A1221
Title: Neural tangent kernel in implied volatility forecasting: A nonlinear functional autoregression approach
Authors: Maria Grith - Erasmus University Rotterdam (Netherlands) [presenting]
Ying Chen - National University of Singapore (Singapore)
Hannah Lan Huong Lai - National University of Singapore (Singapore)
Abstract: Implied volatility (IV) plays a crucial role as a 'visible' measure of volatility in investment, hedging, and risk management, underscoring the importance of accurate IV forecasting. IV exhibits temporal and spatial dependencies due to its reliance on moneyness and maturity, with nonlinear and complex dependence forms in real data. A flexible econometrics modelling framework, the Nonlinear Functional Autoregression (NFAR), is proposed to effectively capture both linear and nonlinear relationships in implied volatility surfaces (IVS) series by leveraging neural networks. The estimation procedure incorporates the Neural Tangent Kernel (NTK) parameterization, enabling the capture of interdependencies among low-dimensional components derived through projections on the covariance operator of the curve time series. The link between NTK and kernel regression is established through rigorous derivation, highlighting NTK's role as a modern nonparametric statistical model. Empirical experiments forecasting the IVS of the S\&P 500 index from January 2009 to December 2021 demonstrate an average improvement of 16\% to 64\% in forecast accuracy for 5 to 20-day-ahead predictions compared to classic alternative and nonparametric variants. The enhanced predictability of IVS significantly impacts trading strategies, resulting in a relative increase in Sharpe Ratio for short straddles between 32\% to 415\% as the investment horizon increases, thus benefiting option market investors.