Despite the growing spatiotemporal density of geophysical observations at subduction zones, predicting the timing and size of future earthquakes remains a challenge. Here we simulate multiple seismic cycles in a laboratory-scale subduction zone. The model creates both partial and full margin ruptures, simulating magnitude M w 6.2-8.3 earthquakes with a coefficient of variation in recurrence intervals of 0.5, similar to real subduction zones. We show that the common procedure of estimating the next earthquake size from slip-deficit is unreliable. On the contrary, machine learning predicts well the timing and size of laboratory earthquakes by reconstructing and properly interpreting the spatiotemporally complex loading history of the system. These results promise substantial progress in real earthquake forecasting, as they suggest that the complex motion recorded by geodesists at subduction zones might be diagnostic of earthquake imminence. Plain Language Summary Large and devastating subduction earthquakes, such as the 2011 magnitude 9.0 Tohoku-oki earthquake (Japan), are currently considered unpredictable. Scientists lack a long enough seismic catalog that is necessary for drawing statistical insights and developing predictions. For this reason, we simulate tens of earthquakes using a small-scale experimental replica of a subduction zone. We show that machine learning (a group of algorithms that make predictions based on the "information" acquired in past "experience") can predict when, where, and how big the next experimental earthquake will be. The "information" in our study is provided by the slow deformation accumulating in the analog tectonic plates during the periods in between earthquakes. Since such slow deformation is also measured by means of space geodesy along real subduction zones, there is the possibility that, in the future, variations of this machine learning approach can predict the timing and size of natural subduction earthquakes.