Generally, predictive marketing is based on opportunities and accounts more so than leads and activities. We used Everstring, and they built a model that was largely built on our closed-won deal data, late stage opportunity data (finalist stage, legal review) coupled with their proprietary database of external intent signals.
They recommended we provide 200 opportunities from the previous two years on which to base our model. There is an assumption that the 200 opportunities represented a significant diversity of accounts. We didn't quite have a one-to-one opportunity-to-account ratio in our mix, but our model was based on 200+ opportunities representing more than 150 different accounts.
The model they built for us was ~85% accurate in the trial period, and 92% accurate in the first full year. In other words, 85% of our closed won business (based on revenue) during the trial period was from an account rated A or B, and 92% of our revenue in the first full year was from an A or B rated account.
There's another use case where they can build a model using a seed list of accounts (they recommend at least 100) external to your current database. Say you're launching a new product--you don't have any closed/won data to use as the base for your model, but you've identified a target list of prospects--maybe a customer list of a competing product. .
.png)