Methods, systems, and apparatus for training model parameters stored in shared memory to predict risk. The method may include obtaining training data that includes a plurality of training data structures that each represent attributes of an entity, wherein each training data structure represents (i) features derived from a first set of categories defined by a first model and from a second set of categories defined by a second model, and (ii) a risk-level associated with the entity. For each respective training data structure, providing the training data structure as an input to the model, receiving an output from the model based on the model's processing of the training data structure, determining an amount of error between the output of the model and the risk-level of the training data structure, and adjusting a parameter value of the model stored in a shared memory based on the determined error.