When the Fed Speaks: Dynamics and Forecasts of the Volatility Surface

Battle Count: 69

Highly relevant for quantitative trading. The paper demonstrates that ML models can forecast the entire IV surface and that incorporating FOMC announcement dates as exogenous features significantly improves predictions on announcement days. The pre-announcement IV premium (up to 50% for 0DTE options) is a well-documented trading opportunity. The impulse response analysis shows the model captures the elevated IV pattern. However, the paper acknowledges limitations in statistical significance globally and the need for trading environment validation. The findings are directly applicable to options trading strategies around scheduled macro events, volatility surface arbitrage, and risk management during policy announcements.

Moderate to high complexity. Requires: (1) CBOE options data processing with no-arbitrage filtering, (2) IV surface construction via three methods (linear, SVI, AHBS) on a fixed 41×21 grid, (3) 2D ConvLSTM architecture with two branches (surface + macro input), (4) autoregressive multi-horizon forecasting (h=1,2,5,10), (5) Diebold-Mariano and Mann-Whitney U statistical tests, (6) careful handling of 0DTE options and wing extrapolation. The SVI fitting per maturity slice and the dual-branch network architecture add implementation complexity. Training is relatively lightweight given the small model size.

Code is publicly available on GitHub. However, the CBOE options data is subscription-based and cannot be shared directly. The methodology is well-documented with specific hyperparameters (lookback=5, learning rate=0.00005, kernel=3x3, filters=32, batch=32, dropout=0.1). Results are reported across 6 random seeds with standard deviations. The FOMC calendar is sourced from the Fed website. Risk-free rate from FRED. Training/validation/test split is clearly defined (2016-2023/2024/2025).