When Alpha Disappears: A One-Switch Benchmark for Decision-Time Leakage in Financial Backtests

Battle Count: 121

Highly relevant to quantitative trading. The paper directly addresses a critical failure mode in financial ML backtesting: protocol-induced inflation that makes strategies appear profitable when they are not. The finding that centered temporal features and same-day-open execution with post-open information cause massive Sharpe inflation (up to 29 points) is directly actionable for practitioners. The benchmark provides a concrete diagnostic tool for validating backtest integrity before deploying trading strategies. The EXEC_OPEN violation (using full OHLCV bar while assuming same-day open fill) is a particularly common and finance-native error in practice.

The benchmark design is conceptually straightforward (one-switch paired comparison) but requires careful implementation of five distinct protocol violations, proper walk-forward splits, feature-schema management with admissibility timestamps, graph construction variants, and execution/label alignment. The formal definitions in Appendix C provide implementation-level detail. Running the full grid (2 markets × 6 models × 2 horizons × 6 variants × 3 cost settings × 9 years) requires significant computational infrastructure. The diagnostic nature means no complex optimization is needed, but ensuring strict one-rule attribution across all components demands careful engineering.

The paper provides detailed formal definitions of the clean reference, five protocol switches, metric definitions, feature schema, model hyperparameters, walk-forward splits, and portfolio construction. Data panels (S&P 500, CSI 300) are publicly available. Split files and feature-schema files are mentioned as released. However, no explicit GitHub repository URL is provided in the extract. The benchmark design is highly structured with fixed hyperparameters across variants, aiding reproducibility.