This paper proposes extending temporal matrix factorization to incorporate neural network regularization for time series forecasting. The intuition is to capture global and local structure to make forecasts. The work is interesting because it bridges more classical forecasting ideas with new state of the art deep learning approaches. The ideas presented in this paper seem novel as the authors take existing building blocks for deep learning and combine them in a creative way to capture interesting structure of time series. This is in contrast to simply applying an existing deep learning approach such as seq2seq. There were some criticisms of the experimental evaluation which the authors seem to have addressed in their response and will include new comparisons that the reviewers asked for. The outstanding concern is in regards to understanding what each component of the proposed model is contributing to the increased performance as the current experiments do not elucidate this. However, I think that the ideas presented in this paper are very interesting and that the improved forecasting abilities of the model are good enough for publication now with a deeper understanding being followup work.