This study is the first to examine the real estate-specific determinants of REIT bond risk premia. Using a dataset of 33,857 U.S. REIT bond yield spreads and 24 explanatory variables, we predict REIT bond yield spreads with a non-parametric artificial neural network algorithm and interpret the model’s predictions using the explainable machine learning method Accumulated Local Effect Plots (ALE). We report evidence of a direct real estate factor for U.S. REIT bond yield spreads proxied by real estate market total return and REIT property type. In addition, we find a property-type diversification risk premium for REIT bonds, indicating that there is no economic benefit in the form of lower cost of bond debt for most property-type diversification at the REIT-level. We argue that this is due to higher management and valuation complexity of diversified REIT portfolios. This study’s findings have relevant implications for REIT portfolio strategy and REIT capital structure decisions, as we show that specialized REITs generally have lower bond debt costs compared to diversified REITs. Moreover, a better understanding of the drivers influencing REIT bond risk premia helps investors to effectively manage bond portfolio risks.