Traditionally, Lead Scoring is done using a rule based approach. For example, leads from North America get 10 out of 10 points for having a “good geo”, leads from larger companies get 20 out of 20 points for having a “good company size”.
The flaw with such an approach is that the outcome is linear - meaning, the outcome will not be able to model for pockets of high conversion (eg. Government leads with smaller org size have higher likelihood of conversion). Such, pockets of high conversion can’t be modeled by a rule based approach which assumes a straight line and fits everything around it.
Our approach to delivering a highly accurate and robust lead scoring was to ditch the traditional rule based approach and apply machine learning to predict conversion likelihood. We leveraged the historical lead conversion data along with enrichment using Clearbit Enrich. This enriched dataset now had profile as well as firmographic information for each lead. This data functioned as input for our machine learning models.
In order to make the prediction robust (i.e. having a good prediction even when some elements are absent) we developed numerous smaller machine learning models which are individually good at predicting certain cases only. But, when put together in a stacked ensemble , they perform better than any one model and are able to handle many different exceptions. This robust setup allowed us to achieve 75% accuracy in predicting whether a lead is a Sales Qualified Lead.