Time on market and likelihood of sale prediction
A facility for estimating the value of a distinguished home, estimating the length of time a home or other property will be on the market at a listing price, and predicting the likelihood of sale of a home at a listing price is described.
1 . A method in a computer system for estimating a length of time from listing a home for sale at a listing price to sale of the home, the method comprising:
accessing, for each of a first plurality of homes that were listed for sale, values of attributes for the home including the listing price at which the home was listed for sale, and a listing information including a length of time the home was on a real estate market and a result of at least one home listing transaction;
collecting a plurality of sale transactions in a geographic area, wherein each sale transaction of the plurality of sale transactions is associated with a subject home;
training a model for estimating a time from listing to sale using the values of attributes for the home, the listing information, and sale transactions associated with the first plurality of homes in the plurality of sale transactions, wherein training the model also comprises testing the model and assigning relative weights to a plurality of decision trees associated with the model, wherein assigning the relative weights comprises:
for each decision tree of the plurality of decision trees associated with the model,
applying the decision tree to a second plurality of homes distinct from the first plurality of homes to generate estimated times on market for the second plurality of homes;
calculating an overall error measure based on a comparison of the estimated times on market for the second plurality of homes and actual times on market for the second plurality of homes; and
assigning a relative weight that controls a contribution of each of the decision trees of the plurality of decision trees to subsequent model outputs to each decision tree that is inversely related to the overall error measure associated with the decision tree;
using the model comprising the plurality of decision trees weighted according to the relative weights,
computing, for each sale transaction of the plurality of sale transactions, a ratio of the listing price for the subject home of the sale transaction to at least one value of the subject home of the sale transaction,
determining an aggregate of computed ratios, wherein the aggregate of the computed ratios includes a mean of the computed ratios or a median of the computed ratios, and
for at least one sale transaction of the plurality of sale transactions in the geographic area,
determining that the ratio computed for the at least one sale transaction is more than a threshold distance from the aggregate of the computed ratios, and
discarding, in response to determining that the ratio computed for the sale transaction is more than the threshold distance from the aggregate of the computed ratios, the at least one sale transaction from the plurality of sale transactions;
retrieving an automatically-determined estimate of a value of a first home in the geographic area;
estimating, using the model that includes the weighted decision trees, for the first home in the geographic area, a number of days from listing to sale based upon (a) the automatically-determined estimate of the value of the first home in the geographic area and (b) sale transactions of the plurality of sale transactions that have not been discarded; and
providing for display to a first user a graph that includes,
for each of a plurality of listing prices for the first home,
a first indication of a sale probability for the first home at the listing price for the first home, and
a second indication of an estimated number of days on market for the first home at the listing price for the first home.
2 . The method of claim 1 , wherein the result of the at least one home listing transaction is one of sale at some price, relisting, or withdrawal from the real estate market.
3 . The method of claim 1 , wherein the model incorporates listing price history and cumulative time on market of at least one home that has been listed for sale more than once.
4 . The method of claim 1 , wherein the model incorporates one or more valuations of at least one home that has been listed for sale.
5 . The method of claim 1 , wherein the model incorporates data characterizing the real estate market in which at least one home was listed for sale.
6 . The method of claim 5 , wherein the data characterizing the real estate market in which at least one home was listed for sale indicates average listing prices.
7 . The method of claim 5 , wherein the data characterizing the real estate market in which at least one home was listed for sale indicates average times from listing to sale.
8 . The method of claim 1 , wherein:
the model comprises a random forest regression model comprising the plurality of decision trees;
training the model, using the values of the attributes and the at least one home listing transaction, comprises, for each of the plurality of decision trees:
selecting a subset of the attributes that includes the listing price;
generating a trained decision tree by training the plurality of decision trees using the subset of the attributes, such that each leaf node of the trained decision tree represents a distinct combination of ranges of values of the subset of the attributes and a distinct subset of homes, each of the first plurality of homes being represented by exactly one leaf; and
storing, in connection with each leaf node, the estimated length of time from listing to sale based on the at least one home listing transaction of each of a subset of homes represented by a corresponding leaf node.
9 . The method of claim 8 , wherein storing the estimated length of time from listing to sale in connection with each leaf node comprises storing, for each of the subset of homes represented by the corresponding leaf node, the length of time the home was on the real estate market and the result of the at least one home listing transaction.
10 . The method of claim 8 , further comprising:
accessing, for each of the second plurality of homes that were listed for sale, the second plurality of homes being distinct from the first plurality of homes, second values of attributes for the home including a second listing price at which the home was listed for sale, and second listing information including a second length of time the home was on the real estate market and a second result of the at least one home listing transaction; and
for each trained decision tree:
for each home:
identifying a second leaf node representing attribute value ranges containing the second values of the attributes including the second listing price;
using the second leaf node to generate a second estimated length of time from listing to sale for the home; and
comparing the second estimated length of time from listing to sale to the second length of time the home was on the real estate market and the second result of the at least one home listing transaction to obtain an error measure for the trained decision tree and the home; and
obtaining the overall error measure for the trained decision tree across the second plurality of homes.
11 . The method of claim 1 , wherein the values of the attributes for the home include at least two imputed values, wherein the at least two imputed values include a value imputed using a median value among a set of values for a continuous variable and a mode value among a set of values for a categorical value.
12 . The method of claim 1 , wherein the model comprises a K-nearest neighbor model.
13 . The method of claim 1 , further comprising:
selecting, from among the plurality of sale transactions that have not been discarded, a number of sale transactions that is between a first threshold and a second threshold.
14 . The method of claim 13 , wherein the second threshold is at least four times the first threshold.
15 . The method of claim 1 , further comprising:
discarding sale transactions from the plurality of sale transactions whose computed ratios identify them as outliers having a top five percent of ratios and a bottom five percent of ratios.
16 . A computer-readable non-transitory medium having instructions that cause a computer to perform a method for estimating a length of time from listing a home for sale at a listing price to sale of the home, the method comprising:
accessing, for each of a first plurality of homes that were listed for sale, values of attributes for the home including the listing price at which the home was listed for sale, and a listing information including a length of time the home was on a real estate market and a result of at least one home listing transaction;
collecting a plurality of sale transactions in a geographic area, wherein each sale transaction of the plurality of sale transactions is associated with a subject home;
training a model for estimating a time from listing to sale using the values of attributes for the home, the listing information, and sale transactions associated with the first plurality of homes in the plurality of sale transactions, wherein training the model also comprises testing the model and assigning relative weights to a plurality of decision trees associated with the model, wherein assigning the relative weights comprises:
for each decision tree of the plurality of decision trees associated with the model,
applying the decision tree to a second plurality of homes distinct from the first plurality of homes to generate estimated times on market for the second plurality of homes;
calculating an overall error measure based on a comparison of the estimated times on market for the second plurality of homes and actual times on market for the second plurality of homes; and
assigning a relative weight that controls a contribution of each of the decision trees of the plurality of decision trees to subsequent model outputs to each decision tree that is inversely related to the overall error measure associated with the decision tree;
using the model comprising the plurality of decision trees weighted according to the relative weights,
computing, for each sale transaction, a ratio of the listing price for the subject home of the sale transaction to at least one value of the subject home of the sale transaction,
determining an aggregate of computed ratios, wherein the aggregate of the computed ratios includes a mean of the computed ratios or a median of the computed ratios, and
for at least one sale transaction of the plurality of sale transactions in the geographic area,
determining that the ratio computed for the at least one sale transaction is more than a threshold distance from the aggregate of the computed ratios, and
discarding, in response to determining that the ratio computed for the sale transaction is more than the threshold distance from the aggregate of the computed ratios, the at least one sale transaction from the plurality of sale transactions;
retrieving an automatically-determined estimate of a value of a first home in the geographic area;
estimating, using the model that includes the weighted decision trees, for the first home in the geographic area, a number of days from listing to sale based upon (a) the automatically-determined estimate of the value of the first home in the geographic area and (b) sale transactions of the plurality of sale transactions that have not been discarded; and
providing for display to a first user a graph that includes, for each of a plurality of listing prices for the first home,
a first indication of a sale probability for the first home at the listing price for the first home, and
a second indication of an estimated number of days on market for the first home at the listing price for the first home.
17 . The computer-readable non-transitory medium of claim 16 , wherein the result of the at least one home listing transaction is one of sale at some price, relisting, or withdrawal from the real estate market.
18 . The computer-readable non-transitory medium of claim 16 , wherein the model incorporates (a) listing price history and cumulative time on market of at least one home that has been listed for sale more than once, or (b) one or more valuations of at least one home that has been listed for sale.
19 . A computing system comprising a processor and memory storing instructions that when executed by the processor, cause the processor to execute a method for estimating a length of time from listing a home for sale at a listing price to sale of the home, the method comprising:
accessing, for each of a first plurality of homes that were listed for sale, values of attributes for the home including the listing price at which the home was listed for sale, and a listing information including a length of time the home was on a real estate market and a result of at least one home listing transaction;
collecting a plurality of sale transactions in a geographic area, wherein each sale transaction of the plurality of sale transactions is associated with a subject home;
training a model for estimating a time from listing to sale using the values of attributes for the home, the listing information, and sale transactions associated with the first plurality of homes in the plurality of sale transactions, wherein training the model also comprises testing the model and assigning relative weights to a plurality of decision trees associated with the model, wherein assigning the relative weights comprises:
for each decision tree of the plurality of decision trees associated with the model,
applying the decision tree to a second plurality of homes distinct from the first plurality of homes to generate estimated times on market for the second plurality of homes;
calculating an overall error measure based on a comparison of the estimated times on market for the second plurality of homes and actual times on market for the second plurality of homes; and
assigning a relative weight that controls a contribution of each of the decision trees of the plurality of decision trees to subsequent model outputs to each decision tree that is inversely related to the overall error measure associated with the decision tree;
using the model comprising the plurality of decision trees weighted according to the relative weights,
computing, for each sale transaction, a ratio of the listing price for the subject home of the sale transaction to at least one value of the subject home of the sale transaction,
determining an aggregate of computed ratios, wherein the aggregate of the computed ratios includes a mean of the computed ratios or a median of the computed ratios, and
for at least one sale transaction of the plurality of sale transactions in the geographic area,
determining that the ratio computed for the at least one sale transaction is more than a threshold distance from the aggregate of the computed ratios, and
discarding, in response to determining that the ratio computed for the sale transaction is more than the threshold distance from the aggregate of the computed ratios, the at least one sale transaction from the plurality of sale transactions;
retrieving an automatically-determined estimate of a value of a first home in the geographic area;
estimating, using the model that includes the weighted decision trees, for the first home in the geographic area, a number of days from listing to sale based upon (a) the automatically-determined estimate of the value of the first home in the geographic area and (b) sale transactions of the plurality of sale transactions that have not been discarded; and
providing for display to a first user a graph that includes, for each of a plurality of listing prices for the first home,
a first indication of a sale probability for the first home at the listing price for the first home, and
a second indication of an estimated number of days on market for the first home at the listing price for the first home.
20 . The computing system of claim 19 , wherein the result of the at least one home listing transaction is one of sale at some price, relisting, or withdrawal from the real estate market.