Predicting coastal sea level 2 weeks to 2 months out is very difficult, but sometimes certain recurrent climate conditions make this prediction problem more tractable. In this work, we use machine learning to identify periods where sea level is more predictable. It turns out that when we group points together into clusters and train one ML model per cluster, we can predict sea level accurately, identify more predictable recurrent climate patterns, and understand underlying mechanisms which make certain periods more predictable than others.