Time series benchmark model

[amcarter]

To assess the performance of the benchmark AR model, I forecast the validation datasets. The model with slope and temperature performed very poorly, even at the sites with lots of data:

The RMSEs for predicting DO mean are high across all sites with an average of ~3.6

I think the poor performance compared to when we fit a single site is because the model is required to select parameters that are pooled across all of these sites, so it is unable to optimize them for a single one, even with lots of data.

Given how poorly this does, I don't think it makes sense to try to optimize it with different variables, etc.