Day ninety. The listing is still live. A price cut has made the asking number look more reasonable, but the owner has yet to meet a buyer who can complete. Another home nearby has sold, which means there is now a comparable transaction to cite. The comparable does not pay this owner's mortgage.
This owner is hypothetical. The blind spot is real. A dataset built around completed deeds can tell us what sold, for how much and sometimes how quickly. It is less willing to tell us about the homes that never made it into a closing record. A listing can be withdrawn, expire, return under a new number or still be active when an analyst stops collecting data. Every one of those outcomes changes the practical meaning of a property forecast.
The owner is not deciding whether the home has a value. She is deciding whether to keep the sign in the yard. Cut again, wait, or take it down? A neighborhood average of days on market sounds like advice. It is not advice until we ask which homes entered the average, and at what point their clocks began.
The denominator matters
Say a hundred owners list homes in a neighborhood. If an analyst studies only the homes that sold, the median sale time might look brisk. Yet the owners whose homes remain listed have been removed from the denominator. For a new owner asking, “Will mine sell by winter?”, the missing listings may be the most informative part of the experience. The numbers in this example are illustrative; the logic is not.
One tool for this problem is a sale hazard. Rather than predict a single sale date, a model estimates the chance of a sale in the next interval, conditional on the house still being on the market. That chance can change over time. The first weeks bring fresh attention. A price reduction may change the buyer pool. New competing homes can arrive. A failed contract can send a listing back with its history attached.
The model also needs to deal carefully with observation that ends before the outcome does. If a house is still listed when the study closes, we know it did not sell during the observed period. We do not know whether it will sell next week. That is different from treating it as a permanent failure and different again from quietly deleting it. In survival analysis, such incomplete observation is called censoring. The word sounds technical. The honesty it demands is simple: do not claim to have seen an ending you have not seen.
There are other competing endings. A seller may choose to rent the property. A listing may be withdrawn for family reasons. Another may reappear under a new identifier after renovations. A model that confuses these outcomes can assign market conditions the blame for a choice it never observed.
The relisted home deserves special attention. Suppose a property is offered through the spring, withdrawn in summer, and returns with new photographs in autumn. If it sells soon after relaunch, a listing-level table can record a quick sale. The same home has already spent months testing the market. Linking episodes at the property level can recover that history, though matching records across agents and vendors is never perfect. Without that link, a slow attempt can reenter the evidence as a fast one.
Withdrawal is not a random censoring mechanism either. Owners with enough cash to wait, owners who cannot accept the market's offers, and owners whose circumstances have changed may leave for different reasons. Treating all departures as if they reveal nothing about demand would be another convenient way to make the simulation too sure of itself. A competing-risk model can give sale and withdrawal separate routes, but its estimates still depend on what the underlying records actually capture.