Non-Spanning Identification of Scheduled Event Risk in Option Pricing

Battle Count: 74

Highly relevant for options market makers, volatility traders, and event-driven strategies. The paper provides a rigorous framework for pricing scheduled macro-announcement risk in short-dated index options, with clear evidence that gains concentrate in event-volatility combinations (straddles, strangles) rather than directional risk reversals. The contaminated-surface stress test is practically important for anyone calibrating surfaces around event dates. The scale-shape attribution informs how much target-event data is needed vs. what can be transferred across events. However, the paper is primarily methodological/identification-focused rather than directly producing trading signals.

Moderate to high complexity. The parametric components (polynomial surface fit, Gaussian/mixture jump calibration) are straightforward with standard optimization (L-BFGS-B). The neural MDN benchmark adds complexity with leave-one-event-out training, conditioning vector construction, and semi-amortized calibration. The event-level bootstrap inference, calendar isolation filters, quote-count floors, and stratified held-out sampling require careful implementation. The main challenge is data preprocessing (OptionMetrics access, PM-settlement filtering, OTM+ATM selection) and ensuring the information-set separation is correctly enforced.

The paper provides detailed implementation specifications (basis functions, hyperparameters, loss functions, calibration procedures, event-level splits). However, the primary data source (OptionMetrics) is proprietary and requires a subscription. The event calendar construction, quote filtering, and held-out sampling procedures are well-documented. Neural MDN architecture details (two-layer width-32 GELU MLP, Adam optimizer, learning rate 1e-3) are specified. No code repository is mentioned.