跳转至

Evidence

This page explains how to reproduce MingCang's key evidence claims offline, and links to the methodology documentation.

Quick Start: Reproduce the Demo Closed Loop

Run the following command from the repository root. No API keys or network access are required.

make reproduce-evidence

What it does:

  1. Points DATABASE_URL at the sample database (examples/sample_db/mingcang_demo.db).
  2. Runs scripts/reproduce_evidence.py, which reads and prints:
  3. The three demo stocks (600519 贵州茅台, 300308 中际旭创, 601318 中国平安)
  4. The ForwardThesis for 300308 — including its three falsification conditions and follow-up metrics
  5. The ReviewCase outcome and attribution
  6. The MemoryPromotionCandidate status (pending — not yet trusted)
  7. The recorded rationale for WEIGHT_QUANT = 0.0

If the demo database does not exist, seed it first:

DATABASE_URL=sqlite:///$(pwd)/examples/sample_db/mingcang_demo.db \
    PYTHONPATH=. python scripts/demo_seed.py

Then re-run make reproduce-evidence.

What the Evidence Covers

1. Quant Layer Off (WEIGHT_QUANT = 0.0)

The quantitative scoring layer is disconnected from the production signal. The decision is based on three hard-gate checks that all failed:

Metric Measured Gate Result
IC (Information Coefficient) 0.0228 >= 0.04 FAIL
ICIR 0.062 >= 0.40 FAIL
Decile monotonicity non-monotonic required FAIL

In addition, a regime sign-flip was detected: the quant factor's direction reversed between bull-market and range-bound windows, making a small non-zero weight worse than zero in expectation.

Full methodology: docs/evidence/m29_quant_off.md

2. Demo Closed Loop

The sample database demonstrates the full L0–L4 research loop:

  • L2 (Thesis): a ForwardThesis with explicit falsification conditions, not a vague "watch this stock"
  • L4 (Review): a ReviewCase recording outcome and attribution
  • L0 (Memory gate): a MemoryPromotionCandidate in pending state — the system surfaces candidates but never auto-trusts them

Full walkthrough: docs/evidence/reproducible_closed_loop.md

3. Forward Validation Methodology

A short methodology note covering:

  • Point-in-Time (PIT) feature discipline
  • Pre-registration before shadow runs
  • The three-phase validation cycle (pre-register → shadow run → promotion gate)

Full methodology: docs/evidence/sample_forward_validation.md

What make reproduce-evidence Does NOT Claim

  • It does not re-run a live backtest. The demo database contains static seed data, not a live market feed.
  • It does not prove statistical significance. Three stocks and one review case are illustrative.
  • The MemoryPromotionCandidate being present does not mean the lesson is trusted or affects any decision. source_trust = "pending" means no production effect.

Production Signal Profile

For reference, the current production signal weights are:

Component Weight
Technical 0.6
Sentiment 0.4
Quant 0.0 (disconnected)
ATR trailing stop 2.5×

These are set in backend/config.py and can be overridden via .env.