ExposureSimulation

The Phase-2 exposure engine: a Monte-Carlo simulation of a netting set's mark-to-market on a time grid, recorded once onto the adjoint tape.

Each path evolves a one-factor Hull-White short rate and a lognormal FX spot, values every trade at every grid date from analytic P(t, T) bonds, applies the CSA (variation margin with a threshold and a margin period of risk), and accumulates the pathwise CVA integrand

CVA = LGD * sum_k  max(V(t_k) - C(t_k), 0) * D(t_k) * ( S(t_{k-1}) - S(t_k) )

against a piecewise-flat counterparty survival curve whose forward hazards are tape inputs. One .greeks() run therefore returns the CVA and its full risk vector — IR delta and rate vega, counterparty CS01 by tenor bucket, recovery and FX — from a single reverse sweep. The bump-and-revalue alternative re-runs cvaOnly once per shocked risk factor.

View source ↗Full module reference on javadoc.io ↗

Constructors

ExposureSimulation(NettingSet nettingSet, int steps)

Methods

ExposureSimulation on(String engine)

String engine()

ExposureSimulation fp64(boolean doublePrecision)

Force the replay precision. Left unset, the precision follows the engine: the Vulkan engine is single-precision, every other engine runs fp64. The money-scaled integrand and the expm1-form marginal-default probability keep the fp32 path within Monte-Carlo error of the fp64 one.

int steps()

double horizonYears()

double stepYears()

NettingSet nettingSet()

CvaResult run(CvaMarket market, long paths, long seed)

CVA and the full CVA gradient from one adjoint sweep on engine, plus the expected-exposure profile. The CVA value and gradient come from a single greeks() kernel over the scalar CVA — the timed sweep. The per-date profile is a separate priceOnly() kernel with many named outputs; that runs on engine where the backend supports it and falls back to cpu-jit otherwise (it is a picture, not a headline number).

double cvaOnly(CvaMarket market, long paths, long seed)

Just the CVA number — the kernel a prescribed-bump sensitivity re-runs.