Using a random sample of 60% of our cross-sectional data on U.S. stocks from 1964 to 2012, we trained four machine learning algorithms to forecast debt paydown over a one-year horizon. An evaluation of these candidate models on half of the hold-out sample (20% of the original dataset) showed that a boosted trees algorithm can forecast debt paydown with up to 70% precision over the next year. This boosted trees model achieved similar results in a second out-of-sample test on the remaining 20% of original data. While information on one-year-ahead equity returns was not used in training or evaluating any of the models, our results show that stocks with a higher estimated probability of paying down debt in the next year also earn higher average returns in that one-year-ahead period. A back-test of the boosted trees model’s forecasts of debt paydown between 1965 and 2012 shows a 10.3 percentage point spread between the average annual returns of portfolios formed from the 10th decile of estimated debt paydown probability versus annual portfolios formed from the 1st decile of estimated debt paydown probability. When the 10th decile is combined with a value investment strategy to focus on cheap leveraged stocks that are most likely to pay down debt, we find a CAPM beta of 1.18 and a statistically significant alpha of 9.3% per year in a four-factor model that controls for the three Fama-French factors and momentum.