Machine Learning
Deep Hedging Framework
Neural hedging policies trained against convex risk measures over simulated markets with transaction costs, accelerated by fused CUDA path kernels, plus a deep BSDE pricer.
Trains feedforward, GRU, no-trade-band, and permutation-equivariant deep-set policies by stochastic gradient descent on Rockafellar-Uryasev CVaR, entropic, and spectral risk over simulated GBM, Heston, Merton-jump, and Dupire local-volatility markets under proportional, bid-ask, and power-law impact costs.
Generates paths on the fly with fused CUDA Philox kernels, one thread per path with state held in registers, sustaining about six billion GBM and three and a half billion Heston paths per second, with whole-episode CUDA-graph capture and a noise-regenerative backward that cuts peak training memory by 12.7 times.
At 40 basis points of cost the learned policy lowers 95 percent expected shortfall by 21 percent against the Black-Scholes delta hedge, and a deep BSDE solver with tilted-Gaussian tail importance sampling prices a 50-dimensional basket call within 1.3 percent of its closed form.