We study the accuracy of homeowners' self-reported estimates using data from the 2001-2019 American Housing Survey. By comparing sellers' estimates with deflated transaction prices for the same properties, we find evidence that American sellers overestimate the value of their homes by 1.3% on average. The mean absolute error is equal to 15.6%. Correcting for selection bias, we find that the (absolute) valuation error is strongly related to several household characteristics, most importantly marriage and education. Moreover, we find that housing market conditions such as recent local price growth and volatility are important determinants for the size and direction of the error. In particular, volatility in house prices decreases the accuracy of homeowners in estimating their property's value. Our results imply that transaction prices are preferred over self-assessed values in several applications such as wealth estimations and hedonic price models.