In the recent years, there have been a lot of research conducted on earthquake prediction using the mathematical models of machine learning. Since the inception of seismology, there has been a constant search for a proper and consistent precursor for predicting earthquakes, but none of such definitive indicator has been found yet. With the recent advancements of machine learning models and frameworks, we get to see a lot of models being trained on earthquake data, which are then used to forecast future earthquakes based on extracted features. But the fundamental problems with such models is that they are reliant on post event features for predicting the earthquakes events rather than using pre-event features. Due to this flaw, the models often achieve artificially high accuracy as they are trained on information that wouldn’t be available in a real predictive scenario. However, in practical real world scenarios, we cannot introduce post event data of an earthquake to such a model because a predictive model cannot use data from an event that has already occurred. In this paper we introduce a new approach for structuring and pre-processing earthquake data in such a format that it doesn’t, in application, reveal any prior information or features about any particular event when training the model. Then we train different machine learning models on this data to try and forecast earthquakes. This helps to maintain the practical applicability of using such models in the real world while maintaining a decent level of accuracy. We train the models using geographically specific location blocks, and then evaluating their zero shot performance in new unseen locations.