REVIEW 3 major objections 3 minor 1 references
Robust Market Making: To Quote, or not To Quote
T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read A market maker trained to sometimes refuse quotes, or quote only one side, earns better returns and Sharpe ratios than one that quotes continuously, while still meeting quoting-ratio requirements.
desk verdict A plausible extension of adversarial-RL market making with real practical appeal, but the provided full text is corrupted and the core claim hinges on whether the simulator charges for absence; worth a referee once the manuscript is readable. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the expanded action space. Instead of forcing the market maker to always post both sides, each time step is a choice among two or four quoting modes, including silence and single-sided quotes; adversarial reinforcement learning turns that expanded space into a policy that withholds liquidity selectively. Quoting-ratio requirements enter through the environment or reward design, so the agent must balance the benefit of silence against the obligation to quote often. The continuously-quoting agent serves as the control that isolates the effect of the refusal option.
What would settle it
Evaluate the trained refusal-enabled policy and the continuous-quoting baseline on a held-out set of adversarial environments never seen during training, or on a real order-book replay, with both agents forced to the same realized quoting ratio, such as 95%. If the refusal-enabled policy does not achieve better returns or Sharpe ratio than the baseline, the paper's central claim is not supported.
Extended reading notes
Core claim
The paper's central claim is that an automated market maker does not have to quote continuously to perform well under adverse conditions. It builds two new agents on top of a continuously-quoting baseline: one whose action set is {post bid and ask, refuse to quote}, and one whose action set is {post bid and ask, post only ask, post only bid, refuse to quote}. Both are trained with adversarial reinforcement learning in a model-driven simulator and compared with the baseline across several adversarial environments. The reported result is that occasional refusal improves returns and/or Sharpe ratios, and that learned quoting ratios are controllable and can reach 99.9% in some cases. The contrib
Load-bearing premise
The load-bearing premise is that the simulated adversarial environments used for training and evaluation capture the real costs, volatility regimes, and order-flow adversities a market maker would face when deciding to stop quoting; if they do not, the measured gains may not transfer out of the simulator.
Editorial extensions
If this is right
- If the claim holds, automated market makers can be made more resilient without changing inventories, pricing models, or exchange rules—only by allowing the policy to say no.
- Quoting-ratio obligations become a tuneable constraint: the same agent can be steered toward high quoting rates, such as 99.9%, or lower ones, while retaining some or all of the performance gain.
- Single-sided quoting is a meaningful intermediate tool, letting the agent stay present in the market while avoiding the more costly side of the book.
- In the tested adversarial environments, the refusal-capable agents outperform the always-quoting baseline, isolating the richer action space as the source of the improvement.
Reading between the lines
- If the reported effect transfers to real markets, minimum-quoting regulations should be treated as an endogenous constraint on action frequency rather than a reason to force continuous quoting; a concrete next step is adding an explicit compliance term to the reward and measuring the trade-off.
- The policy's choice of when to refuse is likely correlated with adverse-selection pressure; inspecting which simulated states trigger refusal—order-flow imbalance, spread width, recent losses—would make the mechanism testable rather than black-box.
- Single-sided quoting may be doing inventory-management work, such as quoting only the side that reduces a skewed book; comparing inventory-dependent action frequencies with an inventory-targeting baseline would test this.
- Because performance is measured inside the training simulator, the strongest extension would be evaluation on held-out adversarial environments or real order-book replays, which lies outside the paper's current scope.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper claims that enriching the action space of an adversarial-RL-trained market maker with the option to occasionally refrain from quoting, or to quote only one side, improves returns and/or Sharpe ratios relative to a continuously quoting market maker, while still satisfying market-rule quoting ratios. Two new agents are introduced: one that can either quote both sides or refuse to quote, and one that can also quote only the bid or only the ask. The comparison is carried out in model-driven adversarial environments built on the authors' prior work. The abstract reports improvements and quoting ratios reaching up to 99.9% in some cases.
Significance. If the central claim is correct, the paper identifies a practically relevant and underexplored degree of freedom in automated market making: the ability to skip quoting in adverse conditions without violating regulatory quoting obligations. The extension of the action space is a natural and potentially valuable contribution to the adversarial-RL market-making literature, and the two new agent designs are clearly described in the abstract. However, the significance cannot currently be assessed because the technical content needed to validate the claim is not available in the supplied manuscript: the body text is corrupted, no methods, environment equations, reward definitions, or experimental protocols are readable, and no statistical evidence is reported. The contribution is therefore plausible but unverified.
major comments (3)
- [Full text / all sections] The supplied full text is largely unreadable mojibake and even carries a different arXiv identifier (2508.16586, physics.bio-ph) rather than the paper's stated identifier. I cannot locate the environment dynamics, state/transition model, reward function, action parameterization, training hyperparameters, or evaluation protocol. This is a load-bearing gap: the central claim that refusing to quote improves returns depends on whether the simulator imposes costs on absence such as queue priority loss, re-entry latency, or information leakage. As it stands, the advantage of the no-quote action may be mechanical if the only obligation is a quoting-ratio constraint with no other absence costs. The authors must supply a clean, readable manuscript with full methods before the claim can be evaluated.
- [Abstract] The abstract's empirical claims are hedged and lack quantitative support: 'improves returns and/or Sharpe ratios' is not accompanied by any effect size, number of seeds, confidence intervals, significance tests, or environment parameter values. The phrase 'reaching up to 99.9% in some cases' is a best-case selection and does not indicate typical or worst-case quoting ratios. The comparison baseline ('continuously quoting MM') is not defined in terms of reward specification or training procedure. These omissions prevent the reader from assessing robustness and are central to the paper's contribution.
- [Evaluation / environments] The evaluation appears to be performed in the same model-driven adversarial environments used for training, and those environments inherit from the authors' prior adversarial-RL line. This creates a risk of in-sample overfitting to the adversarial generator. The manuscript should report out-of-sample or cross-environment robustness checks, including sensitivity to reward weights for quoting-ratio compliance, disturbance distributions, and parameter ranges. Without such evidence, the claimed improvement may be an artifact of the specific simulator rather than a general property of allowing intermittent quoting.
minor comments (3)
- [Abstract / quoting ratio] The term 'quoting ratio' should be defined precisely: is it time-weighted, volume-weighted, or trade-weighted, and over what time window? Also specify the market rules that determine the required ratio.
- [Abstract] 'Basically meet any market requirements' is vague. The paper should report the achieved quoting ratios for all agents, across environments, rather than only the maximum case.
- [General] The repeated formatting errors and the mismatched arXiv identifier in the supplied text need to be fixed before resubmission; the current version cannot be processed by a reader.
Circularity Check
No demonstrable circularity in the available text.
full rationale
The only reliably readable portion of the submission is the abstract; the supplied full text is corrupted mojibake and even carries a different arXiv identifier (2508.16586, physics.bio-ph), so no equations, reward functions, or fitted parameters can be inspected. Based on the abstract alone, the central claim is an empirical comparison: an RL market maker with an expanded action space (occasional refusal or single-sided quotes) is trained in adversarial environments and evaluated on returns and Sharpe ratio. This is a contingent simulation result, not a derivation that reduces by construction to its inputs. The phrase 'we build on this line of work' references prior adversarial-RL market making, but no specific self-citation theorem, uniqueness argument, or fitted-input-called-prediction is quoted. The quoting-ratio compliance claim (up to 99.9%) could be reward-shaped, but the abstract does not describe the reward function, so that would be speculation. Correctness concerns about environment fidelity and in-sample evaluation are external-validity risks, not circularity. Under the hard rule that circularity must be exhibited by quotation and a specific reduction, no circular step can be identified from the recoverable text.
Assumptions & free parameters
free parameters (3)
- Adversarial environment settings
- Reward weights for quoting-ratio compliance
- RL training hyperparameters
assumptions (3)
- domain assumption An adversarial-RL simulation is a faithful proxy for real market-making conditions
- domain assumption Adversarial training confers robustness that transfers outside the training distribution
- domain assumption Registered-Market-Maker obligations reduce to satisfying quoting ratios
Cite this review
Pith. "Pith review of Robust Market Making: To Quote, or not To Quote." pith.science (2026). https://pith.science/paper/SBLAZGQT
@misc{pith2026250816588,
author = {Pith},
title = {Pith review of: Robust Market Making: To Quote, or not To Quote},
year = {2026},
howpublished = {\url{https://pith.science/paper/SBLAZGQT}},
note = {Machine review of arXiv:2508.16588}
}
abstract
Market making is a popular trading strategy, which aims to generate profit from the spread between the quotes posted at either side of the market. It has been shown that training market makers (MMs) with adversarial reinforcement learning allows to overcome the risks due to changing market conditions and to lead to robust performances. Prior work assumes, however, that MMs keep quoting throughout the trading process, but in practice this is not required, even for ``registered'' MMs (that only need to satisfy quoting ratios defined by the market rules). In this paper, we build on this line of work and enrich the strategy space of the MM by allowing to occasionally not quote or provide single-sided quotes. Towards this end, in addition to the MM agents that provide continuous bid-ask quotes, we have designed two new agents with increasingly richer action spaces. The first has the option to provide bid-ask quotes or refuse to quote. The second has the option to provide bid-ask quotes, refuse to quote, or only provide single-sided ask or bid quotes. We employ a model-driven approach to empirically compare the performance of the continuously quoting MM with the two agents above in various types of adversarial environments. We demonstrate how occasional refusal to provide bid-ask quotes improves returns and/or Sharpe ratios. The quoting ratios of well-trained MMs can basically meet any market requirements, reaching up to 99.9$\%$ in some cases.
Reference graph
Works this paper leans on
-
[1]
������� ���������� �� �������� ������� �������� ������ ������� � ��� ������ ��������� ����������� �� ������� �������� ���� �������� �������� ������������� ����������� ������� ����� ������ ������� �������� ��� ��� ���� ������� ������� ���� ������ �� ����������� �� �� ������ ������ ���� �������� ����� ������ �������� �� ����� ���� ����� �� ��������� � �����...
work page Pith review arXiv 2025
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.