I’m going to begin with a distinct and concise clarification of what I mean when I talk about probability versus propensity. Hopefully, these definitions will support my emphasis on this distinction being enforced as an important task as data-driven marketers. We need to be specific when utilising terminology, especially in forecasting.
Rather than using pure dictionary definitions, I feel it is important to describe meaning within the attribution context. Therefore, with any luck our descriptions below enable the comparison to be clearer when examining the difference and relationship between the two when exploring data.
In the context of a future conversion, these two terms have very distinct differences.
The probability of me buying a car, versus the probability of me buying a car in the next two weeks, are in many cases going to be significantly different. At CUBED a variance in propensity is how the algorithm calculates and subsequently forecasts the value of every single interaction – so we needed to define a timeframe within which probability could sit and thus becomes a propensity calculation.
The change of propensity that a singular impression, visit, email, micro-conversion etc., has, is built up of a number of factors that sit beside the CUBED attribution algorithm. Due to all interactions between the consumer and the brand being factored into the calculation (for both successful and unsuccessful consumer journeys) the algorithm learns what success looks like. Just as importantly, it learns also where users are yet to convert.
The fact that CUBED utilises machine learning means that the algorithm doesn’t just have static attributes. This means that when forecasting if a user will convert, it learns based on its success in previous forecasts, ultimately getting smarter!
When building CUBED, early on we decided that the period of time to be used to define propensity change will be two weeks. We wanted to ensure that if we are saying a person’s likelihood of converting is impacted by activity, a marketer could take action within a two week period and see a potential instant impact.
Key to this is the concept of decay and that propensity is not just a static figure. Over time with every day that passes, the propensity for someone to convert will decrease if we don’t see any further interaction with the brand. The rate of this decay is algorithmically calculated based on historical consumer behaviour.
This entire logic is really important in advanced audience segmentation – the propensity of an individual to convert can form a factor of segmentation. This allows brands to not just have static CPA’s based on channel logic, but instead take many elements into consideration.
We are constantly iterating and improving our propensity calculations as part of our ongoing machine learning solution.