Term: Fall 2017
- Team #1
- Project title: Corporación Favorita Grocery Sales Forecasting
- Team members
- Shiqi Duan sd3072@columbia.edu
- Shuyao Hao sh3565@columbia.edu
- Jordan Leung jl4700@columbia.edu
- Jingkai Li jl4756@columbia.edu
- Peter Li pwl2107@columbia.edu
Project Summary: Overall we use Stacking method. That is, we build two layers. In the first layer, we obtain the prediction results from each first-layer model and use these results as features in the second layer to get the final prediction values.
In the first layer, we take use of both Time Series models as well as supervised machine learning models. For Time Series models, we first train them on the subtrain set with different parameters and use the models on the validation set to compare the performance. Then we get the best model and use that to predict the sales on validation set and test set, and call these results as features_valid and features_test respectively. For machine learning models, we train the models using cross-validation on the whole training set and get the best model. Then we also use the best model to do the same thing on validation set and test set.
In the second layer, we use machine learning models (Linear Regression, Random Forest, GBM, and Xgboost) on the features_valid by cross-validation or sampling to find the best model. Then use the best model on features_test to get the predicted scores on test set.
After we get the scores from 2nd Stack, we try to use the mean of predictions from the models with good scores as the final prediction. Then use the final prediction to get the final predicted score.
First Stack
- Time Series models
-
- ETS (Exponential Smoothing State Space Model)
-
- ARIMA
-
- Prophet
- Machine Learning models
-
- XgBoost
-
- Random Forest
Second Stack
- Linear Regression
- Random Forest
- GBM (Gradient Boosting Method)
- XgBoost
**Evaluation: We use the same evaluation formula as on Kaggle evaluation formula.
Main file: main.Rmd or the PDF version.
Contribution statement: (Please see this file for the statement) All team members approve our work presented in this GitHub repository including this contributions statement.
Following suggestions by RICH FITZJOHN (@richfitz). This folder is orgarnized as follows.
proj/
├── lib/
├── data/
├── doc/
├── figs/
└── output/
Please see each subfolder for a README file.