The hedonic method is widely applied because of its ability to recover intuitive price effects of nonmarket goods from readily available data. These valuations become inaccurate when housing markets are too broadly defined, however. We address this issue by developing a novel statisticallearning method to better delineate housing markets across space, time, and attributes and by examining how this decision impacts valuation. Our methodology not only provides insight into the quantity of markets but along which boundaries to define them. We find within our application, where the implicit price for seismic risk is causally identified using the Tohoku Earthquake as a natural experiment, that housing transactions are best categorized into three submarkets defined across commuting time to central Tokyo. More importantly, this specification is statistically preferred and yields aggregate valuations that are 23.7% ($485 million) less relative to a singlemarket approach.