This algorithm finds representative clusters for our dataset. Note that here, we don't want to do this by cohort (if we wanted to, we would first need to subset the data). The horizontal bars reflect the clusters, and the A-B parts on top are used to calculate the discrepancies between the data and the cluster.

Type 1 and Type 3 refer to sequences starting at L1, then moving to L2 at around Week 9, and then remaining for either a few weeks or for many weeks. Both are quite homogeneous, reflecting that the sequences don't diverge much with respect to the representative ones.

Type 2 relates to sequences starting at L1, and then those accounts closing almost immediately. Here, the mean differences with respect to that sequence are even smaller, which is to be expected: once an account is closed, it is not reopened, so we should expect closed accounts to be quite homogeneous.

Type 4 is interesting. It reflects that after opening an L1 account, those clients jump directly to L3. We have two bars, which reflects that the algorithm is finding two large groups—people jumping to L3 in Week 9, and people jumping to L3 in Week 4.