Robust Market Making with Hawkes Order Flow and Price Impact via Adversarial Reinforcement Learning
cs.LG
Quick verdict
Extends adversarial-RL market making with Hawkes flow, price impact, and an LSTM; strong synthetic CVaR gains, but env/architecture confound, no real data.
Summary
This paper extends adversarial reinforcement learning for market making beyond the ARLMM framework of Spooner and Savani: the simulated limit order book gains Hawkes self-exciting order arrivals and trade-induced linear price impact, and the market maker's policy — which observes only inventory and time over a 40-step window — is augmented with a 32-unit LSTM. The authors prove single-stage equilibrium existence (Sion, Glicksberg) for a reduced game and compare four methods (Fixed, Random, ARLMM, Ours) across 19 synthetic regimes using CVaR10% as the primary left-tail metric plus a one-sided bootstrap test for terminal inventory bias. Ours attains higher CVaR10% than ARLMM in 19/19 environments and than Random in 18/19 (losing only in G3_drift_down), e.g., 24.49 vs 15.15 in G6_hawkes_only; the bootstrap finds no directional-inventory confound (p≈0.50 in both comparisons). Main limitations: the comparison confounds a richer adversarial training environment with the LSTM architecture and includes no ablation isolating the latter; evaluation is entirely synthetic with no real order-book data; CVaR gains carry no confidence intervals or significance tests across only five seeds; and the closest prior work combining ARL market making with Hawkes processes (Wang et al., ICAIF 2024) is cited but not discussed.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Extending adversarial market-making training to Hawkes arrivals and price impact, with an LSTM over the 40-step observation window, yields higher CVaR10% than ARLMM in 19/19 and than Random in 18/19 synthetic evaluation environments.
- The extended framework admits Nash equilibria in the reduced single-stage game: pure-strategy under fixed parameters (via Sion) and mixed-strategy in general (via Glicksberg).
- A worst-case-aligned evaluation protocol (CVaR10% primary plus a one-sided bootstrap test of terminal inventory bias, B=20,000) finds no evidence that the left-tail gains are driven by directional inventory exposure (p≈0.50 in both pooled comparisons).
- ARLMM itself is not uniformly more robust than random regime sampling: it beats Random on CVaR10% in only 9 of 19 environments, with failures concentrated in Hawkes-dominated, drift+execution-pressure, and liquidity-structure regimes.
Strengths
Well-motivated, precisely scoped extension of a specific baseline
Table 1 maps exactly which microstructure mechanisms (Hawkes κH/γH, impact ξ, trade-to-intensity feedback) are added to ARLMM's (b, A, k), and Section 2 specifies the discrete-time dynamics (Eqs. 1–8) cleanly, with Poisson recovered as γH=0.
Evaluation protocol aligned with the worst-case training objective
Section 5.1 justifies CVaR10%/CVaR30% over Sharpe for adversarial training and adds a pooled one-sided bootstrap (B=20,000) on terminal-inventory mean shift (Eq. 14) — a targeted check against the classic RL-trading confound of buying return with directional exposure (p=0.504 and p=0.501, both comparisons).
Unusually complete result reporting, including unfavorable cells
Correctly scoped, if routine, equilibrium analysis
Weaknesses
Central attribution claim is confounded by design
Entirely synthetic validation
No inferential statistics on the headline metric
Interpretation exceeds the observation set
Nearest prior work cited but not discussed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper extends ARLMM (Spooner & Savani [6]) — adversarial RL market making over an Avellaneda–Stoikov-style simulator — along two microstructure axes: Hawkes self-exciting bilateral order arrivals (Eq. 4, parameters κH, γH) and linear trade-induced price impact (Eq. 1, coefficient ξ). To handle the resulting non-stationarity, the market maker's policy encodes a K=40 observation window with a one-layer, 32-unit LSTM. Three claims: (i) the richer adversarial environment plus LSTM improves left-tail returns over both ARLMM and random regime sampling; (ii) the reduced single-stage game admits Nash equilibria (Theorems 1–2); (iii) a stricter robustness protocol (CVaR10% primary, plus a terminal-inventory bootstrap test) shows the gains are not purchased by directional inventory exposure. All evidence comes from the authors' own synthetic simulator across 19 test environments; no real order-book data is used.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper add over ARLMM (Spooner & Savani 2020)?
ARLMM trains a market-making agent in a zero-sum game against an adversary that picks adverse environments with Poisson order arrivals and no price impact. This paper extends that adversarial training to Hawkes self-exciting order flow and trade-induced price impact, adds an LSTM over the agent's 40-step observation window, and adopts a left-tail (CVaR10%) evaluation protocol with a bootstrap check on inventory bias. In 19 synthetic regimes the new agent beats ARLMM in all 19 and random regime sampling in 18 of 19 environments.
Does the method work on real market data?
No. All experiments run in the authors' own synthetic limit-order-book simulator with prescribed parameter regimes (19 test environments, five random seeds). The paper cites data-driven testbeds such as mbt-gym but does not evaluate on real order-book data, so real-market performance remains untested.
Why do the authors use CVaR instead of the Sharpe ratio?
Because adversarial training is worst-case: the environment player deliberately selects unfavorable markets each episode. CVaR10%/30% summarize the left tail of the return distribution, matching that objective, whereas Sharpe mixes both tails. The protocol also includes a one-sided bootstrap test to verify that gains are not obtained by holding larger directional inventory.
Is the improvement due to the LSTM or to the richer training environment?
The paper cannot separate the two: the proposed method changes both the training environment (the adversary now picks six regime parameters instead of three) and the policy network (an LSTM encoding the history window), and no ablation isolates the LSTM's contribution. This is the review's principal methodological concern.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The abstract is truncated mid-sentence ('Bootstrap') in the extracted text; part of the contribution statement may be missing.
- Figures 1–2 (terminal-wealth KDE, γH–ξ heatmaps) cannot be analyzed from text; heatmap claims such as 'Ours exceeds ARLMM at all 25 grid points' rest on the prose only.
- Whether reference [36] is cited in the body could not be fully verified — it does not appear anywhere in the extracted text outside the reference list.
- The ± dispersion convention (std over seeds vs over trajectories) and per-cell evaluation sample sizes are not stated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.