Learning the Exact SABR Model

Battle Count: 100

Highly relevant for quantitative trading and risk management in interest rate derivatives. The SABR model is the industry standard for calibrating interest rate volatility surfaces (Caps, Floors, Swaptions). This paper provides a practical, fast, and accurate alternative to the Hagan et al. approximation that eliminates arbitrage in extreme market conditions. The sub-second calibration speed enables real-time pricing, risk measurement (VaR, CVaR, Expected Exposure), stress testing, and what-if analyses for large portfolios. Directly applicable to trading desks, risk management departments, and quantitative research teams at banks and asset managers.

The DNN architecture itself is relatively simple (feed-forward, 5 hidden layers, 64 nodes per layer, standard scikit-learn preprocessing). However, the full pipeline requires: (1) High-performance computing for MC dataset generation (~25,000-30,000 CPU hours), (2) Careful parameter range calibration via iterative procedure, (3) Three separate DNNs for different maturity buckets, (4) Online calibration using L-BFGS-B with 300 random initializations, (5) Volatility stripping from market Cap/Floor prices, (6) Variance interpolation across maturities. The offline training is computationally intensive but one-time; the online calibration is trivially fast.

The paper provides extensive details on DNN architecture (5 hidden layers, 64 nodes, ELU activation), parameter ranges (Table 1), MC simulation scheme (log-Euler discretization), dataset generation procedure, and benchmark Caplet/Floorlet prices (Tables 9-10). However, the 200M+ point dataset requires ~25,000-30,000 CPU hours on 256 CPUs, making full reproduction computationally expensive. The MC discretization scheme is explicitly stated and the framework is noted to be independent of the specific scheme.