Need advice for modeling pipeline and hyper optimization

[ramdhan1989]

hi @petroniocandido
Following my comment in #1. I have 100 series that need one step ahead forecast. for your information, I have other series that can be used as external information. short description about my data

• each series has different behavior. some of them are stationary, non-stationary, some show strong seasonality, and some show drifting (sudden value jumping). to tackle this complexity, I am thinking to build one model for each series (what do you think?)
• data is monthly data where some series are complete and some series have missing values
• data range is very huge. it can be between negative millions and positive millions

I plan to try different approach both univariate (without any external information) and multivariate (with external information).
question apply for both scenarios :

1. I see there are a lot of FTS model implementation. in order to build robust automatic modeling pipeline for my problem, would you mind mapping what FTS model that work best for what kind of problem or type of series ? is there any model that can be used for any type of series but what I need is to get best hyperparameter setting ?
2. would you mind summarizing which model that work for 1 order and which one that can work with more than 1 order ?
3. does the implementation require complete data (no missing values) ? if yes, I am thinking to input the missing values but I build another series indicating whether the row is missing or not. what do you think ? is it the good practice ?
4. I see there is hyperparam search modul but I couldn't find examples to exercise. do you have examples of it ? my experiments on using FTS for my data showing that type of transformation, order, model impact much to the result. different options give very different result. So, I think, automatic search hyperparam would be very helpful. However, I haven't tried a different partitioner.

question apply for multivariate scenario :

1. how does the multivariate approach work here ? does the other series that act as external information can be used as future information (use t+1 external information to predict target y t+1) ?
2. if I can't use future value of external informations (use t + 0 as the latest external information to predict target y t+1). does the implementation support it ?

please advice,

thank you