Multi-card sales are excluded from the sales used to train the model for multiple reasons. The model predicts values per card, not per property; as such, multi-card sales are excluded from training the model, because the model won't know which card to attribute how much of the sale price to. But is there something we can do for modeling multi-card properties?
Some ideas
Constructing a new variable: card's percentage of bldg sqft of total bldg sqft on the parcel. Check with Valuations on whether this proxy makes sense from a valuation perspective. Then can we apportion sales of multi-card PINs, such that the sale price is divvied up per card, and we could use that to train a multicard model?
What about other variables, such as "key card" (largest percentage) and non-primary cards? This might help the model learn which cards are and are not ADUs.
Multi-card sales are excluded from the sales used to train the model for multiple reasons. The model predicts values per card, not per property; as such, multi-card sales are excluded from training the model, because the model won't know which card to attribute how much of the sale price to. But is there something we can do for modeling multi-card properties?
Some ideas