Memory, Roughness, and Information Persistence in Financial Markets: A Structural Approach to Volatility Forecasting

Battle Count: 117

Highly relevant for quantitative trading applications. The paper demonstrates that persistence-augmented volatility forecasts improve out-of-sample performance by 4.6-8.2% MSE reduction relative to HAR benchmarks, with gains reaching 17-19% during stress regimes. The volatility-managed portfolio application shows Sharpe ratio improvement from 0.65 to 1.36 during COVID. The findings directly inform risk management, position sizing, volatility trading across maturities, and regime-aware strategy design. The persistence-by-stress interaction framework provides actionable signals for distinguishing temporary volatility spikes from prolonged uncertainty regimes. However, the implementation requires Bloomberg data access and careful rolling estimation infrastructure.

Moderate-to-high complexity. The framework requires: (1) rolling semi-parametric long-memory estimation (GPH and local-Whittle) with 750-day windows updated weekly, (2) rolling Hurst exponent estimation from Parkinson realized variance, (3) construction of a multi-dimensional feature vector including own-stock, cross-sectional, sectoral, and interaction persistence features, (4) walk-forward out-of-sample forecasting with expanding windows, (5) panel-aware statistical inference with HLN-corrected Diebold-Mariano tests, (6) comparison across 8+ model specifications, and (7) volatility-managed portfolio construction. The econometric components (fractional integration estimation, spectral methods) require specialized knowledge. The layered forecasting ladder and robustness battery add significant implementation burden.

The paper provides detailed descriptions of the econometric methodology, rolling estimation design, forecasting protocol, and data construction. However, no code repository is mentioned. Data is sourced from Bloomberg Terminal (accessed April 21, 2026), which is proprietary. The sample period (Nov 2001 - Apr 2026), 115-stock panel, 750-day rolling windows, and all model specifications are clearly documented. The high-frequency subsample (Oct 2025 - Apr 2026) is also proprietary. Reproduction would require Bloomberg access and careful implementation of the layered forecasting ladder.