Andersen–Broadie primal–dual bounds
Computes lower and upper estimates for early-exercise option prices by simulation.
How it works
The primal step values a feasible early-exercise policy, producing a lower bound in expectation. The dual step uses a martingale adjustment to construct an upper bound. Together they quantify uncertainty about the optimal exercise value without requiring an exact solution of the original stopping problem.
Worked example
Suppose independent simulation gives an exercise-policy value of 4.92 and a dual estimate of 5.10. Their difference is 0.18 in price units before taking simulation error into account. Improving the policy raises the lower bound when it captures more worthwhile exercise. Improving the martingale approximation can tighten the upper bound.
Validation and limitations
Use separate training and evaluation samples and report sampling error alongside the gap. Check the conditional expectations used to construct the martingale. A single estimate below or above the other on a finite sample can reflect noise. Compare least-squares Monte Carlo.