A Details of proSVD algorithm We follow the notation of [ 20

Matrix sizes are listed for convenience in Table 2. Table 2: Matrix dimensions for incremental SVD. On subsequent iterations, the procedure is as follows: 1. Observe a new n b data matrix X What is most important to note in this is that the solution to (14) is not unique. Our implementation of Bubblewrap neither normalizes nor assumes a scale for incoming data. When predicting more than one time step ahead, we use sequential sampling for both models. SVD) to the experimental datasets used in the main text.