> an enormous set of available training data (all traffic happening everywhere across the country).
I'm not sure this is very helpful to predict traffic patterns over the next 20-30 years in a policy-sensitive way... All the historical data in the world wouldn't have predicted the rise of Uber, the pandemic, the maybe-eventual rise of self-driving cars etc.
The model won't tell you how likely these scenarios are, but one advantage of this class of model over black-box ML function-fitting is that you can do scenario planning. You can do models runs like "what happens if the price of oil shoots way up?" or "what happens if 50% of the cars are taxis?" (e.g. Uber-driven future) or "what happens if transit all of a sudden gains a huge perceived cost?" (e.g. respiratory-virus pandemic) in the third and fourth stages of the model. In the early stages of the model you can look at stuff like "what happens if firms abandon city centers?" or "what happens if 50% of white-collar work is done from home?" etc.
I'm not sure this is very helpful to predict traffic patterns over the next 20-30 years in a policy-sensitive way... All the historical data in the world wouldn't have predicted the rise of Uber, the pandemic, the maybe-eventual rise of self-driving cars etc.