Option-Implied Signals and Crash Risk: Predictability and Machine-Learning Evidence from U.S. Equity Options, 2015–2026

Battle Count: 63

Highly relevant to quantitative trading. The paper directly addresses cross-sectional return predictability from option-implied signals, crash-risk forecasting, and regime-conditional factor performance. Key findings (smirk decay, IV spread and risk-neutral skewness persistence, regime-dependent dominant features, XGBoost outperforming linear models only in AI/mega-cap regime) have immediate implications for options-based alpha strategies, tail-risk hedging, and risk management. The finding that canonical hand-engineered signals are subsumed by primitive IV-surface features in ML models suggests that quant desks should move beyond smirk/IV-spread factors to raw surface features. The regime-conditional nature of predictability is critical for strategy deployment timing.

Moderate-to-high complexity. Requires access to OPRA end-of-day option data (privately licensed), construction of a 12M+ row firm-day panel with multiple IV-surface aggregations, two-way fixed-effects panel regressions with double-clustered SEs, rolling expanding-window ML evaluation with XGBoost/Elastic Net/OLS, permutation feature importance computation, and binary crash classification with class up-weighting. The pipeline runs ~6 hours on a 10-core/16GB workstation. Key challenges: memory management for panel regressions, proper regime splitting, feature engineering from raw contract-level data, and out-of-sample protocol implementation. No public code repository provided.

The paper provides a four-stage pipeline (preprocessing, descriptives, panel regressions, ML benchmark) that runs in ~6 hours on a 10-core desktop with 16 GB RAM. Post-filter firm-day parquets and full processing pipeline are available on request. However, the raw OPRA data is privately licensed and not publicly available. No GitHub repository is linked. The risk-neutral skewness uses a proxy rather than the full BKM estimator. Subsampled regressions (1.5M rows) are used due to memory constraints. SHAP attribution is deferred due to environment issues. No CRSP/Compustat merge is performed.