In the last chapter, we examined and built a convolutional neural network (CNN). We even combined it with the LSTM architecture to test whether we could outperform the LSTM models. The results were mixed, as the CNN models performed worse as single-step models, performed best as multi-step models, and performed equally well as multi-output models.
Now we’ll focus entirely on the multi-step models, as all of them output the entire sequence of predictions in a single shot. We’re going to modify that behavior and gradually output the prediction sequence, using past predictions to make new predictions. That way, the model will create rolling forecasts, but using its own predictions to inform the output.
This architecture is commonly used with LSTM and is called autoregressive LSTM (ARLSTM). In this chapter, we’ll first explore the general architecture of the ARLSTM model, and then we’ll build it in Keras to see if we can build a new top-performing multi-step model.