For processing and analysis, you're centralizing the compute of complex analysis and storing the result so downstream decision support systems can use the result as a criterion in their analysis - and not have to distribute, and maintain, that logic throughout the set of applications. A contrived example: is_valued_customer. This is a simple boolean, but its computation can be involved and you wouldn't want to have to replicate and maintain this logic throughout all the applications. But at the time, it likely has no business being in the OLTP store.
You might persist that value as an optimisation, but if you make it your source of truth, and discard your inputs, you better make sure you never ever ever ever have a bug in deriveValuedCustomer() or else you have lost data permanently
Good point - you wouldn't want to discard your inputs. You're going to need them should you ever redefine deriveValuedCustomer() - which is likely for a system that will be in production for 10-20 years or more.