Onaopepo Adekunle, Chang Sun, Arno Riedl and Michel Dumontier, Maastricht University, Netherlands

Recommender systems are ubiquitous in various domains ranging from e-commerce to financial services to deliver personalized products and services at scale. This work designs and develops a proof of concept for personalized pension recommendations. PenRec, a novel pension savings recommender system, takes the combination of individuals’ demographic data and risk preferences data as input. PenRec combines collaborative filtering and deep learning methods to provide individual savings recommendations based on individuals with similar characteristics as pension savings advice. To estimate the risk preferences of individuals, we implemented a survey, including incentivized economic experiments, among the Dutch working population (N = 4, 282); upon which we employed supervised learning techniques to estimate the risk preferences of the general population of the Netherlands based on register data from Statistics Netherlands. We present an analysis of the comparison between traditional collaborative filtering and deep learning for pension savings recommendations. Additionally, we investigate whether the inclusion of a behavioral measure as a feature, estimated risk preference in particular together with demographic data is important for pension savings recommendation. We find that traditional collaborative filtering tends to recommend average values of individual savings for specified categories of people such as self-employed, income source, and province with no deviations for such categories; while the deep learning-based recommender system can capture varying categories of people in the group of similar individuals and hence provides recommendations that track the real savings more closely relative to the traditional collaborative filtering method. In addition, we observe the inclusion of risk preference variable as a feature with which to compute individual similarity impacts the recommended savings marginally and encourages the addition of behavioral measures for a robust recommendation.

Recommender System, Pensions, Collaborative filtering, Deep Learning