Contests for Predictive Algorithms: Ensembling, Interdependency, and Optimal Rewards Design

With the increased availability of data and widespread use of AI/ML technologies, firms are increasingly using crowdsourcing contests to acquire predictive algorithms for business use. A unique feature of these contests is that the designer can create an ensemble of submitted algorithms and improve upon the predictive accuracy (or quality) of individual algorithms. Given this departure from conventional contests, we ask how the optimal rewarding schemes should be in the presence of ensembling and interdependent algorithms. To answer this question, we develop a stylistic model of a contest for a predictive algorithm with two participants. As opposed to the single best contribution typically sought out in conventional contests and the winner-takes-all reward scheme, we find that both the winner and the runner-up could receive rewards in contests for algorithms, depending on whether the ensembled algorithms substitute or complement each other as determined by the algorithm interdependency. We show that the degree of substitution or complementarity can fundamentally alter the structure of the optimal reward scheme. We highlight our results using a real-world algorithm contest for predicting breast cancer using digital mammograms, we demonstrate the practical applicability of our framework.