While I think the model's overall time-trending is performing pretty well, it never hurts to test some tweaks/improvements. One thing we can do to check the model's performance/understanding of time is look at the errors for each fold from rolling-origin CV. To improve performance, we can:
[ ] Try adding more complex features (lagged predictors, lagged outcome, etc.). See somethreads on this.
[ ] Test creating a separate index to use as a model feature
Calculate the index by area (Census tract, neighborhood, etc.)
Allows you to use a dedicated time series model for better/more performant forecasting
Helps get overall price trends correct in the main model
[ ] We can also revisit something we tried in the past: weighting more recent data using a decay function
While I think the model's overall time-trending is performing pretty well, it never hurts to test some tweaks/improvements. One thing we can do to check the model's performance/understanding of time is look at the errors for each fold from rolling-origin CV. To improve performance, we can: