Distribution: A. Approved for public release – distribution is unlimited QuakeCast is a novel system for short-term earthquake prediction using global ionosphere Total Electron Content (TEC) data. In the last 20 years, earthquakes have caused over 800,000 deaths and $650 billion in economic damage. While current seismic early warning systems provide up to a minute’s notice by detecting an earthquake’s non-damaging P-waves prior to the damaging S-waves, earlier warnings could allow more significant emergency preparations. Electromagnetic ionospheric phenomena have recently been observed many days before major earthquakes, notably in 2015 in Nepal. QuakeCast explores whether such signals could be used to predict earthquakes in a global dataset of ionosphere and earthquake data. The ionosphere data consists of global TEC data from NASA’s Crustal Dynamics Data Information System (CDDIS), for the years 2005-2015 at 15 minute intervals. The earthquake data consists of over 10,000 events of at least M5 from the International Seismological Centre (ISC-GEM) Global Instrumental Earthquake Catalogue over the same time period. We used the dataset to explore the following classification problem: given a 24-hour sequence of ionosphere TEC data in a 30° latitude by 30° longitude window, is the sequence preseismic (an earthquake occurs in the next time step) or nominal (no earthquake)? We built two models to address this. The first is a classical logistic regression model trained on radially binned data. Analysis of the classifier weights indicate a localized effect in the ionosphere: TEC data spatially and temporally closer to the event contributes more significantly to the prediction. The second model is a deep learning ConvLSTM autoencoder trained to reproduce nominal TEC data sequences. Since autoencoders reproduce data which resemble their training data better than data which does not, we used the reconstruction error to classify sequences as anomalous (preseismic) or nominal. Both methods were found to perform significantly better than a random null classifier, indicating that the ionosphere contains information useful for predicting earthquakes at least 15 minutes in advance. While further research is needed to incorporate additional data features and address noise, we believe that these results are a promising development toward forecasting earthquakes using geoelectric signals.