This paper develops a pairs trading strategy via unsupervised learning. Unlike conventional pairs trading strategies that identify pairs based on return time series, we identify pairs by incorporating firm characteristics as well as price information. Firm characteristics are revealed to provide important information for pair identification and significantly improve the performance of the pairs trading strategy. Applied to the US stock market from January 1980 to December 2020, the long-short portfolio constructed via the agglomerative clustering earns a statistically significant annualized mean return of 24.8% and a Sharpe ratio of 2.69. The strategy remains profitable after accounting for transaction costs and removing stocks below 20% NYSE-size quantile. A host of robustness tests confirm that the results are not driven by data snooping.