In this article, we examine the accuracy and bias of market valuations in the U.S. commercial real estate sector using properties included in the NCREIF Property Index (NPI) between 1997 and 2021 and assess the potential of machine learning algorithms (i.e., boosting trees) to shrink the deviations between market values and subsequent transaction prices. Under consideration of 50 covariates, we find that these deviations exhibit structured variation that boosting trees can capture and further explain, thereby increasing appraisal accuracy and eliminating structural bias. The understanding of the models is greatest for apartments and industrial properties, followed by office and retail buildings. This study is the first in the literature to extend the application of machine learning in the context of property pricing and valuation from residential use types and commercial multifamily to office, retail, and industrial assets. In addition, this article contributes to the existing literature by providing an indication of the room for improvement in state-of-the-art valuation practices in the U.S. commercial real estate sector that can be exploited by using the guidance of supervised machine learning methods. The contributions of this study are, thus, timely and important to many parties in the real estate sector, including authorities, banks, insurers and pension and sovereign wealth funds.