Next best product
The Next Best Product (NBP) forecasts are designed to support insurance professionals in proactively identifying additional insurance needs among their clients.
The model uses machine learning to predict which insurance products a customer is expected to purchase or need in the next 12 months.
These insights enable advisors to approach customers at the right time with a personalized and relevant offer, increasing both customer satisfaction and commercial effectiveness.
Definition of need
The model predicts the probability that a customer will voluntarily take out a particular product. This perceived closing probability counts as a lower bound of the actual need: after all, a customer may also have a need for a product without closing it himself, for example, due to a lack of knowledge, timing or awareness.
For this reason, the raw model output is scaled up using a market average, determined based on insights from insurance professionals and business intelligence. This ensures that the output is representative of the actual need in the market.
Data usage
To make a reliable prediction, insurance products are first grouped by content and function. This makes it possible to build more specific models than at the main industry level. More than 40 characteristics are used to train the model, drawn from the following data categories:
Policy Information
Information on current and historical policies by customer.
Relationship characteristics
Consider age, family composition, place of residence and customer duration.
External data
Such as demographic data or geographic risk profiles.
Model training and validation
The model is trained with year-level historical customer data, with each data point representing whether a customer purchased a particular product in a specific year. From this, the model learns patterns and characteristics that are indicative of subsequent purchases.
A separate predictive model is trained for each group of insurance products. This allows the models for each product group to be optimized and validated separately.
The validation is based on the most recent full calendar year, so that the performance of the model matches as realistically as possible the situation in which it is actually applied.