Validation Digest - 2026-07-27

This digest summarizes the follow-up validation and discovery surface for v0.2.1. It extends the v0.2.0 validation digest with entrypoint fixes, visibility tracking, and lessons from external technical discussions.

It is not a trading-performance claim.

Scope

The digest covers changes and operating notes after v0.2.0:

• Colab Baostock demo repository URL fix
• Hugging Face paper entrypoint confirmation
• public discovery and traffic pulse
• external validation discussions around replay safety, rate limits, and backtest contracts
• current contributor funnel for benchmark and public-data reports

Current Public Reports

Report	Source	Universe	Purpose	Status
validation_synthetic_20260716.md	Synthetic	Generated GBM panel	Cross-platform CLI and metric reproducibility	Merged
validation_baostock_20260716.md	Baostock	A-share public data	Public-data validation on a China-market source	Merged
public_data_mini_reproduction.md	yfinance	ETF mini example	Factor IC and public-data smoke reproduction	Merged
Issue #22 blocker report	yfinance	ETF-50 attempt	Documents HTTP 429 rate limiting	Documented

Entrypoint Fix

The Baostock Colab demo now clones the canonical repository:

git clone https://github.com/initial-d/ml-quant-trading.git

This matters because Colab is a high-intent path: a user who opens the notebook is already trying to run the project. Sending that runtime to an old repository URL weakens reproducibility and makes traffic harder to attribute.

Outreach-Derived Validation Themes

Recent external technical discussions repeatedly pointed to the same validation contracts:

• separate valuation prices from execution prices when bars are missing;
• keep rebalance-calendar semantics explicit;
• report data-source blockers instead of producing empty benchmark claims;
• record data vintage, source, calendar, and universe assumptions;
• distinguish synthetic, fixture, public-data, and production-data evidence;
• keep paper/live execution claims separate from research backtests;
• require post-solve or post-backtest audit artifacts for optimizers.

These themes match the project's current stance: the repository should be easy to run and audit before it is impressive.

Current Public Discovery Pulse

Recorded on 2026-07-27:

• Repository pulse during outreach: 55 stars, 28 forks, 4 watchers.
• GitHub traffic snapshot in docs/visibility_status.md: 1,206 views and 246 unique visitors over GitHub's rolling 14-day window.
• Clone snapshot: 465 total clones and 249 unique cloners over the same window.
• High-interest paths include the Chinese README, factor handbook, issue #22, discussion #13, and contributor PRs.

Contributor Gaps

The most useful next reports are still:

• CUDA GPU tensor-factor benchmark.
• Linux CPU benchmark on a common cloud instance.
• Apple Silicon benchmark with a larger panel.
• A rerun of ETF-50 validation after yfinance rate limits clear.
• Another public-data case study with clearly documented data provenance.
• A clean Colab/Baostock run report from a fresh runtime.

Reproduction Entry Points

python scripts/benchmark_tensor_factors.py --device auto

python scripts/public_data_validation.py \
  --source synthetic \
  --models equal_weight,momentum_20,alpha101_mean \
  --epochs 1 \
  --batch-size 4096 \
  --hidden 32 \
  --cost-grid-bps 0,7,15,30 \
  --bootstrap-samples 100

python scripts/public_data_validation.py \
  --source baostock \
  --preset cn-large-25 \
  --start 2021-01-01 \
  --end 2025-01-01 \
  --models equal_weight,momentum_20,alpha101_mean \
  --epochs 1 \
  --batch-size 4096 \
  --hidden 32 \
  --cost-grid-bps 0,7,15,30 \
  --bootstrap-samples 100

Report results in Discussions #13 or the pairing issue #22.