Are Your Propensity Models Compromised By Outlier Leakage?

Many banks have come to realize that relying on code-heavy basic propensity models and signal targeting will generate single engagements while misdirecting and mistiming the marketing message. 

The age of personalization and digital engagement has meant that marketers operating within financial services must satisfy an increasing number of unique and entwined variables of data across customer data, business objectives and market conditions. Purveyors of basic rule-based propensity models argue that these systems have served marketers well but increased data insights and volatility across this triumvirate highlights the models’ limitations, in particular the need to re-code on a quarterly basis when banks really need to be able to recognize change instantly. Not only are outliers in customer data increasingly instable, basic propensity models serve up next best actions without providing likely outcomes of different actions and the damage done by message fatigue. 

Basic propensity models search for data outliers as signals that an action should be served against, whether that’s a year-end bonus showing up in an account, decreased spending or a customer clicking on a generic marketing message. Within the personalized marketing space propensity models are based on pre-built rules assigned to each user’s historical behavior and expectations. For instance, an account with a year-long balance of $100 to $1,000 jumping up to $40,000 is a significant signal. Perhaps the model will tell the bank that it should push mortgage offers. Conversely, if the account suddenly drops to zero the model might instruct the bank to push personal financial management (PFM) tools or financial health literature.  

An immediate question raised with this approach is: what then? Typical basic propensity models require full retraining when customer behavior, market context or business requirements change, and generally continue to push the same message or offer to a customer on the basis of that initial signal until hard coding and training is complete. 

Banks are now turning to personalization facilities that make use of contextual multi-arm bandit models, that automatically learn and deepen relationships outside of rigid rules. These systems build a contextual understanding of each individual and the broader marketing environment, assessing much more granularly what drives behavior in real-time, and weighing up likely responses to different creatives much further along in the relationship than the first signal. This contextual understanding of both individuals and the environments they operate in is what allows AI driven solutions to move away from optimizing at the full population level and towards true personalization.  

Unlike strict propensity models that focus on actions that weren’t happening at the time of model training, optimization engines that utilize contextual multi-arm bandit approaches learn constantly about signals and adjust expectations without the need to retrain.  

Next generation marketing personalization optimization based on true AI does not rely on a static view of the world, instead it works off the basis that humans and the environment around them change. By learning constantly about how those human needs and wants are changing within the wider world marketers can truly personalize and deliver on a continual basis.