Joshi and Kwong in this paper propose a strategy for calculating XVAs in American Monte-Carlo that avoids using the regression to calculate directly the exposure. Instead they use it only to obtain the exposure sign. They report that this gives a good improvement in accuracy.
Has anyone tried this and compared it to the standard approach over a wide range of products in a portfolio, to assess its practicality? In particular I’m wondering about its feasibility as being done at portfolio level, it may lead to regressions in a large number of dimensions.