Partisan sorting by social groups is believed to increase affective polarization and decrease group-level leverage in representation. Mounting evidence suggests that social groups are increasingly polarized in voting behavior, but how reliable are demographic labels as predictors of vote choice? We test for demographic sorting, using public opinion surveys between 1952--2020 and applying tree-based machine learning models to calculate out-of-sample predictions of presidential voting decisions. We calculate predictions based on voters' demographics and then gradually incorporate more information to test whether the electorate is becoming more predictable. Demographics alone typically can predict 63.5% of vote choices correctly. But contrary to the sorting hypothesis's implications, demographics have not grown more predictive over time, while partisanship has. Additional information about voters, such as issue positions or candidate perceptions, continue to be necessary for obtaining out-of-sample error rates of 5% or less. However, their added value decreases as partisanship's predictive power grows.