Risk Scoring using Fuzzy Bandits with Knapsacks

Bandits with Knapsacks (BwK) is a MultiArmed Bandit (MAB) problem under supply/budget constraints. Risk scoring is a typical limited-resource problem and as such can be modeled as a BwK problem. This paper tries to solve the triple problem in risk scoring – accuracy, fairness, and auditability by proposing FuzzyBwK application in risk scoring. Theoretical assessment of FuzzyBwK is made to establish whether the regret function would perform better than StochasticBwK and AdversarialBwK functions. An empirical experiment is then set with secondary data (Australian and German credit data) and primary data (Kakamega insurance data) to determine whether the algorithm proposed would be fit for the proposed problem. The results show that the proposed algorithm has optimal regret function and from the empirical test, the algorithm satisfies the test of accuracy, fairness and auditability.