Earthquake is one of the most destructive natural disasters, with catastrophic results such as landslides, tsunamis, fires, and fault ruptures. Therefore, anticipating earthquakes ahead of time has economic and societal benefits. However, earthquake prediction is not a trivial task, since earthquakes can occur in a variety of magnitudes and frequencies and can exhibit different types of behavior depending on the location, time, and geological setting. This variability makes it challenging to develop a single model that can accurately predict earthquakes across different regions and timescales. In light of this, we predict the probability, severity level (class 0–4) and location of the earthquakes with an end-to-end framework named Shock-Alert by exploiting spatial-temporal correlation. The framework is comprised of parallel probability prediction and classification blocks, comprised of transformer and convolution networks, respectively, as the transformer will capture the temporal features due to the self-attention mechanism and the convolution network will capture the spatial features. To assess the effectiveness of the proposed framework extensive experiments are performed on the California dataset, and F1-score of 94.4% and 93.2% are obtained for severity level prediction, respectively; outperform baseline and state-of-the-art models.