TAPES: Terminal Adjoint-Propagated Error Scheduling for Financial Diffusion Samplers
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: "TAPES: Terminal Adjoint-Propagated Error Scheduling for Financial Diffusion Samplers".
Warren: Here's the gist of the paper. Financial diffusion samplers are often reused with a small number of function evaluations (NFE) inside scenario and risk pipelines, where numerical time-grid error can distort terminal quantities such as covariances, tails, and drawdown proxies TAPES estimates signed local defects by Richardson comparisons, transports them through the remaining discrete flow and a terminal-report adjoint, and optimizes the squared norm of their sum before a capped direct refinement TAPES gives the lowest design- and audit-report errors in all 24 dataset–budget–report rows within the reported five-pipeline panel Pruning redundant reasoning tokens The paper introduces TAPES (Terminal Adjoint-Propagated Error Scheduling) to estimate signed local defects by Richardson comparisons, transport them through the remaining discrete flow and a terminal-report adjoint, and optimize the squared norm of their sum before a capped direct refinement This method is designed for offline grid construction for a frozen probability-flow ODE where deployment NFE remains fixed The authors show that this transported signed sum is the first-order terminal-report error TAPES outperforms related methods like ART, AYS, EDM, and UNIF across six budgets on a regime-switching simulation and chronologically held-out Mag7 contexts How the training signal is built The study fixes the trained score network, optional conditioning, solver family, report normalizers, paired noises, and deployment budget while only changing the grid along the frozen probability-flow trajectory The schedule is parameterized as a time change where a uniform grid in u induces diffusion-time intervals Δk = ∫ (k+1)/K k/KthetaK(u) du The objective function minimizes LTAPES = E ∑ k=0 Λ + βdiag E sum k∥wk∥2Λ + ρeuc E sum k∥be k∥2 Where the gains come from The core contribution is report-aware grid construction, which optimizes a positive grid by transporting signed Richardson defects through a terminal-report adjoint, squaring their sum, and applying capped direct refinement The authors show that the transported signed sum is the first-order term in terminal report error Proposition 4 demonstrates that for aligned defects with common direction, the diagonal criterion is inefficient with a factor ρ ≥ 1 This inefficiency factor quantifies how much loss is incurred by ignoring cross-step reinforcement Downstream sensitivity The study validates TAPES across six budgets in two data regimes, showing that cross-budget mean relative audit-error reductions are about 10–11% on Regime-SV and about 1% on Mag7 The downstream sensitivity analysis traces a 21-day numerical bridge and report an exploratory, post-selected allocation comparison under one Mag7 protocol The results show that the Sharpe and Sortino ordering matches the report-error ordering (TAPES > ART > AYS > EDM > UNIF) in Table 4 Within those boundaries the evidence is consistent at every layer Theory identifies the transported signed sum as the first-order terminal-report error; the aligned model gives an inefficiency factor qualitatively consistent with the larger Regime-SV gains, plus a transport mismatch for the untransported defect criterion Within the reported five-pipeline panel, TAPES has the lowest mean paired error in all 24 rows at fixed deployment NFE TAPES changes no trained generator weights, solver, or deployment budget; it reallocates numerical effort where the terminal report is most sensitive The paper concludes that this method is relevant when low-NFE discretization error is appreciable and a high-NFE numerical target is affordable References
1: Roland Becker and Rolf Rannacher.
Warren: An Optimal Control Approach to A Posteriori Error Estimation in Finite Element Methods.
Ruth: Acta Numerica 10 (2001), 1–102.
Title and authors: Ruth: So we're looking at this paper called "TAPES: Terminal Adjoint-Propagated Error Scheduling for Financial Diffusion Samplers". It sounds like it's focused on improving how we construct the time grids for these diffusion samplers, which are used a lot in scenario and risk pipelines.
Warren: Right. Basically, they're tackling the problem where using a small number of function evaluations inside those pipelines means numerical time-grid errors can mess up important final numbers like covariances or drawdown proxies.
Iris: So, the core idea seems to be about finding a way to estimate these errors and then using that information to optimize the grid construction itself, even when you have a fixed number of function evaluations.
Ruth: Exactly. The paper introduces TAPES, which estimates signed local defects using Richardson comparisons, transports those defects through the remaining discrete flow and some kind of terminal-report adjoint.
Warren: And then it optimizes the squared norm of that sum before applying a capped direct refinement process to get a final grid design. It's all about minimizing terminal report errors across different budgets and reports.
Iris: What I find interesting is their claim that this transported signed sum actually ends up being the first-order term in terminal report error, which gives them a solid mathematical footing for their approach.
Ruth: That’s what makes it strong. Because of that theory, they can quantify the inefficiency of simpler methods, specifically showing how diagonal criteria are inefficient when defects are aligned with each other in a certain way.
Warren: And they show this inefficiency factor is greater than or equal to one when the defects have a common direction, meaning you're losing some accuracy by ignoring cross-step reinforcement. That’s a concrete cost to consider for anyone running these pipelines.
Iris: And they validate this across six different budgets on a regime-switching simulation and some chronologically held-out data contexts, showing TAPES gives the lowest design and audit report errors in all twenty-four rows of the five pipeline panel.
Ruth: That’s significant because it shows consistent performance regardless of whether you're looking at a regime-switching simulation or a Mag7 context. It suggests this method is robust across different financial modeling setups.
Warren: And the downstream sensitivity analysis they did, tracing a twenty-one day numerical bridge and comparing an exploratory allocation under one Mag7 protocol, showed that the Sharpe and Sortino ordering matches the report-error ordering, with TAPES leading ART, AYS, EDM, and UNIF across six budgets.
Iris: That implies that optimizing for terminal report error actually aligns with optimizing for downstream performance metrics like Sharpe and Sortino ratios in this context. What I don't see is a major limitation stated upfront about the scope of this work beyond what they mention about the fixed deployment number of function evaluations.
Ruth: Right, so the main thing to keep in mind is that this method is designed for offline grid construction where deployment number of function evaluations stays fixed, and it doesn't change any trained generator weights or solver family. It reallocates numerical effort based on where the terminal report is most sensitive.
Warren: So what does that mean for us practically? It means if you have low function evaluations, this method can make a bigger difference in terminal quantities like covariances and tails, which are often distorted by time-grid errors.
Iris: For someone building these pipelines, it suggests that if the discretization error is significant and you can afford a high function evaluation target numerically, TAPES is a good candidate for optimizing your grid construction.
Ruth: It really seems to be about smart allocation of computation rather than just using more of it blindly. We're looking at how this method might affect how we set up our entire risk pipeline for better results under those constraints.
Warren: Exactly, and the paper's title, "TAPES: Terminal Adjoint-Propagated Error Scheduling for Financial Diffusion Samplers", really tells you exactly what it does—it schedules the error based on terminal reports using an adjoint method.
Iris: Before we wrap up this discussion on TAPES, I want to say that while the results are strong under the aligned-defect model, they don't fully address how this performs when the defects aren't perfectly aligned. That’s a limitation they point out in their work regarding cross-step reinforcement.
Ruth: So, to wrap up, TAPES seems to be a method for constructing grids that minimizes terminal report errors by transporting signed Richardson defects through an adjoint structure and applying capped direct refinement. It shows this is effective across multiple budgets and contexts.
Warren: It means if you're constrained by function evaluations, you can use this approach to allocate your numerical effort where it matters most for the final report.
Iris: I think the main implication is that report-aware grid construction, which optimizes a positive grid by transporting those signed Richardson defects, is a powerful tool for offline error control in these complex simulation settings.
Ruth: So that's TAPES in a nutshell—a method that reallocates numerical effort based on terminal report sensitivity when function evaluations are fixed. We have covered the title, the core summary, and what makes it different from other scheduling approaches.
The paper's summary: Ruth: So we're looking at TAPES, which is this method for constructing time grids for financial diffusion samplers that focuses on minimizing terminal report errors when you have a fixed number of function evaluations inside your pipeline.
Warren: Right, so it's about optimizing where you spend your computational budget because those small discretization errors mess up the final risk numbers like covariances and tails.
Iris: The core mechanism is this TAPES method estimates signed local defects using Richardson comparisons and then transports those defects through a terminal report adjoint.
Ruth: And it optimizes the squared norm of that transported sum before applying a capped direct refinement process to get your final grid design.
Warren: What’s interesting here is the theory behind it, they show that this transported signed sum ends up being the first-order term in terminal report error.
Iris: That gives them a solid mathematical foundation, allowing them to show how diagonal criteria are inefficient when defects are aligned with each other in a certain way.
Ruth: They quantify this inefficiency by showing a factor greater than one when the defects have common direction, meaning you lose accuracy by ignoring cross-step reinforcement.
Warren: And they validated this across six different budgets on a regime-switching simulation and some held-out data contexts, showing it gives the lowest design and audit report errors in all twenty-four rows of their five pipeline panel.
Iris: That consistency is pretty strong, especially since it holds up whether you're looking at a regime-switching simulation or a Mag7 context.
Ruth: They also did this sensitivity analysis tracing a twenty-one day numerical bridge and comparing an exploratory allocation under one Mag7 protocol, which showed the Sharpe and Sortino ordering matched the report-error ordering with TAPES leading other methods.
Warren: So what this changes for us is that if you're constrained by function evaluations, TAPES lets you reallocate your numerical effort based on where the terminal report is most sensitive.
Iris: It means if you can afford a high function evaluation target numerically, this method provides a way to optimize your grid construction offline for better report fidelity.
Ruth: It’s about smart allocation of computation rather than just blindly using more of it, and we need to see how this plays out when you're talking about continuous time mean-variance reinforcement learning for policy gradient allocation.
The paper's improvements: Tom: So we're looking at how TAPES improves on older methods by focusing on report-aware grid construction instead of just local discretization errors.
Ruth: That means they are not just trying to fix small errors where you sampled; they are optimizing the whole grid based on what actually matters for the final report.
Warren: The mechanism involves transporting those signed Richardson defects through a terminal-report adjoint, squaring their sum, and then applying a capped direct refinement process to design the grid.
Iris: The key insight here is that this transported signed sum turns out to be the first-order term in terminal report error, which is a big deal for identification.
Ruth: Because of that theory, they can show exactly how much loss you incur by ignoring cross-step reinforcement when you only look at diagonal criteria.
Warren: They prove that if defects have aligned directions with each other, the diagonal criterion is inefficient by a factor greater than one, meaning you're missing some accuracy.
Iris: And this inefficiency factor quantifies how much loss happens when you ignore those cross-step reinforcements.
Ruth: The results show this method achieves the lowest design and audit report errors across all twenty-four rows in their five pipeline panel, regardless of whether it’s a regime-switching simulation or a Mag7 context.
Warren: That consistency across different scenarios is what really tells you this approach is robust for financial modeling.
Iris: Their downstream sensitivity analysis also confirmed this by showing that the Sharpe and Sortino ordering matches the report-error ordering with TAPES leading other methods across six budgets.
Ruth: So what this means for a trader, it’s that optimizing your time grid based on terminal report error actually aligns with optimizing your downstream performance metrics like Sharpe and Sortino ratios.
Warren: It suggests that spending computational effort where the final report is most sensitive is a better strategy than just chasing the lowest raw sampling error.
Iris: The method does optimize a positive grid by using this transport mechanism to minimize those terminal report errors, which is a significant step for offline grid construction when function evaluations are fixed.
Ruth: And they do flag that this method works best when the low function evaluation discretization error is appreciable and you can afford a high function evaluation target numerically.
Warren: So it’s not a magic fix for everything, but it’s a way to reallocate your resources intelligently within those constraints.
Conclusion: Ruth: So we’re wrapping up on TAPES, which is this paper about terminal report error scheduling for financial diffusion samplers that focuses on optimizing your time grid when you have a fixed number of function evaluations.
Warren: It seems like they've shown that by transporting those signed Richardson defects through an adjoint structure and applying a capped direct refinement, you can get the lowest report errors across all twenty-four rows in their panel.
Iris: The main implication is that this method provides a rigorous way to construct grids offline for frozen probability-flow ODEs where deployment function evaluations stay fixed.
Ruth: It’s about making sure the grid construction itself is optimized for what matters—the final report—not just the local simulation error.
Warren: They also showed that the transported signed sum is essentially the first-order term in terminal report error, which gives them a solid theory behind why this allocation strategy works better than others.
Iris: And they proved that for aligned defects, using a diagonal criterion is inefficient by a factor of one or greater because you’re ignoring cross-step reinforcement.
Ruth: So for someone building these pipelines, it means if the discretization error is significant and you can afford a high function evaluation target numerically, TAPES is a way to allocate your numerical effort where it matters most.
Warren: It’s about smart allocation of computation rather than just blindly using more of it, which is something every quantitative trader needs to consider when running these strategies.
Iris: The paper does flag that this method works best when the low function evaluation discretization error is appreciable, and you need to be able to afford that high function evaluation target.
Ruth: So for us, it’s a tool for offline error control in these complex simulation settings where the goal is report accuracy above all else.
Warren: I think it really shows how important it is to think about the entire pipeline and not just the individual simulation step when you're dealing with financial data.
Iris: It does provide a consistent framework, but like any method, they focus on offline construction for a frozen probability-flow ODE, so deploying it dynamically might need further exploration.
The Hong Kong Polytechnic University
q-fin.CP, cs.LG
Submitted: 2026-11-12
Updated: 2026-11-16
Comments: 9 pages. Accepted at the 7th ACM International Conference on AI in Finance (ICAIF 2026)
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 69/100
The gist: Financial diffusion samplers are often reused with a small number of function evaluations (NFE) inside scenario and risk pipelines…
Key concepts
- Signed Local Defects
- These are small numerical inaccuracies or 'defects' that occur when using a discrete time grid in simulations. TAPES identifies these defects by comparing the results of different numerical schemes, allowing it to pinpoint exactly where the simulation is inaccurate.
- Terminal-Report Adjoint
- This is a mathematical tool used to trace how errors made early in a process (like grid selection) propagate through the entire system to affect the final output or 'terminal report.' It acts like a backward pass, showing which initial choices cause the largest final errors.
- Report-Aware Grid Construction
- This is the core optimization technique. Instead of just minimizing local errors, TAPES uses the propagated defects to construct an optimal time grid. It transports these defects through the adjoint method and applies refinement to create a grid that specifically minimizes terminal report inaccuracies.
- Inefficiency Factor (rho >= 1)
- This factor quantifies how much error is lost when only considering errors in one direction (diagonal criterion). The paper shows that ignoring cross-step reinforcement leads to inefficiency, meaning the optimized method captures more important error information by considering interactions between different steps.
Summary
Financial diffusion samplers are often reused with a small number of function evaluations (NFE) inside scenario and risk pipelines, where numerical time-grid error can distort terminal quantities such as covariances, tails, and drawdown proxies TAPES estimates signed local defects by Richardson comparisons, transports them through the remaining discrete flow and a terminal-report adjoint, and optimizes the squared norm of their sum before a capped direct refinement TAPES gives the lowest design- and audit-report errors in all 24 dataset–budget–report rows within the reported five-pipeline panel
Pruning redundant reasoning tokens
The paper introduces TAPES (Terminal Adjoint-Propagated Error Scheduling) to estimate signed local defects by Richardson comparisons, transport them through the remaining discrete flow and a terminal-report adjoint, and optimize the squared norm of their sum before a capped direct refinement This method is designed for offline grid construction for a frozen probability-flow ODE where deployment NFE remains fixed The authors show that this transported signed sum is the first-order terminal-report error TAPES outperforms related methods like ART, AYS, EDM, and UNIF across six budgets on a regime-switching simulation and chronologically held-out Mag7 contexts
How the training signal is built
The study fixes the trained score network, optional conditioning, solver family, report normalizers, paired noises, and deployment budget while only changing the grid along the frozen probability-flow trajectory The schedule is parameterized as a time change where a uniform grid in u induces diffusion-time intervals Δk = ∫ (k+1)/K k/KthetaK(u) du The objective function minimizes LTAPES = E ∑ k=0 Λ + βdiag E sum k∥wk∥2Λ + ρeuc E sum k∥be k∥2
Where the gains come from
The core contribution is report-aware grid construction, which optimizes a positive grid by transporting signed Richardson defects through a terminal-report adjoint, squaring their sum, and applying capped direct refinement The authors show that the transported signed sum is the first-order term in terminal report error Proposition 4 demonstrates that for aligned defects with common direction, the diagonal criterion is inefficient with a factor ρ ≥ 1 This inefficiency factor quantifies how much loss is incurred by ignoring cross-step reinforcement
Downstream sensitivity
The study validates TAPES across six budgets in two data regimes, showing that cross-budget mean relative audit-error reductions are about 10–11% on Regime-SV and about 1% on Mag7 The downstream sensitivity analysis traces a 21-day numerical bridge and report an exploratory, post-selected allocation comparison under one Mag7 protocol The results show that the Sharpe and Sortino ordering matches the report-error ordering (TAPES > ART > AYS > EDM > UNIF) in Table 4
Conclusion
Within those boundaries the evidence is consistent at every layer Theory identifies the transported signed sum as the first-order terminal-report error; the aligned model gives an inefficiency factor qualitatively consistent with the larger Regime-SV gains, plus a transport mismatch for the untransported defect criterion Within the reported five-pipeline panel, TAPES has the lowest mean paired error in all 24 rows at fixed deployment NFE TAPES changes no trained generator weights, solver, or deployment budget; it reallocates numerical effort where the terminal report is most sensitive The paper concludes that this method is relevant when low-NFE discretization error is appreciable and a high-NFE numerical target is affordable
References
[1] Roland Becker and Rolf Rannacher. 2001. An Optimal Control Approach to A Posteriori Error Estimation in Finite Element Methods. p. 7 References
[1] Roland Becker and Rolf Rannacher. An Optimal Control Approach to A
Posteriori Error Estimation in Finite Element Methods. An Optimal Control Approach to A
Posteriori Error Estimation in Finite Element Methods. Acta Numerica 10 (2001),
1–102.Where the paper says it
Where the paper says it
[2] Tim Bollerslev. 1986. Generalized Autoregressive Conditional Heteroskedasticity. p. 7 1017/S0962492901000010
[2] Tim Bollerslev. Generalized Autoregressive Conditional Heteroskedasticity. Generalized Autoregressive Conditional Heteroskedasticity. Journal of Econometrics 31, 3 (1986), 307–327.Where the paper says it
Where the paper says it
[3] Yang Cao and Linda Petzold. p. 7 Journal of Econometrics 31, 3 (1986), 307–327. [3] Yang Cao and Linda Petzold. [3] Yang Cao and Linda Petzold. A Posteriori Error Estimation and Global Error
Control for Ordinary Differential Equations by the Adjoint Method. A Posteriori Error Estimation and Global Error
Control for Ordinary Differential Equations by the Adjoint Method. SIAM Journal
on Scientific Computing 26, 2 (2004), 359–374.Where the paper says it
Where the paper says it
Where the paper says it
[4] Jehanzeb Hameed Chaudhry, Don Estep, Simon Tavener, Varis Carey, and Jeff Sandelin. p. 7 SIAM Journal
on Scientific Computing 26, 2 (2004), 359–374. 1137/S1064827503420969
[4] Jehanzeb Hameed Chaudhry, Don Estep, Simon Tavener, Varis Carey, and Jeff
Sandelin. 1137/S1064827503420969
[4] Jehanzeb Hameed Chaudhry, Don Estep, Simon Tavener, Varis Carey, and Jeff
Sandelin. A Posteriori Error Analysis of Two-Stage Computation Methods
with Application to Efficient Discretization and the Parareal Algorithm. A Posteriori Error Analysis of Two-Stage Computation Methods
with Application to Efficient Discretization and the Parareal Algorithm. SIAM
Journal on Numerical Analysis 54, 5 (2016), 2974–3002.Where the paper says it
Where the paper says it
Where the paper says it
[5] Defang Chen, Zhenyu Zhou, Can Wang, and Siwei Lyu. p. 8 TAPES: Terminal Adjoint-Propagated Error Scheduling for Financial Diffusion Samplers ICAIF ’26, November 14–17, 2026, Milan, AA, Italy
[5] Defang Chen, Zhenyu Zhou, Can Wang, and Siwei Lyu. Geometric Regularity in Deterministic Sampling Dynamics of Diffusion-Based Generative Models. TAPES: Terminal Adjoint-Propagated Error Scheduling for Financial Diffusion Samplers ICAIF ’26, November 14–17, 2026, Milan, AA, Italy
[5] Defang Chen, Zhenyu Zhou, Can Wang, and Siwei Lyu. Geometric Regularity in Deterministic Sampling Dynamics of Diffusion-Based Generative Models.Where the paper says it
Where the paper says it
[6] Yuzhu Chen, Fengxiang He, Shi Fu, Xinmei Tian, and Dacheng Tao. 2024. Adaptive Time-Stepping Schedules for Diffusion Models. p. 8 1088/1742-5468/ae17ac
[6] Yuzhu Chen, Fengxiang He, Shi Fu, Xinmei Tian, and Dacheng Tao. Adaptive Time-Stepping Schedules for Diffusion Models. Adaptive Time-Stepping Schedules for Diffusion Models. In Proceedings of the Fortieth
Conference on Uncertainty in Artificial Intelligence (Proceedings of Machine Learning Research, Vol.Where the paper says it
Where the paper says it
[7] Rama Cont. 2001. Empirical Properties of Asset Returns: Stylized Facts and Statistical Issues. p. 8 html
[7] Rama Cont. Empirical Properties of Asset Returns: Stylized Facts and
Statistical Issues.Where the paper says it
[8] Rama Cont, Mihai Cucuringu, Renyuan Xu, and Chao Zhang. 2025. Tail-GAN: Learning to Simulate Tail Risk Scenarios. p. 8 [8] Rama Cont, Mihai Cucuringu, Renyuan Xu, and Chao Zhang. Tail-GAN:
Learning to Simulate Tail Risk Scenarios.Where the paper says it
[9] Jamie Fairbrother, Amanda Turner, and Stein W. Wallace. 2018. Scenario Generation for Single-Period Portfolio Selection Problems with Tail Risk Measures: Coping with High Dimensions and Integer Variables. p. 8 00936
[9] Jamie Fairbrother, Amanda Turner, and Stein W. Scenario Generation for Single-Period Portfolio Selection Problems with Tail Risk Measures:
Coping with High Dimensions and Integer Variables.Where the paper says it
[10] Michael B. Giles and Endre Süli. p. 8 0790
[10] Michael B. Giles and Endre Süli. Giles and Endre Süli. Adjoint Methods for PDEs: A Posteriori
Error Analysis and Postprocessing by Duality.Where the paper says it
Where the paper says it
[11] Paul Glasserman. 2003. Monte Carlo Methods in Financial Engineering. p. 8 1017/S096249290200003X
[11] Paul Glasserman. Monte Carlo Methods in Financial Engineering.Where the paper says it
[12] Andreas Griewank and Andrea Walther. p. 8 Monte Carlo Methods in Financial Engineering. [12] Andreas Griewank and Andrea Walther. [12] Andreas Griewank and Andrea Walther. Evaluating Derivatives: Principles
and Techniques of Algorithmic Differentiation (2 ed. Evaluating Derivatives: Principles
and Techniques of Algorithmic Differentiation (2 ed. Other Titles in Applied
Mathematics, Vol.Where the paper says it
Where the paper says it
Where the paper says it
[13] Ernst Hairer, Syvert P. Nørsett, and Gerhard Wanner. p. 8 9780898717761
[13] Ernst Hairer, Syvert P. Nørsett, and Gerhard Wanner. Nørsett, and Gerhard Wanner. Solving Ordinary
Differential Equations I: Nonstiff Problems (2 ed. Solving Ordinary
Differential Equations I: Nonstiff Problems (2 ed. Springer Series in Computational
Mathematics, Vol.Where the paper says it
Where the paper says it
Where the paper says it
[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. p. 8 1007/978-3-540-78862-1
[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic
Models. Denoising Diffusion Probabilistic
Models. In Advances in Neural Information Processing Systems, Vol. Denoising Diffusion Probabilistic
Models. In Advances in Neural Information Processing Systems, Vol. Denoising Diffusion Probabilistic
Models. In Advances in Neural Information Processing Systems, Vol. Denoising Diffusion Probabilistic
Models. In Advances in Neural Information Processing Systems, Vol.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[15] Yilie Huang. 2025. Continuous-Time Reinforcement Learning for Asset–Liability Management. p. 1 Continuous-time reinforcement learning (RL) connects stochastic control to data-driven decision making [38, 39]. Financial applications include mean–variance portfolio selection [16, 18] and
asset–liability management [15]. Continuous-Time Reinforcement Learning for Asset–Liability
Management. In Proceedings of the 6th ACM International Conference on AI in
Finance.Where the paper says it
Where the paper says it
[16] Yilie Huang, Yanwei Jia, and Xun Yu Zhou. 2022. Achieving Mean–Variance Efficiency by Continuous-Time Reinforcement Learning. p. 8 3770337
[16] Yilie Huang, Yanwei Jia, and Xun Yu Zhou. Achieving Mean–Variance Efficiency by Continuous-Time Reinforcement Learning. Achieving Mean–Variance Efficiency by Continuous-Time Reinforcement Learning. In Proceedings of the Third
ACM International Conference on AI in Finance (New York, NY, USA).Where the paper says it
Where the paper says it
[17] Yilie Huang, Yanwei Jia, and Xun Yu Zhou. 2025. Sublinear Regret for a Class of Continuous-Time Linear–Quadratic Reinforcement Learning Problems. p. 8 3561760
[17] Yilie Huang, Yanwei Jia, and Xun Yu Zhou. Sublinear Regret for a
Class of Continuous-Time Linear–Quadratic Reinforcement Learning Problems.Where the paper says it
[18] Yilie Huang, Yanwei Jia, and Xun Yu Zhou. 2026. Mean–Variance Portfolio Selection by Continuous-Time Reinforcement Learning: Algorithms, Regret Analysis, and Empirical Study. p. 8 1137/24M1695075
[18] Yilie Huang, Yanwei Jia, and Xun Yu Zhou. Mean–Variance Portfolio Selection by Continuous-Time Reinforcement Learning: Algorithms, Regret Analysis,
and Empirical Study. Mean–Variance Portfolio Selection by Continuous-Time Reinforcement Learning: Algorithms, Regret Analysis,
and Empirical Study. Management Science (2026). Management Science (2026). https://arxiv. https://arxiv. org/abs/2412.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[19] Yilie Huang, Wenpin Tang, and Xun Yu Zhou. 2026. Adaptive Reparametrized Time for Score-Based Diffusion Sampling. arXiv:2607. p. 8 Adaptive Reparametrized
Time for Score-Based Diffusion Sampling. arXiv:2607. Management Science (2026). https://arxiv. LG] https://arxiv. org/abs/2607.Where the paper says it
Where the paper says it
Where the paper says it
[20] Yilie Huang, Wenpin Tang, and Xun Yu Zhou. 2026. ART for Diffusion Sampling: A Reinforcement Learning Approach to Timestep Scheduling. p. 8 02137
[20] Yilie Huang, Wenpin Tang, and Xun Yu Zhou. ART for Diffusion Sampling:
A Reinforcement Learning Approach to Timestep Scheduling.Where the paper says it
[21] Yilie Huang and Xun Yu Zhou. 2026. Amortized Guidance for Image Inpainting with Pretrained Diffusion Models. p. 8 [21] Yilie Huang and Xun Yu Zhou. Amortized Guidance for Image Inpainting
with Pretrained Diffusion Models.Where the paper says it
[22] Yilie Huang and Xun Yu Zhou. 2026. Data-Driven Exploration for a Class of Continuous-Time Indefinite Linear–Quadratic Reinforcement Learning Problems. p. 8 [22] Yilie Huang and Xun Yu Zhou. Data-Driven Exploration for a Class of
Continuous-Time Indefinite Linear–Quadratic Reinforcement Learning Problems. Data-Driven Exploration for a Class of
Continuous-Time Indefinite Linear–Quadratic Reinforcement Learning Problems. IEEE Transactions on Automatic Control (2026).Where the paper says it
Where the paper says it
[23] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. 2022. Elucidating the Design Space of Diffusion-Based Generative Models. p. 8 [23] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating
the Design Space of Diffusion-Based Generative Models.Where the paper says it
[24] Hans R. Künsch. 1989. The Jackknife and the Bootstrap for General Stationary Observations. p. 8 [24] Hans R. The Jackknife and the Bootstrap for General Stationary
Observations.Where the paper says it
[25] Shujian Liao, Hao Ni, Marc Sabaté-Vidales, Łukasz Szpruch, Magnus Wiese, and Baoren Xiao. 2024. Sig-Wasserstein GANs for Conditional Time Series Generation. p. 8 1214/aos/
1176347265
[25] Shujian Liao, Hao Ni, Marc Sabaté-Vidales, Łukasz Szpruch, Magnus Wiese, and
Baoren Xiao. Sig-Wasserstein GANs for Conditional Time Series Generation.Where the paper says it
[31] Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland Vollgraf. 2021. Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time Series Forecasting. p. 8 [31] Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland Vollgraf. Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time
Series Forecasting. Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time
Series Forecasting. In Proceedings of the 38th International Conference on Machine
Learning (Proceedings of Machine Learning Research, Vol.Where the paper says it
Where the paper says it
[32] Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. 2024. Align Your Steps: Optimizing Sampling Schedules in Diffusion Models. p. 8 html
[32] Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align Your Steps:
Optimizing Sampling Schedules in Diffusion Models. Align Your Steps:
Optimizing Sampling Schedules in Diffusion Models. In Proceedings of the 41st
International Conference on Machine Learning (Proceedings of Machine Learning
Research, Vol.Where the paper says it
Where the paper says it
[33] Timur Sattarov, Marco Schreyer, and Damian Borth. 2023. FinDiff: Diffusion Models for Financial Tabular Data Generation. p. 8 html
[33] Timur Sattarov, Marco Schreyer, and Damian Borth. FinDiff: Diffusion
Models for Financial Tabular Data Generation. FinDiff: Diffusion
Models for Financial Tabular Data Generation. In Proceedings of the Fourth ACM
International Conference on AI in Finance.Where the paper says it
Where the paper says it
[34] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. p. 8 3626876
[34] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano
Ermon, and Ben Poole. Kingma, Abhishek Kumar, Stefano
Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations.Where the paper says it
Where the paper says it
[35] Dejan Stancevic, Florian Handke, and Luca Ambrogioni. 2025. Entropic Time Schedulers for Generative Diffusion Models. p. 8 id=PxTIG12RRHS
[35] Dejan Stancevic, Florian Handke, and Luca Ambrogioni. Entropic Time
Schedulers for Generative Diffusion Models.Where the paper says it
[36] Vinh Tong, Trung-Dung Hoang, Anji Liu, Guy Van den Broeck, and Mathias Niepert. 2025. Learning to Discretize Denoising Diffusion ODEs. p. 8 52202/085713-1474
[36] Vinh Tong, Trung-Dung Hoang, Anji Liu, Guy Van den Broeck, and Mathias
Niepert. Learning to Discretize Denoising Diffusion ODEs.Where the paper says it
[37] Masaki Utumi, Ryuji Takaki, and Toshio Kawai. 1996. Optimal Time Step Control for the Numerical Solution of Ordinary Differential Equations. p. 8 [37] Masaki Utumi, Ryuji Takaki, and Toshio Kawai. Optimal Time Step Control
for the Numerical Solution of Ordinary Differential Equations.Where the paper says it
[38] Haoran Wang, Thaleia Zariphopoulou, and Xun Yu Zhou. 2020. Reinforcement Learning in Continuous Time and Space: A Stochastic Control Approach. p. 8 1137/S003614299426358X
[38] Haoran Wang, Thaleia Zariphopoulou, and Xun Yu Zhou. Reinforcement
Learning in Continuous Time and Space: A Stochastic Control Approach.Where the paper says it
[39] Haoran Wang and Xun Yu Zhou. 2020. Continuous-Time Mean–Variance Portfolio Selection: A Reinforcement Learning Framework. p. 8 html
[39] Haoran Wang and Xun Yu Zhou. Continuous-Time Mean–Variance Portfolio Selection: A Reinforcement Learning Framework.Where the paper says it
[40] Daniel Watson, William Chan, Jonathan Ho, and Mohammad Norouzi. 2022. Learning Fast Samplers for Diffusion Models by Differentiating through Sample Quality. p. 8 12281
[40] Daniel Watson, William Chan, Jonathan Ho, and Mohammad Norouzi. Learning Fast Samplers for Diffusion Models by Differentiating through Sample Quality.Where the paper says it
[41] Magnus Wiese, Robert Knobloch, Ralf Korn, and Peter Kretschmer. 2020. Quant GANs: Deep Generation of Financial Time Series. p. 8 id=VFBjuF8HEp
[41] Magnus Wiese, Robert Knobloch, Ralf Korn, and Peter Kretschmer. Quant
GANs: Deep Generation of Financial Time Series.Where the paper says it
[42] Christopher Williams, Andrew Campbell, Arnaud Doucet, and Saifuddin Syed. 2024. Score-Optimal Diffusion Schedules. p. 8 1730426
[42] Christopher Williams, Andrew Campbell, Arnaud Doucet, and Saifuddin Syed. Score-Optimal Diffusion Schedules.Where the paper says it
[43] Shuchen Xue, Zhaoqiang Liu, Fei Chen, Shifeng Zhang, Tianyang Hu, Enze Xie, and Zhenguo Li. 2024. Accelerating Diffusion Sampling with Optimized Time Steps. p. 8 52202/079017-3428
[43] Shuchen Xue, Zhaoqiang Liu, Fei Chen, Shifeng Zhang, Tianyang Hu, Enze Xie,
and Zhenguo Li. Accelerating Diffusion Sampling with Optimized Time
Steps. Accelerating Diffusion Sampling with Optimized Time
Steps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition. Accelerating Diffusion Sampling with Optimized Time
Steps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition.Where the paper says it
Where the paper says it
Where the paper says it
[44] Liangyu Yuan, Ruoyu Wang, Tong Zhao, Dingwen Fu, Mingkun Lei, Beier Zhu, and Chi Zhang. 2026. Few-Step Diffusion Sampling Through Instance-Aware Discretizations. p. 8 00792
[44] Liangyu Yuan, Ruoyu Wang, Tong Zhao, Dingwen Fu, Mingkun Lei, Beier Zhu,
and Chi Zhang. Few-Step Diffusion Sampling Through Instance-Aware
Discretizations.Where the paper says it
[45] Xinyu Yuan and Yan Qiao. 2024. Diffusion-TS: Interpretable Diffusion for General Time Series Generation. p. 8 pdf
[45] Xinyu Yuan and Yan Qiao. Diffusion-TS: Interpretable Diffusion
for General Time Series Generation.Where the paper says it
[16175 Accepted for publication; preprint available at arXiv:2412.16175]
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
-
Report-aware grid construction optimizes a positive grid by
transporting signed Richardson defects through a terminal-report adjoint, squaring their sum, and applying capped direct refinement.
This allows for an offline construction process that minimizes terminal report errors rather than just local discretization errors. -
Report-error theory shows that the
transported signed sum is the first-order term in terminal report error,
which provides a rigorous foundation for understanding why certain grid allocation strategies are superior to others. This theory enables the derivation of closed-form optimal grids under an aligned-defect model, quantifying the inefficiency of diagonal criteria with a factor rho greater than or equal to 1. -
The system can achieve lowest design and audit report errors across all 24 dataset–budget–report rows in a five-pipeline panel, showing cross-budget mean relative audit-error reductions of about 10–11% on Regime-SV and about 1% on Mag7. This demonstrates that the method consistently performs well regardless of the specific simulation regime or data context.
-
The system can perform exploratory allocation sensitivity analysis by testing a
H = 21 numerical bridge and exploratory selected setting allocation sensitivity,
showing that TAPES grids maintain the lowest design and audit errors across successive trading-day blocks under a fixed decision pipeline. This suggests that the report-space fidelity carries through to downstream decision layers. -
The AI system can be configured to use a continuous-time mean–variance reinforcement learning methodology for policy gradient allocation, where the grid differences are held fixed across sampling grids. This allows for scenario-trained RL agents to select optimal weights based on terminal reports derived from the optimized grids, rather than just raw simulation outputs.
Abstract
Financial diffusion samplers are often reused with a small number of function evaluations (NFE) inside scenario and risk pipelines, where numerical time-grid error can distort terminal quantities such as covariances, tails, and drawdown proxies. We study offline grid construction for a frozen probability-flow ODE: the score network, conditioning, Euler solver, report normalization, and deployment NFE remain fixed. TAPES (Terminal Adjoint-Propagated Error Scheduling) estimates signed local defects by Richardson comparisons, transports them through the remaining discrete flow and a terminal-report adjoint, and optimizes the squared norm of their sum before a capped direct refinement. We show that this transported signed sum is the first-order terminal-report error and quantify, under an aligned-defect model, the inefficiency of diagonal criteria. Across six budgets on a regime-switching simulation and chronologically held-out Mag7 contexts, TAPES gives the lowest design- and audit-report errors in all 24 dataset--budget--report rows within the reported five-pipeline panel. Cross-budget mean relative audit-error reductions are about 10--11% on Regime-SV and about 1% on Mag7, with Bonferroni intervals excluding zero; all displayed winners persist under a doubled-resolution reference, and a small Regime-SV Heun probe preserves the TAPES-first audit ordering. TAPES changes only the offline schedule and retains the prescribed deployment NFE; claims concern numerical fidelity to frozen generators.
Sources
- Mean--Variance Portfolio Selection by Continuous-Time Reinforcement Learning: Algorithms, Regret Analysis, and Empirical Study
- Adaptive Reparametrized Time for Score-Based Diffusion Sampling
Related papers
- AFNO: Universal Certifiably-Arbitrage-Free Operator Learning Framework for European Call-Price Surfaces
- Diffusion models for dynamic volatility surface generation and data-driven hedging
- Rethinking Synthetic Scenario Realism: Compatibility, Not Fidelity, Drives Hedging Performance
- CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining
- Diversification Across Predictors: Heterogeneous Ensemble Learning for Cross-Sectional Return Prediction
- DSINC beyond Heston: the 3/2 and other non-affine stochastic-volatility models