Possibility of Location-Specific Latent Factors

[jbogomolovas2]

First of all big fan of mvgam! My to go package for time series analysis.

As this is not an issue, decided to post in discussion section.

So I have time series from several locations and three predictors-latent factors. Latent factors were derived using PLS ignoring the location. Now I want to use mvgam to "bend" those latent factors to particular location to fit and subsequently predict data better. So I run something like this:
model_static <- mvgam( mating_doy ~ s(latent1, series, bs = "fs", k = 5) + s(latent2, series, bs = "fs", k = 5) + s(latent3, series, bs = "fs", k = 5))
The issue is that some locations have a lot of missing data (bloom_doy; not latent factors) and would benefit from modeling as a state-space model. However it is not possible directly specify such a smooth structure under trend_formula. I am wondering is there a smart way around it? Any suggestions would be greatly appreciated.

[nicholasjclark]

Hi @jbogomolovas2, thanks for the nice comments about the package and for the helpful discussion. If I understand correctly, your series here indicates the different locations, and your goal is to try and learn nonlinear relationships between the factors and the mating_doy . Is this correct? If you want to take that formula and use it in trend_formula, you need to replace series with trend, i.e.

model_static <- mvgam( 
   formula = mating_doy ~ 1,
   trend_formula = ~ s(latent1, trend bs = "fs", k = 5) + 
           s(latent2, trend bs = "fs", k = 5) + 
           s(latent3, trend, bs = "fs", k = 5),
   ...,
)

Please let me know if that is helpful or if I'm not understanding correctly :)

[jbogomolovas2]

Thank you for your prompt response @nicholasjclark! Worked like a charm; just copying series under new name trend allowed to use trend_formula. State-space modeling indeed improved out of sample CRPS compared to simple model.

[nicholasjclark]

Great to hear @jbogomolovas2. I'll go ahead and close the discussion in that case, but feel free to ask more questions as you need to. All the best,
Nick