A-Share Medium/Low-Frequency Multi-Factor Alpha Research

An offline, reproducible study of cross-sectional A-share stock selection. The system combines 18 price, volatility, liquidity and technical factors with a Transformer ranking model, then evaluates a five-session Top-50 portfolio with turnover, transaction-cost and risk controls.

The public repository contains the research package, report generators, tests and the complete set of result tables and figures. Raw market data, trained models and prediction caches are intentionally excluded from version control. Once those private local artifacts are placed at the documented paths, no market-data API is required to reproduce the reports.

Research design

Component	Specification
Data	8,813,885 daily observations, 5,696 stocks, 2018-01-02 to 2025-12-31
Training / factor selection	2018-2021, with forward-target boundary purging
Validation	2022, retained as a full bear-market evaluation segment
Final test	2023-2025, excluded from model and checkpoint selection
Features	18 selected reversal, momentum, volatility, liquidity and technical factors
Model	20-session Transformer sequence model
Target	Signal after date-T close; buy at next adjusted open and exit at the sixth adjusted open
Portfolio	Top 50; Top-100 retention buffer; minimum 10% replacement per rebalance
Rebalance	Every five observed market sessions
Cost model	15 bps per one-way executed turnover
Risk control	15% volatility target; 10-rebalance lookback; 50% exposure cap after 15% prior drawdown
Benchmark	Equal-weight return of the available common stock universe

The factor pipeline applies point-in-time time-series feature construction, cross-sectional preprocessing and market-cap neutralization before sequence inference. The portfolio engine accounts for both stock replacement and exposure resizing in executed turnover. Maximum drawdown is calculated from a wealth curve that includes initial capital of 1.0.

See methodology for the full research contract and research integrity for evaluation boundaries, uncertainty analysis and known limitations.

Main results

The primary result is the risk-controlled portfolio at 15 bps. Validation and final-test results are reported together so that the 2022 bear market remains visible rather than being omitted from the headline evidence.

Period	Rebalances	IC	ICIR	Annualized net	Benchmark	Annualized net excess	IR	Net MDD
Validation 2022	48	0.0880	1.225	-3.62%	-12.49%	10.13%	0.729	24.09%
Final test 2023-2025	145	0.0752	0.851	28.39%	17.45%	9.31%	0.692	21.08%

Source: backtest.csv.

Calendar-year stability

Year	Research segment	IC	Annualized net	Benchmark	Annualized net excess	IR	Net MDD
2022	Validation	0.0880	-3.62%	-12.49%	10.13%	0.729	24.09%
2023	Final test	0.0633	24.27%	5.34%	17.96%	1.578	9.52%
2024	Final test	0.0654	3.41%	2.00%	1.38%	-0.027	18.07%
2025	Final test	0.0975	64.80%	51.14%	9.04%	0.790	16.49%

IC is positive in every reported year, while portfolio-level excess performance varies materially by market regime. Source: yearly_performance.csv.

Baseline comparison

All models use the same stock-date observations, target definition, rebalance schedule and no-risk portfolio engine. The comparison isolates ranking-model behavior from the portfolio risk overlay.

Final-test model	IC	Annualized net	IR	Net MDD
Equal-weight factor score	0.0914	-34.01%	-2.519	78.73%
Ridge (cross-sectional)	0.0778	19.58%	0.198	45.55%
Ridge (lagged)	0.0717	38.27%	0.987	36.93%
Transformer	0.0752	44.06%	1.771	32.51%

The Transformer does not have the highest mean IC, but it has the strongest portfolio-level return and information ratio among the declared baselines. Source: baseline_comparison.csv.

Robustness and attribution

Transaction costs

Final-test cost	Annualized net	Annualized net excess	IR	Net MDD
15 bps	28.39%	9.31%	0.692	21.08%
25 bps	25.17%	6.57%	0.487	21.44%
35 bps	22.02%	3.89%	0.280	21.81%

Performance remains positive at higher assumed costs, but the declining IR shows material turnover sensitivity. Source: cost_sensitivity.csv.

Risk ablation

Final-test configuration	Annualized net	IR	Net MDD	Average exposure
No risk overlay	44.06%	1.771	32.51%	100.0%
Volatility targeting only	29.82%	0.863	21.08%	77.5%
Drawdown scaling only	35.23%	1.124	24.30%	87.9%
Full overlay	28.39%	0.692	21.08%	73.9%

The overlay reduces drawdown and market exposure, but also reduces return and IR. It should therefore be interpreted as an explicit return-risk trade-off. Source: risk_ablation.csv.

Universe ablation

Excluding the daily bottom 10% by point-in-time circulating market capitalization leaves final-test IC at 0.0734 and annualized net return at 24.74%, versus 0.0752 and 28.39% in the full universe. The result is not solely driven by the smallest stocks, although the filter increases net MDD from 21.08% to 25.31%. Source: universe_ablation.csv.

Statistical uncertainty

Two thousand deterministic circular moving-block bootstrap draws produce the following 95% intervals:

Period	Net excess estimate	95% interval	P(excess > 0)	IC estimate	95% IC interval
Validation 2022	10.13%	[-13.48%, 41.32%]	76.65%	0.0880	[0.0718, 0.1049]
Final test 2023-2025	9.31%	[-4.52%, 24.43%]	89.70%	0.0752	[0.0602, 0.0900]

IC intervals remain positive, while excess-return and IR intervals include zero. The evidence is stronger for persistent ranking power than for a statistically conclusive portfolio-level excess-return claim over this sample. Source: bootstrap_uncertainty.csv.

Project structure

A_Share_Alpha_System/
├── data/raw/                  # Local A-share daily Parquet panel
├── models/                    # Factor configuration and Transformer checkpoint
├── artifacts/cache/           # Validated prediction ledger and metadata
├── reports/
│   ├── tables/                # Summary metrics and robustness tables
│   ├── ledgers/               # Rebalance-level audit trails
│   └── figures/               # Publication-ready charts
├── src/a_share_alpha/         # Research, model, backtest and reporting package
├── tests/                     # Integrity and regression tests
├── docs/                      # Methodology and research-integrity notes
├── pyproject.toml
└── requirements-research.txt

Generated tables use UTF-8 with BOM for convenient opening in Chinese-language spreadsheet software. Large local data and prediction-cache files are ignored by Git.

Offline reproduction

Python 3.11 is required. To create an isolated environment with uv:

uv venv --python 3.11 .venv
uv pip install --python .venv/bin/python -r requirements-research.txt

The following private artifacts are required for a full rerun and are not distributed through GitHub:

data/raw/a_share_daily_full.parquet
models/factor_config.pkl
models/transformer_alpha.pth
artifacts/cache/prediction_ledger.parquet
artifacts/cache/prediction_ledger.metadata.json

The committed result tables and figures can be inspected without these files.

Run the primary backtest and derived reports from the project root:

PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src .venv/bin/python -m a_share_alpha.backtest
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src .venv/bin/python -m a_share_alpha.cost_sensitivity
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src .venv/bin/python -m a_share_alpha.excess_metrics
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src .venv/bin/python -m a_share_alpha.risk_ablation
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src .venv/bin/python -m a_share_alpha.universe_ablation
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src .venv/bin/python -m a_share_alpha.robustness_report
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src .venv/bin/python -m a_share_alpha.research_audit

The baseline comparison is a separate, slower full-panel run:

PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src .venv/bin/python -m a_share_alpha.baseline_comparison

Run the test suite:

PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m pytest -q

These commands use only the local Parquet, model files and prediction cache. Use a_share_alpha.backtest --rebuild-cache only after an intentional change to source data, selected factors or the checkpoint.

Known limitations

• Historical ST labels, limit-up/limit-down tradability and suspension-state fields are unavailable in the local panel. ST filtering is therefore not evaluated or approximated using current security names.
• The target advances by each stock's observed rows. Its entry or exit date differs from the common market calendar for 0.61% of valid samples, including 0.55% in validation and 0.18% in the final test. See horizon_integrity.csv.
• The final test contains three calendar years. Portfolio-level bootstrap intervals remain wide.
• Results use one saved Transformer checkpoint; multi-seed training stability is not claimed.
• The benchmark is a research universe return rather than a directly investable index with constituent history and its own trading costs.

This repository documents a research backtest, not live-trading performance or investment advice.