Sberbank Kaggle Competition

Overview

This is my repository for the Kaggle competition involving Sberbank, the state bank of Russia. The objective of this competition is to predict realty prices in Russia, and doing so will involve capturing complex macroeconomic trends. The deadline is June 29, 2017.

Methodology

Data Cleaning

The Data Processing.R code shows how I prepped the data for XGBoost modeling.

• Null values are set to -1e6
• Yes/no variables are set to 1/0
• Training predictors are joined to the macroeconomic variables on the timestamp field
• Certain features are selected based on initial gut feelings (i.e.: I included variables that included raion because they describe features of the neighborhood, as described in data/data_dictionairy.txt)
  • Other predictors may be included later, but this will provide an initial look

Gradient Boosting

My first approach was a standard XGBoost model, which can be seen in the Initial GBM.R code. , I used caret to train another GBM using 4-fold cross-validation. I initially used a random parameter search to provide some guidance on parameter optimization. Root mean squared error is used as the evaluation metric, but it may be beneficial to use the metric by which submissions are scored (RMSLE).