GoAnt: Quality-Diversity Multi-Agent Search for Alpha Factor Discovery in Market Microstructure Data
cs.AI
Quick verdict
GoAnt claims quality-diversity multi-agent alpha mining lifting yield 57% and 97% on A-share data; protocol-rich but five-seed evidence and an unreported control keep it unverified.
Summary
GoAnt is a quality-diversity multi-agent framework for discovering symbolic alpha factors in market microstructure data. Non-communicating Explorer, Exploiter and Connector islands populate a growing MAP-Elites-style archive (Mental Map) whose cell centres are fixed at birth, while a compact orchestrator (Queen), a 1.5B Qwen2.5 model distilled from eight frontier LLMs, allocates the evaluation budget. A layered objective separates search fitness from post-cost net returns, countering the paper's 'Execution Trap'. On private A-share price-volume and order-book data (2023-2026), GoAnt reaches quality-weighted effective yields of 41.8 and 47.6 versus 25.1 and 24.2 for the strongest baselines, and out-of-sample retention of 0.64/0.67 versus 0.61/0.63 for a static map. Main limitations: only five seeds per arm with confidence intervals and significance tests not shown, a capacity-matched control experiment described but never numerically reported, and private data with no code release, making independent verification difficult.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 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
- LLM-driven alpha mining is jointly limited by an Execution Trap (high-RankIC factors turn unprofitable after costs) and a Reasoning Trap (general-purpose LLM coordinators homogenize populations), and both appear across price-volume and order-book data.
- A growing MAP-Elites-style archive with centres fixed at birth (the Mental Map) outperforms static, re-fitted and moving-centre maps under matched budgets, with adaptive cardinality - not centre motion - responsible for the gain (Tables 1-2).
- The distilled 1.5B Queen orchestrator beats UCB, EMA and Round-Robin schedulers under an identical Atlas, and a decision-level audit (Table S2) certifies it is not a UCB impersonator.
- GoAnt reaches quality-weighted effective yields of 41.8 (price-volume) and 47.6 (order-book), improving the strongest baseline (AutoAlpha: 26.6/24.2) by 57% and 97% under a matched 500-evaluation budget.
- Locked populations retain 0.64/0.67 of in-sample quality out of sample versus 0.61/0.63 for a static map, and optimizing raw net return directly collapses yield (Arm 7: 9.9/6.8).
Strengths
Unusually rigorous attribution protocol for the 'dynamic archive' claim
The 2x2 factorial (Table 2) separates centre mobility from adaptive cardinality; the static control is enforced by abort-on-violation end-of-run invariants (Table S1: centre displacement 0, cells 96=96) rather than by configuration label; the siting/capacity objection is addressed by pre-registered quantization-error admissibility (Appendix C).
Map-independent scoring metric with pre-registered thresholds
The effective-yield protocol (Appendix A) counts quality-filtered, pairwise-decorrelated factors (q* = 95th percentile of a random-genome null, frozen pre-run; rho = 0.7 with a 0.3-0.9 sweep) from the raw evaluation record, so archive-based and map-free arms share one ruler.
The Queen is audited, not asserted
The layered-objective claim is tested, not just argued
Mechanism-level ablations with per-role attribution
Weaknesses
Five seeds underpin every headline number, with no dispersion shown
The capacity-matched control is described but its results are never reported
Baseline LLM identity and teacher-overlap contamination are unaddressed
Private data, no code or data availability statement
Orchestrator cost is not accounted
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
GoAnt (arXiv 2609.08719v2) targets automated discovery of symbolic alpha factors under a fixed evaluation budget, framing two failure modes: the *Execution Trap* (strong-RankIC factors flip negative after execution costs, severe on order-book data) and the *Reasoning Trap* (general-purpose LLM coordinators homogenize populations). The system couples three non-communicating worker islands (Explorer/Exploiter/Connector), a MAP-Elites-style Mental Map that grows cardinality online with centres fixed at birth over six leakage-free descriptor axes, and a 1.5B Qwen2.5 'Queen' LoRA-distilled from eight frontier teachers. The central claim is architectural: adaptive capacity - explicitly separated from centre mobility and from the orchestrator - drives yield gains over LLM pipelines, non-LLM searchers, QD baselines and its own ablations, under matched budgets on A-share price-volume (PV) and L2 order-book data (2023-2026).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 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 is GoAnt?
GoAnt is a quality-diversity multi-agent search framework that discovers symbolic alpha factors (trading signals) in market microstructure data. Three non-communicating worker agents (Explorer, Exploiter, Connector) populate a growing MAP-Elites-style archive called the Mental Map, while a compact distilled language model called the Queen allocates the evaluation budget among them.
What is the 'Execution Trap' in alpha factor mining?
It is the paper's term for the failure mode where a factor with strong statistical predictive power (e.g., high RankIC) produces negative net returns once execution costs, slippage and market impact are priced in. The effect is mild on daily price-volume data and severe on intraday order-book data.
How much better is GoAnt than existing methods?
The paper reports quality-weighted effective yields of 41.8 and 47.6 on price-volume and order-book tasks, respectively - 57% and 97% above the strongest baseline under an identical evaluation budget. Out-of-sample retention is 0.64/0.67 versus 0.61/0.63 for a static-map variant. All figures are five-seed means on private A-share data (2023-2026).
Can GoAnt's results be reproduced?
Not independently at present: the paper gives extensive implementation detail but no code or data release, and all experiments use proprietary A-share price-volume and order-book data, so verification currently depends on the authors.
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.
- Only the main body, references and supplementary appendices (A-I) were extracted; figure 1 is rendered as caption text only. No claim in this review relies on the figure's visual content.
- Tables are linearized by PDF extraction; cell-to-column alignment was verified against in-text numbers (e.g., the 57%/97% margins against 41.8/26.6 and 47.6/24.2) before citing.
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.