Transient Local Spectral Sensitivity for Drawdown Early Warning: A Frozen-Protocol Audit

A Frozen-Protocol, Returns-Only Audit of Market-State Summaries and Local Sensitivity for Drawdown Early Warning This study examines whether a combination of rotation, participation entropy, stress measures, and a fragility feature (F) can describe market behaviour, and whether F improves the prediction of severe drawdowns beyond existing early-warning measures. The analysis uses daily value-weighted returns for ten U.S. industry portfolios from the Kenneth R. French Data Library, covering July 1, 1926, to September 30, 2026. The sample is split into an estimation period (1926–1949), a calibration period (1950–1969), and a testing period (1970–2026). Volatility states are estimated using a diagonal Kalman filter. A recursive least-squares response map is fitted online, with F defined as its largest five-step gain. All decision rules, thresholds, and sample splits were fixed before the main analysis and recorded using cryptographic hashes. Results H1: Describing market behaviour The results are inconclusive under the prespecified decision rule. Participation entropy remains close to ten, suggesting little evidence of a strongly concentrated market structure, while rotation is largely indistinguishable from noise. F is the only feature to pass the original block-shuffle surrogate test, with a Holm-adjusted p-value of 0.006. However, this result is sensitive to implementation choices. Using a 10−610^{-6} variance floor raises the adjusted p-value to 0.0804, while treating exact zero returns as missing gives 0.0306. The observed median F is also lower than the shuffled null distribution, suggesting less amplification rather than more. H2: Drawdown forecasting There is no evidence that F improves out-of-sample drawdown forecasts. Adding it to the richer baseline model reduces AUC by 0.009, with an adjusted interval of [-0.024, 0.004]. This falls short of the prespecified improvement threshold of 0.02. None of the 27 sensitivity settings reaches that threshold. Overall, the findings do not establish F as a useful addition to existing drawdown early-warning measures. Limitations and Scope The analysis uses a single historical return path and includes only 21 event clusters in the test period. Smaller effects may therefore be difficult to detect. A co-movement of R, H, S, and F around early 2020 is noted as a post hoc observation and is not used as evidence for either hypothesis. The study does not propose a trading strategy, make investment performance claims, or provide investment advice. Reproducibility and Data Availability The research archive includes SHA-256 hashes for the source files, derived data, frozen outputs, research protocol, analysis plans, amendment log, and executable files. Implementation corrections and two deviations in data access are also documented. The original Kenneth R. French Data Library files are not redistributed because of their source terms. A downloader is provided instead. Since the upstream archive may have changed, downloading the data again may not reproduce the exact historical snapshot used in the study. Code repository: https://github.com/FFTryhard/Avarent The code is released under the MIT License. The paper, figures, and derived results are released under CC BY 4.0. Competing Interests and AI Disclosure I previously developed a proprietary quantitative framework called Avarent, which has since been decommissioned. AI tools assisted with coding, analysis scripts, and manuscript drafting. The author reviewed the materials and takes full responsibility for the study and its conclusions. Included Files The research package contains the main paper (PDF), a supplementary methods and reproducibility document (PDF), and an archive of the code and derived results.