X2Strategy Learning Lab

RAG × AI × Quant × Agent

Research into
strategy evidence.

A learning lab for understanding how research sources become structured hypotheses, reviewed code, historical backtests and reproducible diagnoses.

Trace the workflow →
Research documents flowing through retrieval, structured strategy extraction, human review, code, validation, backtest and reporting
Learning project · No live-trading authority
Origin under studyALAGENT-HKU open-source project
Academic contextHKU computer science × finance
Our roleStudy · reproduce · benchmark · adapt
Primary sourceInspect upstream ↗
01 / Research workflow

Specification before implementation

From a source to a testable claim.

  1. 01

    Parse

    Read papers, reports, drafts, notes or bounded research inputs.

  2. 02

    Extract

    Convert rationale, indicators, signals, execution and risk into a typed StrategySpec.

  3. G1

    Human Review

    Challenge economic logic, missing assumptions, ambiguity and data availability.

  4. 03

    Generate

    Create a code draft from the reviewed specification, not from an unconstrained prompt.

  5. 04

    Validate

    Check syntax, structure, indicators, timing, leakage and reproducibility risks.

  6. G2

    Human Review

    Confirm specification fidelity, timing, costs and metric correctness.

  7. 05

    Backtest

    Run a real historical engine and preserve configuration and artifacts.

  8. 06

    Diagnose

    Explain results, deviations, risk concentration and the next experiment.

02 / Two systems

Intelligence meets deterministic research

The Agent workflow cannot replace the Quant engine.

AI research layer

Understand and structure.

  • Retrieval and source context
  • Hypothesis extraction
  • Strategy specification
  • Code drafting
  • Research memory and reporting
Quant research layer

Compute and verify.

  • Governed data and validation
  • Features and signals
  • Position and cost rules
  • Historical backtest
  • Evaluation and benchmark
Core principleAgent proposes structure. Deterministic tools produce evidence. Humans retain research judgment.
03 / Benchmark

Does the Harness improve research?

Compare the workflow, not the marketing.

Workflow

Time, completion, first-run success, human edit rate, cost

Research

Hypothesis completeness, citations, specification fidelity

Validation

Leakage, timing, indicator and metric error detection

Quant

Return, Sharpe, drawdown, turnover, costs and robustness

04 / Pengyi integration

Learn upstream, own the interfaces

One case study. Six system contracts.

PRAG

Evidence retrieval

Source quality, retrieval and RAG evaluation.

PAT

Agent Harness

State, tools, permissions, review and traces.

PDAT

Data truth

Point-in-time datasets and quality contracts.

PAAT

Quant research

Features, Alpha, backtest and evaluation.

PBT

Benchmark governance

Comparable protocols and release evidence.

PCPT · PRT · PET

Production boundary

Portfolio, risk and execution remain separately governed.

Adopt

Typed specification, review questions and artifact lineage.

Adapt

Parser, validation and diagnosis for A-share multifactor research.

Rewrite

Point-in-time universe, costs, neutralization and portfolio handoff.

Reject

Silent defaults, ungoverned fallback and direct execution authority.

First real research case

Active trade intensity from reconstructed orders.

A 19-page sell-side factor report becomes a page-level evidence map, typed StrategySpec, tested order-reconstruction component and explicit replication-gap report.

Open Case 001 →

Public disclosure boundary

Study is not authorship.

This is an independent learning and integration project based on public upstream material. It does not claim affiliation, authorship, upstream contribution, live performance or production-trading capability.

Visit the official ALAGENT site ↗