Despite governors’ crucial roles in shaping important policies, including abortion, education, and infrastructure, forecasters have paid little attention to gubernatorial elections. We posit that institutional idiosyncrasies and lack of public opinion data have exacerbated the classic problem facing all election forecasts: there are too many predictors and too few cases, leading to overfitting. To address these problems, we combine new governor and state-level presidential approval data with a machine-learning approach, LASSO, for variable selection. LASSO examines numerous variables but retains only those that substantively improve model performance. Results demonstrate the efficacy of gubernatorial and presidential approval ratings measured two quarters preelection in predicting both incumbent-party vote share and election winners in out-of-sample predictions. For 2022, our approach outperformed the Cook Political Report’s Partisan Voting Index and compared well with 538’s Election Day prediction. For 2024, our LASSO-Popularity model predictions indicate that it will likely be a difficult year for Democrats in gubernatorial contests.