Loopera

Factor Mining, Infinite Insight
An intelligent agent for fundamental factor research
_{Fundamental research, engineered for evidence.}

Product Overview · Product Framework · Research Inputs · Case Study · Boundaries

Repository scope: This repository contains public product and technical materials for Loopera; it is not a source-code distribution. It currently provides no runnable code, installer, CLI, or public API. The workflows and case studies in this README describe product capabilities and evidence boundaries.

Product Overview

Loopera works with public-company financial statements and operating information to help fundamental quantitative teams continuously perform research-direction discovery, economic-hypothesis formation, candidate-factor construction, trusted validation, and knowledge accumulation.

Core Value

Who It Is For

• Fundamental quantitative teams: expand research coverage and establish a consistent, reviewable factor-research process;
• Investment-research and fintech institutions: connect internal reports, research views, and data capabilities to a standardized validation pipeline;
• Research leaders: track sources, decision rationales, failure reasons, and next steps—not just final backtest results.

_{Loopera organizes research inputs, fundamental data, hypothesis discovery, factor design, trusted validation, and research memory into a continuously learning Agent research loop.}

Its goal is not to mass-produce formulas or replace research judgment with one attractive backtest. Loopera asks whether a study has clear fundamental logic, survives multiple forms of scrutiny, and contributes information beyond existing research.

Loopera turns fundamental research from isolated manual experiments into a continuous process that can accumulate and evolve.

Why Start from Fundamentals

Prices reflect outcomes; fundamentals explain how a business creates, consumes, and reallocates value.

Loopera studies earnings quality, cash flow, balance-sheet structure, operating efficiency, capital investment, and changes in business conditions. Rather than looking only at market prices, it focuses on developments inside the business that may not yet be fully understood or priced.

Fundamental data is not simple. Relationships must be established across statements; the same accounting number can mean different things across industries and business stages; and the publication time of information determines whether a study could have used it in reality. Loopera brings these issues into one research process so that the Agent behaves more like a fundamental researcher than a generic curve-fitting tool.

What Loopera Can Do

Move from operating phenomena to trusted conclusions, then turn every study into team knowledge.
_{Discover → Hypothesize → Build → Validate → Remember → Evolve}

Three Inputs, One Research Thread

Topic  ·  Prompt  ·  PDF   →   Unified Research Input   →   Hypotheses, factors, and validation

Product Framework

Loopera turns research intent into evidence-qualified, reusable research assets.
_{Understand the business → Form a hypothesis → Test the evidence → Compound the knowledge}

A Closed-Loop Product Workflow

flowchart LR
    A["Research Input<br/>Topic · Prompt · PDF"] --> B["Fundamental Understanding<br/>Fields · Statements · Industry Context"]
    B --> C["Intelligent Research<br/>Phenomena · Mechanisms · Candidates"]
    C --> D["Trusted Validation<br/>Timing · Logic · Robustness · Incremental Value"]
    D --> E["Research Memory<br/>Results · Failures · Lineage"]
    E -. "Feeds the next cycle" .-> C

    classDef input fill:#E8F7FF,stroke:#168CFF,color:#07142E,stroke-width:1.5px;
    classDef research fill:#F1ECFF,stroke:#7C3AED,color:#07142E,stroke-width:1.5px;
    classDef evidence fill:#E8FFF7,stroke:#0F9F8F,color:#07142E,stroke-width:1.5px;
    classDef memory fill:#FFF2FB,stroke:#C13C8A,color:#07142E,stroke-width:1.5px;
    class A,B input;
    class C research;
    class D evidence;
    class E memory;

The endpoint of the product loop is not a factor; it is a reviewable research asset. Every run preserves its question, hypothesis, evidence, conclusion, and next steps.

The public version describes product capabilities and research principles only. Loopera's specific hypothesis structures, validation strategies, evaluation rules, model-collaboration mechanisms, and factor implementations are proprietary and are not disclosed here.

Technical Framework: Hypothesis-Driven, Evidence-Gated, Memory-Augmented

The core capability does not come from letting one model improvise; it comes from the coordination of an Agent, deterministic programs, and research memory.
_{A research environment around the model—not a model wrapped around a backtest.}

How the Three Capabilities Work Together

flowchart LR
    A["AI Agent<br/>Propose · Explain · Revise"] --> B["Research Contract<br/>Hypothesis · Fields · Direction · Constraints"]
    B --> C["Deterministic Harness<br/>Compute · Check · Evaluate"]
    C --> D["Research Memory<br/>Results · Failures · Lineage"]
    D -. "Provides research context" .-> A
    H["Human Researcher<br/>Review · Intervene · Decide"] <--> B

    classDef agent fill:#F1ECFF,stroke:#7C3AED,color:#07142E,stroke-width:1.5px;
    classDef contract fill:#E8F7FF,stroke:#168CFF,color:#07142E,stroke-width:1.5px;
    classDef evidence fill:#E8FFF7,stroke:#0F9F8F,color:#07142E,stroke-width:1.5px;
    classDef memory fill:#FFF2FB,stroke:#C13C8A,color:#07142E,stroke-width:1.5px;
    classDef human fill:#FFF8E6,stroke:#D98E04,color:#07142E,stroke-width:1.5px;
    class A agent;
    class B contract;
    class C evidence;
    class D memory;
    class H human;

• AI Agent expands the research space by proposing, explaining, and revising ideas;
• Research Contract turns natural-language claims into structured objects that can be computed and reviewed;
• Deterministic Harness performs computation and objective checks so that narrative cannot substitute for evidence;
• Research Memory supplies results, failures, and lineage as context for the next cycle;
• Human Researcher retains review, intervention, and final decision rights at critical checkpoints.

System Modules

Nine modules form a three-layer architecture: the top layer handles interaction and orchestration, the middle layer performs research reasoning, and the bottom layer supplies data, evaluation, and long-term memory.
_{Each layer can evolve independently while sharing the same structured research objects.}

Three-Layer System Architecture

flowchart TB
    subgraph L1["01 · Research Interaction & Orchestration"]
      direction LR
      A["Research Input & Material Understanding"] --> B["Agent Orchestration"] --> C["Traces & Reports"]
    end

    subgraph L2["02 · Hypothesis & Factor Research"]
      direction LR
      D["Hypothesis Research"] --> E["Factor Construction"] --> F["Trusted Validation"]
    end

    subgraph L3["03 · Data, Evaluation & Memory"]
      direction LR
      G["Data & Field System"] --> H["Backtesting & Incremental Evaluation"] --> I["Research Memory & Planning"]
    end

    A --> D
    B --> D
    E --> H
    F --> H
    H --> C
    G --> D
    G --> E
    I -. "Feeds back research context" .-> D

    classDef experience fill:#E8F7FF,stroke:#168CFF,color:#07142E,stroke-width:1.5px;
    classDef engine fill:#F1ECFF,stroke:#7C3AED,color:#07142E,stroke-width:1.5px;
    classDef foundation fill:#E8FFF7,stroke:#0F9F8F,color:#07142E,stroke-width:1.5px;
    class A,B,C experience;
    class D,E,F engine;
    class G,H,I foundation;

Modules connect through structured research objects. Models, data sources, and evaluation components can therefore be upgraded independently without rewriting the entire research process.

PDF / Prompt Research Input Framework

Different forms of research material enter the same verifiable, traceable research thread.
_{Bring your question, your thesis, or an entire research report.}

Research Entry Points

From Material to Research Evidence

flowchart LR
    subgraph INPUT["Research Entry Points"]
        direction TB
        A["🎯 Topic"]
        B["✍️ Text Prompt"]
        C["📄 PDF"]
    end

    A --> D["Unified Research Input<br/>Identity · Summary · Source Fingerprint"]
    B --> D
    C --> D
    D --> E["Material Understanding<br/>Claims · Mechanisms · Candidates"]
    E --> F{"Conservative Routing"}
    F -->|Candidate is explicit| G["Priority Validation"]
    F -->|Primarily a viewpoint| H["Hypothesis Context"]
    F -->|Mixed content| I["Mixed Path"]
    G --> J["🛡️ Standard Validation Pipeline"]
    H --> J
    I --> J
    J --> K["🧠 Results · Failures · Source Lineage"]

    classDef input fill:#E8F7FF,stroke:#168CFF,color:#07142E,stroke-width:1.5px;
    classDef process fill:#F1ECFF,stroke:#7C3AED,color:#07142E,stroke-width:1.5px;
    classDef decision fill:#FFF7E6,stroke:#D97706,color:#07142E,stroke-width:1.5px;
    classDef trust fill:#E8FFF7,stroke:#0F9F8F,color:#07142E,stroke-width:1.5px;
    class A,B,C input;
    class D,E,G,H,I process;
    class F decision;
    class J,K trust;

Four Processing Stages

Routing Outcomes

🛡️ Input safety: PDFs and Prompts can provide research evidence, but they cannot rewrite system rules, expand the data scope, or demand that validation be skipped.

✅ One standard: “Priority Validation” means only that a candidate can be constructed. Every candidate, regardless of source, still faces the same data, logic, accounting, historical-performance, and incremental-information checks.

🔗 End-to-end traceability: The final report identifies the source material, candidate-recognition outcome, necessary adaptations, blocking reasons, and final research disposition of every candidate.

How a Research Run Works

A research run is not “generate a formula + view a backtest.” It is an evidence-driven decision chain.
_{Context → Question → Mechanisms → Candidate → Evidence Gates → Memory}

flowchart LR
    A["01  Establish Data Context"] --> B["02  Formulate the Research Question"]
    B --> C["03  Compare Competing Mechanisms"]
    C --> D["04  Construct a Candidate"]
    D --> E["05  Pass Evidence Gates"]
    E --> F["06  Write to Research Memory"]
    F -. "Feeds the next cycle" .-> B

    classDef context fill:#E8F7FF,stroke:#168CFF,color:#07142E,stroke-width:1.5px;
    classDef research fill:#F1ECFF,stroke:#7C3AED,color:#07142E,stroke-width:1.5px;
    classDef gate fill:#E8FFF7,stroke:#0F9F8F,color:#07142E,stroke-width:1.5px;
    classDef memory fill:#FFF2FB,stroke:#C13C8A,color:#07142E,stroke-width:1.5px;
    class A context;
    class B,C,D research;
    class E gate;
    class F memory;

What the Agent Actually Produces

Loopera's smallest deliverable is not “a factor name + a backtest curve,” but a structured research record. Based on current run artifacts, a complete output usually includes:

Full Case Study: From a Fundamental Observation to a Candidate Factor

A real research trace: from an accounting anomaly to an evidence-qualified candidate.
_{The case summary, nine-step trace, validation results, and evidence boundaries are collapsed below.}

flowchart TB
    subgraph P1["01 · Discover & Explain"]
      direction LR
      A["1 Detect an Anomaly"] --> B["2 Compare Mechanisms"] --> C["3 Select a Testable Explanation"]
    end
    subgraph P2["02 · Build & Review"]
      direction LR
      D["4 Form a Research Contract"] --> E["5 Validate Data & Implementation"] --> F["6 Blind-Review the Economic Logic"]
    end
    subgraph P3["03 · Evaluate & Learn"]
      direction LR
      G["7 Evaluate Performance & Incremental Value"] --> H["8 Compare the Candidate Family"] --> I["9 Write to Research Memory"]
    end
    C --> D
    F --> G
    I -. "Feeds the next cycle" .-> A

    classDef discovery fill:#E8F7FF,stroke:#168CFF,color:#07142E,stroke-width:1.5px;
    classDef validation fill:#F1ECFF,stroke:#7C3AED,color:#07142E,stroke-width:1.5px;
    classDef learning fill:#E8FFF7,stroke:#0F9F8F,color:#07142E,stroke-width:1.5px;
    class A,B,C discovery;
    class D,E,F validation;
    class G,H,I learning;

Core Characteristics

Current Research Output

147 candidate attempts · 66 structured hypotheses · 37 research patterns · 5 candidates admitted to the research library
_{The complete candidate funnel, research coverage, and independent Agent evaluation are collapsed below.}

flowchart LR
    A["147 Candidate Attempts"] --> B["130 Valid Candidates"]
    A --> X["17 Did Not Produce Valid Candidates<br/>Execution or Generation Issues"]
    B --> C["125 Explicitly Stopped"]
    B --> D["5 Admitted to the Research Library"]

    classDef total fill:#E8F7FF,stroke:#168CFF,color:#07142E,stroke-width:1.5px;
    classDef stopped fill:#FFF1F2,stroke:#E5484D,color:#07142E,stroke-width:1.5px;
    classDef accepted fill:#E8FFF7,stroke:#0F9F8F,color:#07142E,stroke-width:1.5px;
    class A,B total;
    class X,C stopped;
    class D accepted;

Use Cases

Disclaimer

Public Disclosure Scope

Make Fundamental Research Compound

Start with a question worth explaining, and make every research cycle the foundation for the next.

_{Loopera is intended for research and technical evaluation. It is not investment advice and does not promise future returns.}