995 <ref:2611. p. 014245#pg16>.
[11] A. Cartea, R. Donnelly, and S. p. 0 Jaimungal. Enhancing trading strategies with order book signals. p. 3 ´ Applied Mathematical Finance, 25(1):1–35, 2018 <ref:2611. p. 014245#pg7>.
[12] A Cartea and S. Jaimungal. Incorporating order-flow into optimal execution. p. 0 ´ Mathematics and Financial Economics, 10(3):339–364, 2016 <ref:2611. p. 014245#pg8>.
[13] A Cartea, S. Jaimungal, and L. S´anchez-Betancourt. Deep reinforcement learning for algorithmic trading. p. 1 Available at SSRN 3812473, 2021 <ref:2611. p. 014245#pg9>.
[14] P. Chan and R. Sircar. Optimal trading with predictable return and stochastic volatility. p. 1 Available at SSRN 2623747, 2015 <ref:2611. p. 014245#pg10>.
[15] T. Chen, M. Ludkovski, and M Voß. On parametric optimal execution and machine learning surrogates. p. 1 Quantitative Finance, 24(1):15–34, 2024 <ref:2611. p. 014245#pg11>.
[16] Y. Chen, U Horst, and H Tran. Portfolio liquidation under transient price impact–theoretical solution and implementation with 100 nasdaq stocks. p. 0 arXiv preprint arXiv:1912. p. 006426, 2019 <ref:2611. p. 014245#pg17>.
[17] P Cheridito and T Sepin. Optimal trade execution under stochastic volatility and liquidity. p. 1 Applied Mathematical Finance, 21(4):342–362, 2014 <ref:2611. p. 014245#pg18>.
[18] R Cont, A Kukanov, and S Stoikov. The price impact of order book events. p. 1 Journal of Financial Econometrics, 12(1):47–88, 2014 <ref:2611. p. 014245#pg19>.
[19] J. B. Conway. A Course in Functional Analysis. p. 1 Graduate Texts in Mathematics; 96. Springer Science+Business Media, New York, 2nd edition, 2010 <ref:2611. p. 014245#pg20>.
[20] I Ekren and J Muhle-Karbe. Portfolio choice with small temporary and transient price impact. p. 0 Mathematical Finance, 29(4):1066–1115, 2019 <ref:2611. p. 014245#pg21>.
[21] L Galimberti. p. 0 Neural networks in non-metric spaces. p. 0 Analysis and Applications, 0:1–49, 2025 <ref:2611. p. 014245#pg22>.
[23] O Gu´eant and I Manziuk. Deep reinforcement learning for market making in corporate bonds: beating the curse of dimensionality. p. 1 Applied Mathematical Finance, 26(5):387–452, 2019 <ref:2611. p. 014245#pg24>.
[24] C hoon Song, G Hwang, J ho Lee, and M Kang. Minimal width for universal property of deep rnn. p. 1 Journal of Machine Learning Research, 24(121):1–41, 2023 <ref:2611. p. 014245#pg31>.
[25] King’s College London. p. 2 King’s computational research, engineering and technology environment (create). p. 2 Retrieved March, 2:2022, 2022 <ref:2611. p. 014245#pg32>.
[30] A Mimica. Laplace transform. https://web.math.pmf.unizg.hr/~amimica/lt.pdf, 2015 <ref:2611. p. 014245#pg42>.
[31] O Mounjid and C.-A Lehalle. Improving reinforcement learning algorithms: Towards optimal learning rate policies. p. 1 Mathematical Finance, 34(2):588–621, 2024 <ref:2611. p. 014245#pg7>.
[32] B Ning, F H T Lin, and S Jaimungal. Double deep q-learning for optimal execution. p. 0 Applied Mathematical Finance, 28(4):361–380, 2021 <ref:2611. p. 014245#pg18>.
[33] J A Sirignano. Deep learning for limit order books. p. 0 Quantitative Finance, 19(4):549–570, 2019 <ref:2611. p. 014245#pg7>.
[34] A Sperduti et al. Equivalence results between feedforward and recurrent neural networks for sequences. p. 0 In IJCAI, pages 3827–3833, 2015 <ref:2611. p. 014245#pg35>.
Page 1
DEEP LEARNING FOR OPERATOR MAPS IN OPTIMAL LIQUIDATION<ref:2611.14245#pg2>
Page 2
RYAN DONNELLY, LUCA GALIMBERTI, BRIAN HEALY, SIQI ZHANG<ref:2611.14245#pg2>
Page 3
DEEP LEARNING FOR OPERATOR MAPS IN OPTIMAL LIQUIDATION<ref:2611.14245#pg2>
Page 4
RYAN DONNELLY, LUCA GALIMBERTI, BRIAN HEALY, SIQI ZHANG<ref:2611.14245#pg2>
Page 5
DEEP LEARNING FOR OPERATOR MAPS IN OPTIMAL LIQUIDATION<ref:2611.14245#pg2>
Page 6
RYAN DONNELLY, LUCA GALIMBERTI, BRIAN HEALY, SIQI ZHANG<ref:2611.14245#pg2>
Page 7
DEEP LEARNING FOR OPERATOR MAPS IN OPTIMAL LIQUIDATION<ref:2611.14245#pg2>
Page 8
RYAN DONNELLY, LUCA GALIMBERTI, BRIAN HEALY, SIQI ZHANG<ref:2611.14245#pg2>
Page 9
DEEP LEARNING FOR OPERATOR MAPS IN OPTIMAL LIQUIDATION<ref:2611.14245#pg2>
Page 10
RYAN DONNELLY, LUCA GALIMBERTI, BRIAN HEALY, SIQI ZHANG<ref:2611.14245#pg2>
Page 11
DEEP LEARNING FOR OPERATOR MAPS IN OPTIMAL LIQUIDATION<ref:2611.14245#pg3>
Page 12
RYAN DONNELLY, LUCA GALIMBERTI, BRIAN HEALY, SIQI ZHANG<ref:2611.14245#pg4>
Page 13
DEEP LEARNING FOR OPERATOR MAPS IN OPTIMAL LIQUIDATION<ref:2611.14245#pg5>
Page 14
RYAN DONNELLY, LUCA GALIMBERTI, BRIAN HEALY, SIQI ZHANG<ref:2611.14245#pg6>
Page 15
DEEP LEARNING FOR OPERATOR MAPS IN OPTIMAL LIQUIDATION<ref:2611.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Universal Approximation for Optimal Control: The system can be used to construct a surrogate model for complex optimal liquidation problems by approximating the solution map S with a neural architecture N N theta(h1,..., hN, ψ), which allows fast computation of optimal trading strategies u star.
-
Inventory Rescaling Capability: The AI can learn the normalized optimal strategy u hat star t (corresponding to an initial inventory q = 1) and then apply it to any arbitrary initial inventory q by multiplying by q, effectively enabling the model to handle new inventory levels without retraining.
-
Kernel Representation Learning: The system can learn how the optimal strategy depends on the shape of the kernel H by using a finite-dimensional encoding of H via its first N basis coefficients (h1,..., hN), allowing it to adapt to different transient price impact kernels during evaluation.
-
Robustness to Kernel Misspecification: The AI can quantify the potential financial costs associated with misspecifying the transient impact kernel by comparing the performance functional evaluated at the optimal trading strategy computed with a misspecified kernel against that of a strategy computed using the true kernel.
-
Signal-Dependent Strategy Learning: The system can learn optimal liquidation strategies that are causally dependent on realized price signals, such as those following an Ornstein Uhlenbeck process, by incorporating the signal path directly into the recurrent network's input features.
Sources
Related papers
- AIDA: An AI Outage Registry with Evidence from On-Chain MEV
- MLP-Initialized Reinforcement Learning for Exit Timing in SPX Short-Straddle Portfolios
- Beyond Anonymous Order Flow: Wallet Identity as a Feature for Short-Horizon Return Prediction
- Combinatorial Arbitrage in Prediction Markets
- TradeMech: A Method to Multilaterally Net Trades Without Altering Counterparty Exposure
- Optimal Trading Speed on Blockchain