Source-linked AI summary
DSA: Evidence-Aware LLM-Agent Orchestration for Multi-Market Stock Research
Linsen Zhu, Yi Shi
TL;DR
Operational stock research requires evidence assembly, capability-aware routing, and controls around generated opinions rather than fluent summarization alone. DSA implements this evidence-aware workflow with default and agentic report profiles, and its frozen selected test manifest passed without failures while supporting bounded software-conformance conclusions.
Problem
Operational stock-research systems must handle heterogeneous evidence, unequal market and model capabilities, and controlled use of generated opinions.
Method
DSA orchestrates evidence acquisition, structured context construction, model-routed analysis, optional role and Strategy Skill reasoning, and report generation through profile-specific safeguards.
Results
0 failures, errors, or skips were recorded in 102.532 seconds for the selected frozen contract-test manifest, supporting bounded implementation-conformance conclusions.
Takeaways & Limitations
DSA provides an open-source reference implementation for evidence-aware LLM-agent orchestration, while analytical value remains to be established through benchmarks, comparisons, ablations, and report-quality evaluation.
Takeaways & Limitations
The evaluation establishes software conformance, not factual accuracy, forecast quality, investment performance, or superiority of multi-agent reasoning.
Abstract
from arXiv · showhide
Large language models can summarize financial information, but an operational stock-research system must first assemble heterogeneous evidence, expose unavailable data and model capabilities, and control how generated opinions affect a final report. We present DSA, an evidence-aware orchestration framework for multi-market stock research with large language model (LLM) agents. DSA organizes the workflow into evidence acquisition, structured context construction, model-routed analysis, optional role and Strategy Skill reasoning, and report generation with selected context and diagnostics. A default report profile and an optional agentic profile share evidence and model-routing services but use profile-specific output validation and risk safeguards. In the agentic profile, core role outputs are processed by role-specific parsers, whereas Strategy Skill opinions undergo an additional signal-eligibility partition before synthesis; disagreement is supplied explicitly to the decision agent, followed by a conservative risk override. The reference implementation includes six regional market paths, fifteen bundled Strategy Skills, hosted and local model routes, and multiple execution and delivery surfaces. At a frozen software snapshot, a selected manifest of 1,457 portable offline backend contract tests passed; 596 cases were retrospectively mapped to six contract families central to the reported LLM-agent architecture. This evidence establishes implementation conformance for the tested software contracts, not superior report quality, forecasting accuracy, or investment returns.
1. Introduction
DSA frames stock research as an evidence-to-report systems problem rather than a simple question-answering task. It proposes profile-specific orchestration and controls for heterogeneous evidence, models, capabilities, outputs, and risks.
- Motivation: Operational stock research must retrieve fallible market evidence, calculate indicators, route compatible models, manage analytical roles, and preserve inspectable context.Fluent generation alone is insufficient when unavailable data are treated as neutral evidence or generated opinions gain authority without explicit controls.
- Motivation: Unequal market fields, provider capabilities, and analytical extensions should remain observable instead of being hidden behind a uniform conversational interface.
- Architecture: DSA provides a five-stage evidence-to-report architecture for repeatable LLM-based stock research across heterogeneous markets and data sources.
- Control semantics: DSA separates model routing from research logic and gives configured core roles and optional Strategy Skills different output-admission mechanisms.
- Control semantics: Profile-level semantics govern evidence state, capability admission, output validation, risk handling, and observable failure without requiring identical implementations across profiles.
- Evaluation boundary: The reference implementation and frozen claim-mapped evaluation establish software-contract conformance, not analytical or financial effectiveness.
2. Related Work
Prior financial-agent systems explore domain evidence, memory, tools, specialized roles, collaboration, and trading-oriented evaluation. DSA shares some infrastructure and role-based ideas but focuses on how heterogeneous evidence and generated opinions move through an operational research workflow.
- Financial-agent foundations: FinGPT, FinAgent, and FinMem develop financial-language-model resources involving domain evidence, multimodal information, tools, memory, and character design.
- DSA's position: DSA adopts related components as infrastructure concerns while centering its claim on evidence and generated opinions moving through an operational research workflow.
- Specialized and collaborative agents: FinRobot, FinCon, TradingAgents, and AlphaAgents study layered platforms, managerial coordination, risk control, trading outcomes, equity selection, or portfolio construction.
- General agent infrastructure: ReAct, AutoGen, and AgentScope provide reasoning, action, coordination, and tool-enabled application abstractions, while FinRL and FinRL-Meta support data-driven trading research environments.
3. DSA Framework
DSA is an evidence-to-report workflow with default and agentic execution profiles that share evidence construction and model routing while applying distinct validation and risk controls. It preserves evidence availability, capability differences, role-specific outputs, disagreement, and degradation states as observable parts of analysis.
- Execution profiles: DSA offers default and agentic profiles that reuse evidence and model-routing services but differ in orchestration, validation, and safeguards.The default path uses fewer model calls and dependencies; the agentic path adds role agents, Strategy Skills, disagreement-aware synthesis, and post-synthesis risk control.
- Five-stage workflow: The workflow acquires market evidence, constructs bounded structured context, routes models and capabilities, optionally performs role reasoning, and renders reports with inspection records.Evidence includes quotes, bars, fundamentals, announcements, news, technical features, availability states, and selected diagnostics.
- Evidence and capability control: DSA keeps market-specific coverage and backend capabilities visible instead of assuming equivalent fields, service conditions, model behavior, or tool support.Unsupported operations are reported rather than silently replaced by semantically different operations.
- Evidence-state semantics: Evidence blocks distinguish available, missing, unsupported, fallback, stale, estimated, partial, and failed states so unavailable evidence is not silently treated as neutral.The retained context distinguishes retrieved evidence from supported failure and degradation states.
- Controlled agentic synthesis: Core role opinions are parsed by role-specific mechanisms, while optional Strategy Skill opinions undergo an additional signal-eligibility partition before synthesis.The decision agent receives retained role opinions, eligible Strategy Skill opinions, and an explicit disagreement summary.
- Controlled agentic synthesis: The agentic risk step may preserve or reduce the initial decision, including vetoes or configured downgrades, but cannot upgrade it.RiskGuard operates on the initial state and risk context after synthesis.
4. Implementation and Reproducibility
DSA is organized as a Python reference implementation with explicit interfaces, multiple regional and operational paths, and documented licensing and frozen-environment details. Its reproducibility boundary is the recorded software snapshot and evaluation artifact rather than the moving public repository.
- Architecture: DSA organizes evidence construction, model routing, agents, Strategy Skills, report schemas, and rendering behind explicit interfaces in Python.The default and agentic profiles reuse provider and routing services while retaining distinct orchestration and safeguard implementations.
- Market and analysis scope: The implementation includes market paths for mainland China, Hong Kong, the United States, Japan, South Korea, and Taiwan, plus dedicated screening and portfolio structures.Batch analysis reuses the single-instrument analysis path more directly.
- Operational scope: DSA bundles fifteen Strategy Skills, multiple model routes, and command-line, API, browser, desktop, scheduled-job, and messaging interfaces.These surfaces demonstrate invocation and delivery scope, not evidence of AI quality; the evaluated system does not place or modify broker trades.
- Licensing: DSA-authored code uses the MIT License, while bundled AlphaSift-derived screening components use Apache License 2.0 with attribution and modification notices.The licensing boundary is recorded separately for original and derived components.
- Reproducibility: The conformance evaluation used frozen Git revision 0ca56cbee2dff5cf23b1fc59c16e2d48e61ba85c on 26 August 2026 with Python 3.10.9, pytest 7.1.2, and LiteLLM 1.89.3.The manifest, command, environment record, JUnit XML, checksums, and platform-boundary probe were retained; the public repository is moving.
5. Implementation Conformance Evaluation
The evaluation tests whether a frozen implementation exercises DSA’s described framework semantics, not whether it improves research quality or investment returns. A selected, claim-mapped portable manifest passed its executed cases, while platform, sampling, dependency, and open empirical boundaries constrain interpretation.
- Frozen evaluation scope: 1,457 portable cases executed with zero failures, errors, or skips in 102.532 seconds.The manifest collected 1,458 cases and explicitly deselected one case tied to an unsupported native-Windows route.
- Frozen evaluation scope: The manifest was selected after exploratory runs and is not the complete repository suite or a prospectively registered statistical sample.Counts describe executed test cases rather than independent observations.
- Scope boundaries: Platform-specific checks were separated from portable failures, exposing native-Windows and Unix-shell boundaries rather than supporting universal cross-platform passage.The boundary probe identified unsupported behavior in the Codex App Server route and Unix-shell assumptions in Docker-entrypoint tests.
- Claim-mapped contracts: The six-family subset exercises evidence-state handling, capability reporting, role sequencing, parsing, Strategy Skill validity partitioning, disagreement synthesis, and conservative risk transitions.Broader multi-agent and tool-surface cases remained in the 1,457-case record rather than this retrospective subset.
- Interpretation: Mocked, contributor-authored tests support software-contract conformance but cannot establish semantic quality under live data, models, or credentials.Many assertions exercise related code paths, and dependency sensitivity was observed across LiteLLM versions.
- Open questions: Comparisons of profiles, ablations of major components, provider trade-offs, and financial returns remain unevaluated.The endpoint studied is a research report and inspection record, while translating suggestions into orders and returns would require a separate downstream protocol.
6. Discussion
DSA integrates evidence-aware stock research with optional agent mechanisms while keeping execution profiles, model capabilities, evidence states, and safeguards explicit. Its evaluation supports tested software conformance, but important boundaries remain around analytical validity, temporal consistency, provenance, security, generalization, and reproducibility.
- DSA combines a repeatable evidence-to-report workflow with optional core roles, Strategy Skills, disagreement-aware synthesis, and post-synthesis risk control.
- The two profiles avoid making multi-agent dialogue mandatory for every request and establish a basis for future quality, latency, and cost comparisons.The present evidence does not show that decomposition improves analysis.
- Separating routing from research logic accommodates heterogeneous model capabilities without treating different routes as semantically equivalent.Model heterogeneity affects structured output, tool use, context limits, latency, and failure behavior.
- Current experiments establish software conformance rather than factual accuracy, forecast quality, investment performance, or superiority of multi-agent reasoning.
- Live providers and model endpoints can change, while evidence-state metadata does not create an immutable point-in-time dataset or eliminate look-ahead leakage.
- Optional snapshot persistence retains selected rather than complete context and does not guarantee exact replay, full event traces, or claim-level provenance.
- Bundled technical-analysis perspectives include disputed heuristics, and combining them with an LLM does not establish predictive value.
- External news, model outputs, and Strategy Skill packages may carry prompt injection, poisoned content, unsafe tool requests, or secrets; external packages remain trusted inputs.
7. Conclusion
DSA provides an open-source reference implementation for evidence-aware LLM-agent orchestration in multi-market stock research. Its frozen selected contract-test manifest supports implementation conformance for tested components, while analytical value requires broader point-in-time, quality, ablation, profile, and cost-aware evaluation.
- DSA connects heterogeneous evidence, model and capability routing, report profiles, optional agents, disagreement-aware synthesis, safeguards, and multi-surface delivery.
- The frozen selected contract-test manifest supports implementation conformance for tested components and boundaries, not analytical value.
- Establishing analytical value requires point-in-time benchmarks, profile comparisons, ablations, report-quality evaluation, and cost-aware analysis.
Declarations
The authors disclose their contributions, code availability, AI-assisted preparation, and the system’s research-only financial-use scope.
- Linsen Zhu led system conception, architecture, implementation, validation, and manuscript preparation, while Yi Shi contributed technical review and revision.
- Both authors reviewed the final manuscript and accept responsibility for its contents.
- The evolving implementation and tests are publicly available, but the evaluated revision and environment define the reported snapshot.
- DSA provides non-executable research opinions and allocation suggestions, but does not place, modify, or cancel trades or guarantee returns.
- Generative AI tools assisted language editing and code-to-manuscript consistency checks, while the authors independently verified the technical claims and conclusions.