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!