Abstract
<jats:p>Single-stock studies collapse data availability, strategy rules, execution, forecasting, and investment conclusions into one analytical layer, allowing non-executable strategies to be proxied, in-sample rankings to be treated as deployable, and forecasting failure to be obscured by precise-looking price paths. This paper proposes an auditable framework based on point-in-time data governance, a strategy eligibility registry, out-of-sample competition, statistical and economic thresholds, and halt codes. The infrastructure includes 19 data tables, 13 source classes, and 15 quality gates. VALIDATION-HALT constrains strategy claims, while FORECAST-HALT constrains forecasting claims. A proof of concept uses 6,809 trading days for Yingluohua (000795.SZ) from August 8, 1997, to August 4, 2026. Of 155 candidate strategies, only nine qualified for backtesting. RSI(14) oversold mean reversion achieved an 11.05% compound annual growth rate and a 0.729 Sharpe ratio in the out-of-sample period, but generated only 20 position changes and lacked parameter stability. The 50/200-day moving-average crossover delivered a similar return with a −55.13% maximum drawdown. The inverse-error-weighted forecasting ensemble recorded a 2.604% return RMSE, an out-of-sample R² of 0.328% against a random-walk benchmark, and a Diebold–Mariano p-value of 0.515. Neither result satisfied the preregistered validation rules. The confirmatory design requires at least 2,000 point-in-time A-share stocks, 5 million stock-day observations, 1.5 million out-of-sample forecasts, 60,000 firm-quarter observations, and 200 out-of-sample trades per core strategy, with multiple-testing control, risk-factor adjustment, regime stratification, frozen prospective validation, and license-controlled reproducibility. The framework converts non-comparability, insufficient evidence, and failure to outperform into traceable decision states.</jats:p>