I have seen it written that in Monte Carlo simulation, a prepayment model is used, not PSA. Can someone please explain this?
I spent six years at PIMCO analyzing mortgage-backed securities; for the most part, that’s developing prepayment models.
You build an interest rate tree and sample paths from that tree via Monte Carlo simulation (a 360-month binomial interest rate tree has just over 2.3485 × 10^108 different paths: a few more than is practical to use in their entirety). For each node in each path you choose you use your prepayment model to estimate the amount of the mortgage pool that would be prepaid. Then you aggregate all of the discounted cash flows to determine the price, OAS, effective duration, and so on.
In saying that you don’t use PSA, they mean that you don’t, ex ante, decide what percentage of PSA you’ll receive in prepayments each month. Ex post, for each path you can certainly calculate each month’s PSA of course, but you don’t use PSA to determine the prepayments. (This is like saying that when cutting firewood you don’t set about to make each piece a specific length, but you can measure each piece after you cut it, if you want to.)
S2000, just curious - how good was your estimate? Did you ever go back and compare actual data vs what you had predicted?
I have to do a bunch of software estimation and I was wondering what accuracy I should be satisfied with. The commercial models, God bless their soul, have so many dials and knobs that I can get any answer I want out of them. And there is no established theory backed by experiments. Just academic speculation building towers in the air on three data points.
As I recall, they were pretty good, but I left PIMCO 13 years ago, so I cannot be more specific than that: too few neurons left.
thanks
You’re welcome.