One way to increase the ROI in Custom Audiences is to make use of predictive models in order to find those customers that are more likely to convert due to the ad than other customers from your database.
You will usually be aware of your current best customers. Those that have just purchased. Oftentimes these are good customers to include in your next campaign (sometimes with an offset as some products do not need to be replaced too often). But a more precise way is to identify those that are going to buy soon. In order to find those you need to make use of predictive models (there are many terms for this, e.g. AI has become a term that is often used in this context as well).
The value of a predictive model is often called lift. This is the extra ROI that you are able to generate when making use of these models instead of other analytical means like ABC analysis, RFM criteria or whatever your weapon of choice is.
The lift that a predictive models yields is not something that is certain. A predictive model will yield a certain ROI in that situation. The less advanced your alternative approach the bigger the lift. Some of our customers just reported 75% increase in ROI making use of Facebook custom audience lookalikes (incase you want to find out how that works, have a look here: Björn Goerke’s answer to Are you seeing the same performance on the lookalike audience vs the custom audience? So far I am, which leads me to question what magical algorithm is Facebook using?). The range differs widely. In most cases the potential is very attractive.