Description

Zillow’s Zestimate home valuation has shaken up the U.S. real estate industry since first released 11 years ago.

A home is often the largest and most expensive purchase a person makes in his or her lifetime. Ensuring homeowners have a trusted way to monitor this asset is incredibly important. The Zestimate was created to give consumers as much information as possible about homes and the housing market, marking the first time consumers had access to this type of home value information at no cost.

“Zestimates” are estimated home values based on 7.5 million statistical and machine learning models that analyze hundreds of data points on each property. And, by continually improving the median margin of error (from 14% at the onset to 5% today), Zillow has since become established as one of the largest, most trusted marketplaces for real estate information in the U.S. and a leading example of impactful machine learning.

Zillow Prize, a competition with a one million dollar grand prize, is challenging the data science community to help push the accuracy of the Zestimate even further. Winning algorithms stand to impact the home values of 110M homes across the U.S.

In this million-dollar competition, participants will develop an algorithm that makes predictions about the future sale prices of homes. The contest is structured into two rounds, the qualifying round which opens May 24, 2017 and the private round for the 100 top qualifying teams that opens on Feb 1st, 2018. In the qualifying round, you’ll be building a model to improve the Zestimate residual error. In the final round, you’ll build a home valuation algorithm from the ground up, using external data sources to help engineer new features that give your model an edge over the competition.

Because real estate transaction data is public information, there will be a three-month sales tracking period after each competition round closes where your predictions will be evaluated against the actual sale prices of the homes. The final leaderboard won’t be revealed until the close of the sales tracking period.

Evaluation:

Submissions are evaluated on Mean Absolute Error between the predicted log error and the actual log error. The log error is defined as

logerror=log(Zestimate)−log(SalePrice)$$logerror=log(Zestimate)-log(SalePrice)$$

and it is recorded in the transactions training data. If a transaction didn't happen for a property during that period of time, that row is ignored and not counted in the calculation of MAE.

Submission File

For each property (unique parcelid), you must predict a log error for each time point. You should be predicting 6 timepoints: October 2016 (201610), November 2016 (201611), December 2016 (201612), October 2017 (201710), November 2017 (201711), and December 2017 (201712). The file should contain a header and have the following format:

ParcelId,201610,201611,201612,201710,201711,201712
10754147,0.1234,1.2234,-1.3012,1.4012,0.8642-3.1412
10759547,0,0,0,0,0,0
etc.

Note that the actual log errors are accurate the 4th decimal places, so you can adjust your decimal formats to limit the size of your submission file.

Data Description

In this competition, Zillow is asking you to predict the log-error between their Zestimate and the actual sale price, given all the features of a home. The log error is defined as

logerror=log(Zestimate)−log(SalePrice)$$logerror=log(Zestimate)-log(SalePrice)$$

and it is recorded in the transactions file train.csv. In this competition, you are going to predict the logerror for the months in Fall 2017. Since all the real estate transactions in the U.S. are publicly available, we will close the competition (no longer accepting submissions) before the evaluation period begins.

Train/Test split

• You are provided with a full list of real estate properties in three counties (Los Angeles, Orange and Ventura, California) data in 2016.

• The train data has all the transactions before October 15, 2016, plus some of the transactions after October 15, 2016.

• The test data in the public leaderboard has the rest of the transactions between October 15 and December 31, 2016.

• The rest of the test data, which is used for calculating the private leaderboard, is all the properties in October 15, 2017, to December 15, 2017. This period is called the "sales tracking period", during which we will not be taking any submissions.

• You are asked to predict 6 time points for all properties: October 2016 (201610), November 2016 (201611), December 2016 (201612), October 2017 (201710), November 2017 (201711), and December 2017 (201712).

• Not all the properties are sold in each time period. If a property was not sold in a certain time period, that particular row will be ignored when calculating your score.

• If a property is sold multiple times within 31 days, we take the first reasonable value as the ground truth. By "reasonable", we mean if the data seems wrong, we will take the transaction that has a value that makes more sense.

File descriptions

• properties_2016.csv - all the properties with their home features for 2016. Note: Some 2017 new properties don't have any data yet except for their parcelid's. Those data points should be populated when properties_2017.csv is available.

• properties_2017.csv - all the properties with their home features for 2017 (released on 10/2/2017)

• train_2016.csv - the training set with transactions from 1/1/2016 to 12/31/2016

• train_2017.csv - the training set with transactions from 1/1/2017 to 9/15/2017 (released on 10/2/2017)

• sample_submission.csv - a sample submission file in the correct format