Baostock A-Share Public-Data Validation Report

August 6, 2026 · View on GitHub

Units note. cost_drag in the tables below is cumulative over the run, not annualised — unlike ann_return, gross_ann_return and ann_vol beside it. Divide by the number of years before comparing it with a per-year figure, and do not compare it across reports of different lengths. The field is named cost_drag_cumulative in new output; see the metric glossary.

2026-07-16 · Windows 11 · PyTorch 2.13.0+cpu · Python 3.12.7

Command

python scripts/public_data_validation.py \
  --source baostock \
  --preset cn-large-25 \
  --max-tickers 25 \
  --cost-grid-bps 0,7,15,30 \
  --bootstrap-samples 100

Environment

Item	Value
OS	Windows 11 (10.0.26100)
Python	3.12.7
PyTorch	2.13.0+cpu
CUDA	Not available

Data Coverage

Field	Value
Source	baostock
Tickers	25 (SSE 15 + SZSE 10 A-share blue chips)
Date range	2021-01-04 to 2024-12-31
Trading days	969
Download coverage	all 25 requested tickers returned data
Stocks with no data	none

Results (7 bps effective cost)

strategy	ann_return	sharpe	max_dd	turnover	cost_drag
equal_weight	+1.7%	0.18	28.8%	0.001	0.07%
momentum_20	+0.2%	0.14	41.6%	0.179	24.2%
alpha101_mean	−23.2%	−1.02	73.1%	0.533	72.2%
mlp_alpha101	−11.9%	−0.81	53.5%	0.272	36.8%
transformer_alpha101	−6.0%	−0.38	35.3%	0.282	38.2%

The public A-share run is weak after costs: equal weight is only slightly positive, while higher-turnover strategies deteriorate materially after transaction costs. Treat this as a reproducibility and validation report, not an alpha claim.

Cost sensitivity

At 30 bps, high-turnover strategies suffer disproportionately:

alpha101_mean Sharpe drops from −0.21 (0 bps) → −3.62 (30 bps), with very large cost drag
equal_weight Sharpe barely moves (0.185 → 0.182), because turnover is minimal

Bootstrap CIs (100 samples, block 20 days) included for all scenarios.

Key packages

numpy==2.4.4 · pandas==3.0.2 · scipy==1.17.1 · scikit-learn==1.8.0 · torch==2.13.0+cpu · cvxpy==1.9.2 · scs==3.2.11 · click==8.4.1 · baostock installed (exact version not verified in this PR)