LGD is hard to predict and doesn't correlate with many underwriting variables, as if it did, one could incorporate that information into the pricing. Pretty much whatever is known up front is used for decisioning and pricing, and exogenous events like job loss, death, and family issues determine the severity. Therefore, P[D] is the most common metric. Blending the probability of default and loss given default help you construct an annualized expectation of loss, but it will correlate very highly to P[D]. I simply used the average LGD in my P2P-Picks work. Therefore, Rob's charts aren't flawed in the way contemplated by nonattender.