The ability to accurately forecast the recoveries from defaulted or charged-off loan accounts is important for the profitability of credit providers, calculating regulatory capital requirements, and determining the value of a defaulted loan book. The recoveries are typically modeled using traditional statistical techniques and, recently, machine learning techniques. In this chapter, we utilize a Bayesian framework to analyze and detect whether there are recoveries on defaulted loans from a peer-to-peer lending platform. In this framework, we train a Bayesian logistic regression model with an automatic relevance determination (BL-ARD) prior distribution, and using the Metropolis-Adjusted Langevin Algorithm (MALA), Hamiltonian Monte Carlo (HMC) as well as the No-U-Turn Sampler (NUTS). The BL-ARD model allows us to rank the features automatically in terms of importance. The results indicate that the MALA marginally outperforms HMC and NUTS for the recovery detection problem on unseen data using the Area Under the Receiver Operating Curve performance metric. Furthermore, we find that the purpose of the loan, home-ownership status, and the interest rate charged on the loan are some of the key features in determining which charged-off account will have recoveries or not. This framework can be easily extended to incorporate more complex models, such as Bayesian deep neural networks.