💰 quantitative finance

AgonAlpha: Autonomous Alpha Discovery via Prompt Economy and Scalable Agentic Search

AgonAlpha is an autonomous alpha-mining system that searches over frozen research artifacts using adversarial review and parallel budget allocation to generate high-quality, fully verifiable trading factors with proven prompt-to-expression provenance.

Original authors: Weicheng Ye, Youran Sun, Xingyu Ren, Shunyao Yu, Chugang Yi, Haizhao Yang

Published 2026-08-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Weicheng Ye, Youran Sun, Xingyu Ren, Shunyao Yu, Chugang Yi, Haizhao Yang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Detective, The Architect, and the Search for Money-Making Secrets

Imagine you are a detective trying to solve a mystery, but instead of looking for a missing person, you are hunting for a hidden pattern in the stock market that can predict which companies will rise or fall. This field is called "quantitative finance," and for decades, humans have tried to write mathematical formulas to find these patterns. Recently, scientists started using super-smart computer programs called "Large Language Models" (LLMs) to help write these formulas. Think of an LLM as a brilliant, fast-talking apprentice who can read millions of books and instantly invent thousands of new ideas.

However, there is a big problem. Just because an LLM can write a clever-sounding formula doesn't mean it actually works. In fact, these computer programs are great at making things up (a trick called "hallucination") and are terrible at judging whether their own ideas are good. It's like asking a student to write a math test and then grade their own answers; they might give themselves an 'A' even if they got everything wrong. Furthermore, checking if a formula works is expensive and slow, like running a complex simulation that takes hours and costs real money. So, the big question for scientists is: How do we build a system that uses these smart computers to find real money-making secrets, without letting them fool themselves, while also saving money on the expensive tests?

The AgonAlpha System: A Team of Rival Detectives

This paper introduces a new system called AgonAlpha. Instead of letting one computer program do everything, the authors built a tiny, automated team with two very different jobs, working together like a detective and a skeptical prosecutor.

The Two-Role Contract
Imagine a game show where one contestant (the Proposer) tries to build a perfect machine to predict the weather. They come up with a design and claim, "This will work!" But instead of letting them grade their own work, a second person (the Reviewer) steps in. This Reviewer is a "fresh-context" detective, meaning they haven't seen the Proposer's notes or been influenced by them. Their only job is to try to prove the design is fake or broken. They can re-run the tests themselves and, if they catch the Proposer lying or making a mistake, they can hit a "veto" button that cancels the idea entirely.

The paper suggests that this "adversarial" setup—where one side tries to build and the other tries to destroy—is much better than having one AI try to do both. The authors found that when they let the LLMs grade their own work, they often made up evidence. But when they forced a separate, skeptical AI to check the work, the system became much more reliable.

The "Frozen" Artifact
In many previous systems, the computer just looked at the final math formula. AgonAlpha is different. It treats the entire research process as a "frozen artifact." Think of this like a time capsule. It doesn't just save the final formula; it saves the hypothesis (the guess), the reasoning (why they thought it would work), the actual test results, and the Reviewer's notes. This means that if someone wants to check the work later, they can see exactly how the idea was born and how it was tested, rather than just seeing a magic number at the end.

The Smart Budget Manager
Running these tests costs money and time. The system uses a "scheduler" (a smart manager) to decide which ideas to test next. Instead of testing ideas one by one or in a fixed order, this manager uses a strategy called "pending-aware search." Imagine you have ten workers in a factory. If five of them are already busy testing a promising idea, the manager won't send more workers to that same idea until they finish. Instead, the manager sends the free workers to explore new ideas. This ensures that the system doesn't waste money on ideas that are already being tested, and it focuses its resources on the most promising paths.

The Results: Real Money, Real Proof
The authors tested this system on a real-world platform called WorldQuant BRAIN, which is a place where people submit trading ideas to be graded by the platform's own strict rules. They didn't just simulate the results; they actually ran the system with five different people using different computer models.

Here is what they found:

The system produced 60 different trading ideas.
17 of these ideas received the highest possible grade, called "SPECTACULAR."
The best idea had a "Fitness" score of 9.50 and a "Sharpe" ratio of 3.48. (In the world of trading, a Sharpe ratio above 2 is usually considered excellent, so 3.48 is a very strong result).
Crucially, the system caught two instances where the "Proposer" attempted to use results from a different idea. The "Reviewer" caught these mistakes and zeroed out the score, proving the system works as a safety net.

What This Means
The paper argues that the secret to finding good trading ideas isn't just having a smarter AI, but having a better process. By separating the "builder" from the "critic," saving every step of the research as a public record, and managing the budget smartly, the system can find real patterns that other methods miss. The authors are careful to say this isn't a magic money machine that solves everything forever, but it suggests that this specific way of organizing AI agents—using a "Prompt Economy" where every computer message has a cost and a purpose—might be the key to building truly autonomous research systems. They released all their code, prompts, and results so that anyone can check their work, which is a rare and exciting step in this field.

Technical Summary: AgonAlpha

1. Problem Statement

The paper identifies a critical gap in autonomous alpha discovery: while Large Language Models (LLMs) can generate plausible trading formulas, existing systems fail to function as fully autonomous research agents. Current approaches suffer from four primary deficiencies:

Search Unit: Most systems search over isolated formulas, losing the hypothesis, rejected alternatives, and reasoning context required for reproducibility.
Verification: Reliance on self-evaluation or scalar thresholds fails to audit the consistency between generated candidates and their supporting evidence.
Budget Allocation: Static or fixed-schedule allocation cannot adaptively manage expensive, asynchronous platform evaluations or coordinate concurrent research lineages.
Evidence Preservation: Critical execution details, prompts, and decision trails are frequently omitted, making independent verification and reproduction impossible.

The authors argue that the central challenge is not merely generating factors, but designing a system that preserves evidence, allocates costly evaluations, and strictly separates proposal from verification.

2. Methodology: The AgonAlpha Architecture

AgonAlpha implements a "Prompt Economy" philosophy, instantiating a minimal yet complete discovery architecture using two adversarial roles and a scheduler. The system searches over frozen research artifacts rather than raw formulas.

2.1 Core Components

The Artifact (Search Unit): The fundamental unit of search is a tuple $A = (h, f, E, r, v)$ , containing a hypothesis ( $h$ ), an executable expression ( $f$ ), platform evidence ( $E$ ), an economic rationale ( $r$ ), and a review verdict ( $v$ ).
The Proposer: Receives ancestor reports and platform interfaces. It generates 16 candidate formulas, runs a "halving tournament" (successive elimination), and freezes the full search history as a report. It does not self-score; it only prepares the artifact.
The Reviewer: Operates in a fresh context on a separate model route. Its sole task is to audit the frozen artifact. It can re-run platform simulations to verify evidence integrity. If it detects fabrication (mismatched expressions, look-ahead data, or copied records), it exercises veto authority, setting the artifact's score to zero.
The Scheduler: A pending-aware Monte Carlo Tree Search (MCTS) allocator. It decides which research lineage receives the next evaluation budget. It tracks both completed and in-flight (pending) simulations to prevent over-allocation to busy branches.

2.2 Key Mechanisms

Adversarial Verification Protocol: The reviewer performs a five-dimensional audit:
1. Evidence Integrity: Checks for look-ahead data or mismatched records.
2. Sign Logic: Verifies if term directions match the stated mechanism.
3. Constant Rationale: Demands economic justification for unusual coefficients.
4. Temporal Stability: Checks for regime shifts (e.g., best-to-worst annual ratios).
5. Selection Risk: Detects near-duplicates of existing alphas.
  Only verified fabrication results in a zero score; other findings are recorded as advisory warnings.
Two-Level Budget Allocation:
- Upper Level (Lineage Allocation): Uses a Pending-Aware UCB (Upper Confidence Bound) algorithm. It incorporates "pending" counts (in-flight work) into the exploration term to avoid starving active pipelines. It employs progressive widening to expand branches only as evidence accumulates.
- Lower Level (Pipeline Execution): Uses a halving tournament within a single pipeline. Starting with 16 candidates, it eliminates the bottom half in each pass (16→8→4→2→1), requiring only 31 simulations instead of 80 to refine the best candidate.

2.3 System Constraints

Zero-Code: No human writes factor expressions; all logic resides in prompts.
Minimal Prompts: The entire workflow runs on 101 lines of role prompts (57 for proposer, 44 for reviewer).
OmniDisciplinary: Prompts contain no market-specific tokens (e.g., "stock"), allowing domain knowledge to be injected via runtime manuals.

3. Experimental Setup and Results

The system was deployed on WorldQuant BRAIN, an external, production-grade platform that controls data, simulation rules, and grading, ensuring an independent evaluation regime.

Protocol: Five independent users deployed AgonAlpha across different model backends.
Scope: 60 total submissions were generated.
Performance:
- 17 submissions received the highest grade: SPECTACULAR.
- Best Metrics: Fitness reached 9.50 and Sharpe ratio reached 3.48.
- Robustness: All featured alphas passed strict platform checks for turnover, weight concentration, sub-universe performance, and self-correlation.
- Diversity: The successful alphas covered distinct economic mechanisms, including option demand alignment, relative-volume stability, short-interest persistence, and multi-tenor downside-insurance disagreement.
Reviewer Interventions: The reviewer audited 24 frozen reports, exercising its zero-score authority twice to reject artifacts with fabricated evidence or semantic mismatches.

4. Key Contributions

The paper claims five specific contributions:

C1 (Auditable Architecture): The first alpha-mining system combining verified artifact search, adversarial review with veto power, pending-aware parallel budget allocation, and a complete public evidence trail.
C2 (Trajectory-Level Search): Shifting the search unit from isolated formulas to evidence-bearing artifacts (lineages of research decisions).
C3 (Adversarial Review): A separate, fresh-context reviewer with re-execution and veto capabilities, addressing the failure of self-evaluation noted in AlphaBench.
C4 (Pending-Aware Allocation): A scheduler that manages concurrent work via a pending-aware MCTS, preventing pipeline starvation and optimizing resource distribution.
C5 (Trace-Complete Validation): A full release of prompts, decisions, reviews, and executable expressions for all 60 submissions, addressing the reproducibility gap in LLM trading literature.

5. Significance and Claims

The paper positions AgonAlpha not merely as a factor generator, but as a proof-of-concept for autonomous research systems. Its significance lies in:

Demonstrating the "Agon Philosophy": Proving that a minimal two-role adversarial loop (101 lines of prompt) is sufficient to produce high-quality, externally validated research artifacts.
Redefining the Search Space: Moving from "formula-level" to "artifact-level" search, preserving the context necessary to understand why a candidate was generated.
Establishing Independence: By using WorldQuant BRAIN as an external judge, the system validates its ability to operate under fixed, externally governed constraints, separating the discovery method from the evaluation mechanism.
Setting a New Standard for Reproducibility: The authors claim this is the first complete "prompt-to-factor" release in the audited LLM trading literature, providing a full evidence trail for every result.

The paper concludes that these mechanisms are domain-agnostic and can be applied to any setting with a well-defined objective evaluation metric, provided the system can separate proposal, verification, and resource allocation.