This was a two-week competition to predict apartment sale prices over the following three months from historical transaction data. I focused on recovering price-relevant features with extensive missing values by combining architectural domain knowledge with external data. On a four-person team, I handled EDA, feature engineering, and external data collection; model selection and tuning were collaborative.
Using domain knowledge to guide EDA
Gross floor area, building footprint, parking capacity, floor level, and XY coordinates all had substantial missingness, so we initially removed them. I brought them back because they contained information needed to explain price. Conversely, I excluded features with little relevance to sale price, such as road-name address and title-registration filing date. I also changed the definition of a newly built property from the existing 15-year cutoff to the 20-year threshold used for aged and substandard buildings. Rather than reducing features indiscriminately, I used architectural domain knowledge to decide which values to retain and which to discard.
In QGIS, I extracted building footprint, gross floor area, site area, and parcel-centroid coordinates for each address and joined them to the original data. For values that external data could not fill, I used linear interpolation or Iterative Imputer according to each column's characteristics.
Validating data against official records
I recomputed floor area ratio and building coverage ratio from the dataset's gross floor area, building footprint, and site area, then compared them with the ratios in official records. I also checked the average number of floors. When the values in a row could not be reconciled, I classified the inconsistent values as outliers, removed them, and imputed them again. The priority was to clean data that could not be explained in the architectural domain before changing the model.
Score progression
Root mean squared error (RMSE) fell from an initial 5894.0 to 5326.6 after the first round of EDA, then to 4922.5 after filling missing values with external data. An XGBoost model with 1,000 trees and a maximum depth of 10 scored 3850.0. A later experiment that also added the base interest rate and unit-size categories reduced it further to 3761.9. We placed first overall.