« A practical guide to model back-testing | Main | Trying to Tame the Unknowable »


John Ivy

I am very interested in your comments for the 7th way "Simulation". I am assuming that this method would be taking an AL model from the past (lets say last month), and feeding it current rates and volumes (maybe even current spreads, prepayments, etc.) then comparing the calculated income in month 1 with what was actually realized. Is this a correct assumption? If so - how far back in time would you go?
I have read your blog a lot over the years...thanks so much for your time and effort.

Brad Olson

...via email...

>> I just spent that last couple of hours reading your blog posts. I had previously found your
>> discussion on ways to back-test models as very helpful. (Especially since the regulators are
>> asking that question of us.) In going through your blog, I was able to pick up the first
>> five ways to back-test a model. Have you written the last two items?
>> I enjoy your perspective on interest rate risk. Thanks, Gary

Brad Olson

Wow, two comments in one week (on the same post)! Maybe it's the start of a trend. Blogging "experts" say that one sure way to generate comments is to leave a list post unfinished. I guess there's something too that.

John – yes, you have the right idea. In this context Simulation means “modeling history” and comparing the results to what actually happened. Typically you’re looking at margin and net income. How far back you go largely depends on how easily you can obtain the data. This process often requires you to query (or pull) historical data from core systems. There’s usually a limit to how far back in the past you can go.

I’ll blow the cobwebs out and put the finishing touches on the post and publish it soon.

Thanks for reading!

The comments to this entry are closed.