I used to build structured credit risk models at JPMorgan and other banks. A few observations: - Excel models are usually built on a pre-existing template, pr

I used to build structured credit risk models at JPMorgan and other banks. A few observations: - Excel models are usually built on a pre-existing template, probably created by someone else. Any errors/hidden assumptions carries through. This is he biggest source of errors. - Any large model usually has a tab devoted to reference tables and assumptions based on external sources, a la dependency injection. Usually you have to use static numbers since you need to email the model around (you don't want your boss seeing a bunch of #NAs due to missing links. Keeping this data updated is another huge source of errors. - A sanity check is your best defense. I'm positive that every CDO model I've ever built has contained at least one error, just like every large software application has at least one bug. The difference is that models have fewer output parameters, which allows you spot incongruities more easily. - People who use big Excel models are not software engineers and will never be. In my opinion, solutions that implement concepts like testing, diffs, commits, etc are a waste of time, because it creates extra work and learning curve.