Seasonal-to-decadal climate prediction is crucial for decision-making in a number of industries, but forecasts on these timescales have limited skill. Climate institutions worldwide have provided an immense amount of climate model output which has made data-driven approaches to seasonal-to-decadal forecasting feasible and popular. One of the most intuitive approaches is “analog forecasting” — making forecasts based on how events with similar initial conditions evolved. In this presentation, I will introduce an interpretable neural network architecture which is trained on a proxy task to analog forecasting. The neural network provides a spatially-weighted mask that quantifies how important each grid point is for determining whether two climate states will evolve similarly. I will show that analogs selected using this weighted mask provide more skillful forecasts than analogs that are selected using traditional spatially-uniform methods. This method is tested on two prediction problems within a perfect model framework using the Max Planck Institute for Meteorology Grand Ensemble: decadal prediction of North Atlantic sea surface temperatures and seasonal prediction of El Niño Southern Oscillation. In addition to improving analog forecasts, the weighted mask provides a means for identifying and exploring sources of predictability.