/Kaggle-M5-Forecasting-Accuracy

My solution to the Kaggle competition of M5 Forecasting Accuracy

Primary LanguageJupyter Notebook

Kaggle-M5-Forecasting-Accuracy

My solution to the Kaggle competition of M5 Forecasting Accuracy

Competition homepage

https://www.kaggle.com/c/m5-forecasting-accuracy

Exploratory data analysis

  • The EDA notebook is included, with highlights on findings of the data

Feature engineering and modeling

  • The most important features are lag features, created by a combination of lags, rolling windows, and aggregation functions on sales and prices.
  • The modeling is performed using LightGBM.
  • The hyperparameter tuning is done via 3-fold time-series cross validation.

Inference

  • The final prediction (unit sales of the next 28 days) is conducted via recursive inference.

Results

  • My final submission ranks top 3% among the 5500+ teams and wins me a silver medal!