All predictions start with an offset

[juancalderonbustillo]

Hi,

I have just been playing with Lightweight for a week or so, only with the example notebooks. I have encountered two "issues" and I wonder whether these are expected or not.

1. All predictions from lightweight start with a tremendous offset relative to the "train" data. Is this expected or is this happening because the signal is starting abruptly? Should one apply some sort of tukey or any other suitable window to the "train" time series?

2. In order to check self-consistency, I tried to use as "target_train" one of the predictions of lightweight, corresponding to a given set of predicted parameters (I'll call this an "injection"). While lightweight can perfectly recover the injected time-series (i.e., the injected target_train) I see that lots of the recovered parameters are inconsistent with those of the injection. Moreover, the summary table reports mostly r_hats way larger than 1. Is this expected?

Cheers,

Juan.

[uomodellamansarda]

1. I guess it is related to the fact that MMM is not a tool for prediction but optimization. It is usually not good at forecasting and it is generally used for scenario planning and optimization.
2. It can happen, but it means that the Markov chain didn't converge, so the results are not reliable. Here more details

Not sure if my answer is helpful, please let me know what do you think.