Academic article on Prosper published - Dynamic Learning and Selection

Seth Freedman and Ginger Zhe Jin from the University of Maryland have published an academic article on Prosper called Dynamic Learning and Selection: the Early Years of

Freedman considers the article a preliminary draft. Since first publishing the article they have already made a couple minor changes to the abstract and the introduction with an effort to to stress that the rate of return they calculate is the expected rate of return as opposed to the realized rate of return.

Abstract - Dynamic Learning and Selection: the Early Years of

This paper studies a new business model on the Internet., the first peer-to-
peer lending website in the US, matches individual lenders and borrowers for unsecured consumer loans. Since its inception in February 2006, Prosper has attracted over 500,000 members and has originated loans of over $100 millions. How do individual borrowers and lenders behave in this new marketplace? On what ground can Prosper survive the competition with traditional banks? Is Prosper positioned to replace the traditional lending market or serve a market that the banks have missed? We attempt to address these questions using the full transaction history since the birth of

Compared to the traditional market, Prosper decreases operation and search costs, but may face information costs associated with anonymous online interaction. To overcome the information problem, Prosper has implemented a number of policies over the past two years. Our primary goal is to document the dynamics on both sides of the Prosper market, while accounting for changing Prosper policies and the macro environment.

We have several findings. On the borrower side, we find that the overall observable risks have worsened over time for the pool of listings. While part of this change is driven by the macro environment, Prosper policies are effective in countering this trend, especially for the sub-prime credit grades. On the lender side, we find that the risk perception that lenders apply to key borrower attributes is by and large consistent with how these attributes correlate with the loan’s ex-post performance, but there are significant exceptions. Over time, lenders exhibit significant selection and learning. This learning includes better understanding of the risk of low credit grades and group member loans.

Overall, we conclude that Prosper is evolving from a comprehensive market toward a market that primarily serves borrowers who have access to traditional credit. Using the estimated relationship between loan attributes and loan performance, we estimate the rate of return that a fully rational lender could expect if he can perfectly predict the probabilistic distribution of loan performance conditional on these attributes. If Prosper loans continue to perform according to what we have predicted from their existing performance, this average annual expected return on the funded loans will be approximately 6%. This return has varied by time and has been increasing as the composition of the funded listings shifts toward better credit risks.

Please see details about the empirical methods and assumptions in the full paper. Freedman said they will update the article with more current information later in the summer. The current data set used runs through the end of 2007.