Hi Gonçalo Abreu, thanks for the feedback!
It is a first approach by TensorFlow (TF) in giving off-the-shelf probabilistic methods to help on fitting and forecasting structural time series models. It is a different approach from the one that we decided to use, nevertheless it is also a trendy one. Adding up to that, they already have interesting features like the use of external variables or multiple seasonality effects.
It is an advantage to run these models in TF’s platform, as you can process efficiently many time series in parallel.
What it misses in our opinion:
- Extract the maximum knowledge that you can from the time variable — you can see on the Forecasting Demand for Electricity that there are still autocorrelations in the residuals, meaning that there is still information to extract;
- Leverage information that can be nested in sub-structures of the data;
- Combine optimally forecasts for different levels of aggregation. For us this is quite important, as brands don’t take decisions based on their overall sales, but on specific categories, gender, age group, product type, etc.
Looking at the future, as we shared in our post, hybrid models are the trend today in the field (check the M4 competition).
It will be very interesting to see what is coming next in TF!