AAD
Adjoint algorithmic differentiation for many sensitivities at near-constant reverse cost.
Define the numerical question and error budget.
Record the pricing computation, then propagate sensitivities backwards through it.
Adjoint algorithmic differentiation for many sensitivities at near-constant reverse cost.
Specify the estimator or discretization.
Notation and units
Decimal rates and volatilities, year-fraction time and continuous compounding unless stated otherwise.
Expose convergence and stability conditions.
The numerical target, discretization and stopping rule must be fixed before comparing outputs.
Convergence is assessed against bias, variance or residual tolerances rather than visual smoothness.
Finite precision, boundary treatment and input conditioning can dominate model error.
“An unstated convention is a future reconciliation break.”— THEQUANTBATEMAN
Connect controls to an observable output.
Industry-standard technique for large-scale Greeks where the implementation supports it.
Monitor bias, variance, and failure modes.
Report the Frontier number with its convergence evidence. A stable-looking output can still carry discretization bias or an ill-conditioned input.
Ask Bateman about this topic →