Reality Check And Validation Status

This project is best described as a reproducible research-engineering baseline. It is not a live trading system and does not claim independently verified, deployable alpha.

What Is Already Real

• The repository contains a runnable end-to-end factor-to-backtest pipeline.
• The factor engine, masking rules, bias-correction plumbing, ML baselines, Markowitz portfolio construction, and backtest metrics are implemented as auditable Python modules.
• CI runs linting, tests, and a CLI smoke test across Python 3.9, 3.10, and 3.11.
• The benchmark tooling reports environment details such as CPU, GPU, PyTorch, CUDA, OS, and command line.
• Public-data examples use reproducible commands and document expected outputs.

What Is Still A Smoke Test

• configs/small.yaml is a synthetic quick start. It proves that the pipeline runs; it does not prove strategy profitability.
• docs/public_data_mini_reproduction.md uses a tiny 10-name yfinance universe. It validates the public-data API path and factor IC workflow, not a robust trading result.
• The first maintainer benchmark is an engineering throughput benchmark for tensor factor primitives, not a trading-performance benchmark.

What Depends On Restricted Data

The paper-shaped result path targets a Wind / Tushare-style A-share panel. The repository ships loaders and reproduction instructions, but it cannot redistribute proprietary data. External readers therefore need their own licensed data to compare the paper-style empirical results quantitatively.

What Has Industrial Value

The industrial value today is infrastructure value:

• mask-aware cross-sectional factor computation
• handling for halts, missing cells, limit-up / limit-down cases, and tradability
• repeatable ML factor research workflow
• portfolio construction and vectorized backtest organization
• benchmark and validation harnesses that make assumptions explicit

This is useful as a starting point for research teams, students, and engineers who want a clean baseline to audit or extend. It is not sufficient by itself for live deployment.

What Is Required Before Live Use

Before anyone interprets results as deployable, they should add:

• licensed point-in-time data with delisted names and historical universe membership
• corporate-action, survivorship-bias, and lookahead-bias controls
• walk-forward validation across market regimes
• transaction-cost, slippage, borrow, liquidity, and capacity models
• risk limits, exposure controls, attribution, and monitoring
• independent reproduction by someone other than the maintainer

Current Validation Ladder

Level	Status	Meaning
CI smoke test	Complete	The package installs and core workflows run across supported Python versions.
Synthetic end-to-end pipeline	Complete	The full engineering path runs without private data.
Tensor factor benchmark	Started	One maintainer CPU baseline exists; more CPU/GPU reports are needed.
Public-data IC note	Started	A small yfinance example validates the public-data factor workflow.
Larger public-data walk-forward benchmark	Started	One 100-stock yfinance maintainer run exists; community runs on larger universes are invited.
Paper-style empirical reproduction	Data-gated	Requires Wind / Tushare-style data that cannot be redistributed.
Industrial production readiness	Not claimed	Needs execution, risk, monitoring, and independent validation layers.

The goal is to make every rung explicit so users can tell the difference between engineering reproducibility, public-data diagnostics, paper reproduction, and production readiness.