pfnet-research/pfhedge

Purpose and Public Surface

PFHedge is a PyTorch library for Monte Carlo deep hedging of financial derivatives under market frictions. Users compose instruments, a hedging model, input features, and a risk criterion; then train and price through Hedger. Package exports are rooted at pfhedge/__init__.py, with main public domains in pfhedge/instruments/__init__.py and pfhedge/nn/__init__.py. Runnable usage entry points are the scripts/notebook in examples/, especially examples/example_readme.py.

Instrument Domain Model

The instrument layer separates tradable simulated underliers (“primary” instruments) from contingent claims (“derivatives”). Shared abstractions establish the contract and likely device/dtype/state behavior at pfhedge/instruments/base.py, pfhedge/instruments/primary/base.py, and pfhedge/instruments/derivative/base.py. Derivative implementations bind one or more underliers and define payoff/contract-specific attributes, e.g. pfhedge/instruments/derivative/european.py, pfhedge/instruments/derivative/lookback.py, and pfhedge/instruments/derivative/variance_swap.py.

Market Simulation Pipeline

Primary instruments model paths and market costs; concrete models include geometric Brownian, stochastic-volatility, rates, and jump processes in pfhedge/instruments/primary/brownian.py, pfhedge/instruments/primary/heston.py, and pfhedge/instruments/primary/rough_bergomi.py. The underlying stochastic-process implementations are factored into pfhedge/stochastic/engine.py, pfhedge/stochastic/brownian.py, and pfhedge/stochastic/heston.py. This boundary lets instruments expose finance-facing objects while stochastic modules generate tensor-valued paths.

Feature Assembly

Feature objects turn derivative state, underlier paths, and prior hedge decisions into model inputs. The extensible feature contract and lookup mechanism live in pfhedge/features/_base.py and pfhedge/features/_getter.py; composition/container handling is in pfhedge/features/container.py. Features are central to the recurrent control loop because prev_hedge makes each time-step decision depend on the prior position.

Hedging and Optimization Orchestration

pfhedge/nn/modules/hedger.py is the main application controller and PyTorch Module: it simulates/evaluates a derivative, constructs sequential feature tensors, invokes the strategy, derives portfolio P&L including transaction costs, applies a risk criterion, and supports fitting/pricing. Strategy networks and controls include pfhedge/nn/modules/mlp.py, pfhedge/nn/modules/clamp.py, and pfhedge/nn/modules/naked.py. Training delegates parameter updates to standard PyTorch optimizers; GPU and dtype movement follow PyTorch module/tensor semantics.

Models, Baselines, and Risk Objectives

Risk measures used as training objectives reside in pfhedge/nn/modules/loss.py, including entropic risk and expected shortfall. Analytic/model-based benchmark strategies are organized under pfhedge/nn/modules/bs/, with generic Black-Scholes machinery in pfhedge/nn/modules/bs/_base.py and derivative-specific formulas such as pfhedge/nn/modules/bs/european.py. Transaction-cost-aware Whalley–Wilmott logic is in pfhedge/nn/modules/ww.py; SVI-related modeling is in pfhedge/nn/modules/svi.py.

Differentiation, Runtime, and Delivery

pfhedge/autogreek.py differentiates user pricing functions through PyTorch autograd to calculate Greeks, crossing the pricing/autodiff boundary without separate finite-difference infrastructure. The material runtime dependency is PyTorch; no server, database, or external market-data service is represented. Packaging is defined by pyproject.toml; GitHub Actions provides CI, docs, examples, release, and version-bump workflows under .github/workflows/. Documentation is built from docs/source/.