How Does a Firm Adapt in a Changing World? The Case of Prosper Marketplace

We propose a generalized revealed preference approach to infer how a firm adapts to a changing environment and provide a step-by-step guide to explain how to implement it in general. To illustrate this new approach, we apply it to Prosper, which is a peer-to-peer lending platform. We develop a structural model, in which Prosper uses an adaptive learning algorithm to continuously update its predictive models about borrowers’ and lenders’ behavior as more data become available and uses these updated models to help assign loan ratings over time. To infer which adaptive learning algorithm Prosper may adopt, we consider a set of algorithms motivated by the machine learning literature. For each algorithm, we use observed Prosper loan-rating decisions to estimate the structural parameters of Prosper’s objective function. By comparing the goodness-of-fit of these algorithm-specific models, we find that Prosper most likely uses an ensemble algorithm, which selects past observations based on their economic conditions. We conduct counterfactual experiments to shed light on: (i) How does an exclusive focus on either accurately reporting loan risk or expected current revenue influence Prosper’s decision making? (ii) What is the value of adaptive learning for Prosper? (iii) Is there any potential for Prosper to improve its current adaptive learning algorithm?