PaperlayerView exact revision on arXiv

Source-linked AI summary

GoAnt: Quality-Diversity Multi-Agent Search for Alpha Factor Discovery in Market Microstructure Data

Stella Zhao, Tommy Sha

arXiv:2609.08719v1cs.AI

TL;DR

Factor discovery can optimize predictive proxies that fail after execution costs, while redundant exploration limits behavioral diversity. GoAnt addresses this with a quality-diversity multi-agent system using an adaptive MAP-Elites Mental Map, decoupled workers, and a distilled Queen orchestrator. Under matched budgets, it is evaluated with a map-independent metric that compares archive-based and map-free systems on the same terms.

  • Problem

    Predictive proxies can fail after execution costs, and single-agent search can collapse into localized sub-optima.

  • Method

    GoAnt couples decoupled generative Worker Ants and a distilled Queen Ant orchestrator with a MAP-Elites Mental Map that enforces diversity.

  • Results

    Under matched budgets, GoAnt is scored against external coordinators using a metric computed from evaluation records alone, while a map-free ablation isolates the niche-structured memory's contribution.

  • Takeaways & Limitations

    The attribution protocol separates centre motion from cardinality growth, holds the search operator fixed across geometry arms, and quantifies the price of adaptive structure.

  • Takeaways & Limitations

    The matched-budget analysis identifies the mechanism responsible for a difference without establishing how the result would change with the budget.

Abstract

from arXiv · show

Automated alpha factor discovery searches symbolic trading signals from price-volume panels and order-book data under a fixed evaluation budget. Existing single- and multi-agent program-search systems can overfit predictive proxies that fail after execution costs and repeatedly explore redundant factor families, limiting execution robustness and behavioral diversity. We introduce GoAnt, a quality-diversity multi-agent search framework that combines non-communicating Explorer, Exploiter and Connector workers with a shared adaptive Mental Map and a compact Queen dispatcher. The Mental Map organizes candidates by leakage-free execution profiles and retains one elite per niche, while the Queen reallocates the evaluation budget from explicit search-state summaries. We also define a map-independent effective-yield protocol that counts high-quality, mutually nonredundant factors directly from each method's evaluation records, giving archive-based and map-free systems the same ruler. On real A-share microstructure data spanning 2023--2026, GoAnt reaches quality-weighted yields of 41.8 and 47.6 in price-volume and order-book settings, improving the strongest baseline by 57% and 97% under matched budgets. Its locked populations retain 0.64 and 0.67 of in-sample quality out of sample, compared with 0.61 and 0.63 for a static map.

Introduction

GoAnt addresses execution-aware alpha discovery by combining decoupled multi-agent generation with an adaptive quality-diversity archive. Its central contribution is separating archive mechanisms and evaluating them across both price-volume and order-book data under matched budgets.

  • Strong predictive power can become negative net return after hidden execution costs, slippage, and market impact.
  • General-purpose multi-agent coordinators can suffer consensus bias and context squeezing, producing poor budget allocation and homogenized factors.
  • GoAnt uses Explorer, Exploiter, and Connector worker islands, a shared adaptive-capacity MAP-Elites Mental Map, and a distilled Queen Ant.
  • The paper separates centre mobility from adaptive cardinality and attributes them independently with a 2×2 factorial design.
  • A map-independent effective-yield protocol scores archive-based and map-free systems using quality-filtered, pairwise nonredundant factors from common evaluation records.

Related Work

Prior work explores automated factor search, collaborative LLM coordination, and quality-diversity archives, but typically fixes search granularity, solution-space partitions, and allocation policies. GoAnt instead adapts archive capacity while holding roles and generators fixed, and evaluates behavioral coverage across two data categories.

  • Existing factor-discovery systems operate at a single granularity and generally omit execution cost from the search objective.
  • Collaborative LLM frameworks vary communication topology but fix solution-space partitions and allocation policies in advance.
  • Quality-diversity methods differ in archive geometry, including fixed partitions, adaptive emitters, moving boundaries, and archive-free novelty search.
  • GoAnt fixes centres at birth, grows cardinality online, separately ablates centre mobility and adaptive capacity, and reports behavioral coverage rather than cross-sectional correlation.

Methodology: The GoAnt Framework

GoAnt performs budget-constrained quality-diversity search with independent worker islands, a shared leakage-free Mental Map, and a distilled Queen that allocates evaluations. Candidates are admitted as niche elites under execution-aware evaluation, while archive geometry and orchestration are assessed with explicit invariants and ablations.

  • GoAnt separates search into a MAP-Elites Mental Map, non-communicating Worker Ant generators, and a distilled Queen budget-allocation policy.
  • The Execution Trap arises because predictive proxies such as RankIC can diverge from execution-aware net return once costs are priced.
  • The Mental Map: The Mental Map embeds factors by leakage-free execution profiles and retains one elite per niche, using six standardized descriptor axes.
  • Ablations and scoring: At a matched budget, the study separates adaptive cardinality from centre motion, while fixed-grid scoring prevents each arm from being judged on its own partition.
  • Archive invariants: A candidate replaces a niche elite only when it has strictly higher score, yielding monotonic elite quality while fixed centres preserve this invariant throughout the run.
  • Worker Ants: Explorer, Exploiter, and Connector ants respectively pursue radical mutation, localized repair, and semantic crossover across distant niches.
  • Queen Ant: The Queen allocates evaluations among islands using state summaries and a distilled policy trained from frontier teacher models.
  • Queen Ant: Coverage counts a newly founded cell only after its elite clears the quality floor, so exploration consumes evaluation budget before producing credited coverage.

Experiments

The experiments compare coordination and memory mechanisms under matched conditions across price–volume and order-book data. They isolate archive structure, dynamism, evaluation, and out-of-sample retention using controlled ablations and locked populations.

  • Experimental design: Every arm uses the same data split, evaluator, feasibility gates, search fitness, and evaluation budget, varying only coordination and memory structure.Each arm is run with five independent seeds, with means and bootstrap confidence intervals reported over seeds.
  • Experimental design: The two categories share instruments and periods but use category-specific descriptor scaling and cost models fixed before arm comparisons.The data comprise real A-share microstructure observations from 2023–2026 in price–volume and order-book settings.
  • Evaluation protocol: The effective-yield ruler scores every arm from evaluation records using a common quality floor and fixed diversity grid, including map-free systems.Fitness is identically cost-deflated and redundancy is assessed consistently before comparison.
  • Coordination comparison: The coordination spectrum tests single-agent, static-topology multi-agent, non-LLM search, quality-diversity, random-search, GoAnt, and mechanism-ablated systems under matched budgets.The table separately reports distinct factors, quality-weighted yield, correlation among admitted factors, and archive cells where applicable.
  • Mechanism ablations: The 2 × 2 ablation independently varies centre mobility and adaptive cardinality, with GoAnt using fixed centres and online cell creation.Centre motion is treated as an ablation away from GoAnt, while static and periodically redrawn geometries provide controls.
  • Out-of-sample validation: Locked populations are replayed without refitting, re-selection, or re-ranking on a later disjoint window to measure out-of-sample retention.Retention compares held-out quality with in-sample quality for the same factor set.

Conclusion

The paper presents GoAnt as a matched-budget factor-mining system whose niche-structured memory is evaluated with a common record-based metric. Its attribution protocol separates archive mechanisms so gains can be assigned to specific design properties.

  • Conclusion: GoAnt decouples generative Worker Ants from a distilled Queen Ant and uses a MAP-Elites Mental Map to enforce behavioral diversity.The architecture combines fixed worker roles with shared niche-structured memory.
  • Conclusion: A record-based metric evaluates archive-based and map-free arms on identical terms, while a map-free ablation isolates the contribution of niche-structured memory.The comparison is made under matched evaluation budgets against external coordinators.
  • Conclusion: The attribution protocol separates centre motion from cardinality growth and defines frozen controls through end-of-run invariants.The search operator remains fixed across geometry arms, making the two archive properties independently interpretable.

Supplementary Material

The supplement documents the measurement, enforcement, auditing, ablation, statistical comparison, and implementation procedures supporting the paper’s experiments.

  • Supplementary Material: The supplementary material collects protocols for measurement, static-archive enforcement, orchestrator auditing, descriptor diagnostics, role ablations, seed dispersion, arm comparisons, and implementation details.It also includes hyperparameters and additional supporting material.

A Measurement Protocol

The protocol evaluates map-based and map-free systems with a common, quality- and redundancy-aware yield measure while enforcing auditable static-map, geometry, scaling, and orchestration controls.

  • Map-independent evaluation: The effective yield counts high-quality factors that are pairwise nonredundant, using a frozen q⋆ threshold and ρ = 0.7 on realized signal series.The measure applies equally to systems with and without an archive: a single strong factor scores 1, duplicates are constrained, and uncorrelated noise is filtered by q⋆.
  • Map-independent evaluation: Quality-weighted yield supplements factor counts by weighting retained factors according to their excess over q⋆.The count measures independent directions, while the weighted version captures their quality.
  • Ablation controls: Geometry ablations retain Connector edges, all ant roles, and budget accounting so observed differences remain attributable to centre movement or cell creation.Disabling edge recomputation changes the Connector operator and therefore confounds geometry with search behaviour.
  • Baseline comparability: Descriptor scaling is estimated once on the development split and frozen across arms, preventing adaptive maps from gaining resolution through coordinate rescaling.The static baseline is also tested at the adaptive arm’s realized cell count and is subject to a pre-registered siting admissibility check.
  • Static-control enforcement: Static-map arms keep centres, cell count, and boundaries fixed, admit candidates only into existing cells, and abort when invariants are violated.Out-of-bounds candidates are recorded rather than creating new cells, making the static comparison enforceable rather than nominal.
  • Orchestrator audit: The Queen is audited for parse success, confidence, fallback, and deviation from UCB, with failed anti-impersonation or audit-coverage assertions causing run exclusion.The audit distinguishes a silently default scheduler from a language-model policy that merely imitates UCB.

E Descriptor Contamination Diagnostic

The descriptor diagnostic shows that quality-contaminated axes can collapse archive geometry, motivating leakage-free descriptors and fixed centres that prevent later quality-driven partition deformation.

  • Descriptor contamination: Including RankIC and IC-series statistics in the descriptor made centre updates act like ascent on the fitness gradient and collapsed the archive.The diagnostic observed monotonic alignment between node centroids and ∇S, with axis–score correlations approaching one.
  • Experimental isolation: The ant-role ablation holds the archive, descriptor axes, evaluator, and gates fixed while varying only the active roles.This isolates role composition from descriptor geometry in the comparison.
  • Leakage-free redesign: The present descriptor excludes future-return labels and uses execution, structural, and position-shape axes that cannot be arguments of search fitness.Residual association remains measurable, but the call-graph criterion provides a direct leakage check.

F Ant-Role Composition

The role ablation indicates that the full Explorer–Exploiter–Connector colony performs best, with Connector-mediated recombination supporting broader search beyond already explored niches.

  • Single-role arms: Neither single role is sufficient: Exploiter alone reaches 11.6 q-wtd in PV and 10.4 in L2, while Explorer alone reaches 24.3 and 26.1.Refinement has little material without exploration, whereas exploration alone remains below the coordinated configurations.
  • Pairwise role comparisons: Removing Connector costs 8.7 q-wtd in PV and 10.9 in L2, compared with 6.0 and 7.4 when Exploiter is removed.The pair containing Connector also holds 116 and 128 cells, versus 92 and 101 for the pair without it.
  • Full-colony result: Recombination across niches converts proposals into new territory rather than duplicates, and the full colony is best in all four reported columns.The comparison uses matched archive, evaluator, gates, and budget conditions.

G Per-Seed Dispersion

The study treats seeds as the replication unit and reports uncertainty and arm comparisons from five independent runs. It explicitly limits interpretation when this small sample cannot cleanly separate methods.

  • Replication unit: Five independent seeds, 101 through 105, underlie every main-paper mean.The seed, rather than individual factors within a run, is the replication unit.
  • Dispersion: 95% bootstrap confidence intervals are resampled over seeds, while medians accompany means because factor quality is heavy-tailed.Resampling factors would incorrectly treat one search trajectory as multiple independent observations.
  • Arm comparisons: Two-sided Mann-Whitney U tests compare arms as independent samples, with rank-biserial correlation reported as effect size.Shared seeds do not make the runs meaningfully paired once their dispatch trajectories diverge.
  • Power limits: With five observations per arm, the smallest attainable two-sided p-value is 2/252 ≈0.0079.Conventional significance is reachable only when the two seed sets separate completely.
  • Interpretation: Narrow body-paper comparisons are treated as unresolved at five seeds rather than established findings.Per-seed values, intervals, and test statistics are retained in run records.
  • Budget accounting: Candidates failing any frozen admission gate still consume one evaluation-budget unit.This prevents arms from improving yield by proposing cheaply through rejected candidates.

H Implementation and Hyperparameters

The implementation fixes search settings before evaluation, admits candidates using a deflated predictive-quality score with a positivity modulation, and uses a frozen archive geometry, budget, seed set, and distilled Queen.

  • Fixed settings: All hyperparameters are fixed before runs and are not tuned by arm, seed, or data category.Category-specific quantities are identified explicitly when applicable.
  • Search fitness: The archive admits candidates using a deflated signal-quality score rather than realized return.The score uses the deflated t-statistic of rank information coefficient and the fraction of positive daily ICs.
  • Search fitness: A factor with randomly changing daily IC signs has p+ ≈1/2 and therefore S ≈0, regardless of pooled IC magnitude.Realized net return enters only during final verification, not archive admission.
  • Admission gates: Every candidate must clear identical frozen admission gates shared across arms.Rejected candidates remain budget-consuming evaluations.
  • Archive geometry: Candidates within niche radius 0.65 join the nearest cell; otherwise they found a new cell whose centre is never updated.Descriptor components are robustly standardized and clipped at ±3 standard units.
  • Budget and seeds: Each arm receives 500 evaluations per run across five seeds, counting rejected candidates as evaluator calls.The budget measures total evaluation effort rather than accepted-candidate successes.
  • Orchestrator: The 1.5B-parameter Qwen2.5-Instruct Queen is LoRA-adapted from eight teachers using textual orchestration-state summaries.The state includes pull counts, empirical Q-values, coverage, isolation, and stagnation.
  • Infrastructure: Candidate generation and evaluation are CPU-bound, while four A100 80GB GPUs serve Queen inference and LoRA adaptation.All arms share the symbolic operator grammar and vectorized Robust RankIC backtest.

I Anticipated Questions

The anticipated-questions section explains controls for capacity, descriptor leakage, archive interpretation, coverage inflation, Queen distinctiveness, and category scope. It also states that conclusions are tied to the matched 500-evaluation budget and two-category bracket.

  • Capacity control: The frozen tessellation is sized to the adaptive archive’s final cardinality, equalizing capacity before comparing when cells were created.This separates adaptive timing from a simple larger-archive advantage.
  • Descriptor leakage: Descriptor axes may use only quantities computable without reading future-return labels, verified by inspecting the call graph.Residual marginal association is reported as expected rather than treated as proof of leakage.
  • Archive geometry: A frozen-centre archive is not a pre-specified grid because cells arise at observed candidate locations, although neither moves an established centre.The distinction preserves the fixed-centre property needed by the elitist-monotonicity argument.
  • Diversity measurement: Archive entropy is omitted because one elite per occupied cell makes search-time entropy or coverage a restatement of cell count.Diversity is instead computed on a fixed external grid that no arm can influence.
  • Coverage control: A new cell counts toward coverage only after its elite clears the preregistered quality floor q⋆.This makes outlier-driven exploration contingent on subsequent quality rather than immediate occupancy.
  • Allocation objective: Coverage, quality, and redundancy are rescaled to [0, 1] so the convex weights α, β, γ cannot dominate through differing units.The normalization makes the allocation objective unit-balanced.
  • Queen validation: A deviation floor, parse-failure cap, and fallback cap test whether the Queen differs from a UCB bandit rather than assuming agreement demonstrates competence.The protocol measures both agreement and meaningful deviation.
  • Budget scope: The conclusions are tied to one matched budget of 500 evaluations, and the adaptive-versus-frozen crossover at other budgets remains future work.Growing archives trade evaluation effort between resolution and refinement, so the authors do not claim budget independence.
Loading 2609.08719v1…