code for Kaggle [M5 Forecasting - Accuracy] (https://www.kaggle.com/c/m5-forecasting-accuracy) Related data also can be downloaded from Kaggle.
A ensemble method of lightGBM and NN model is adopted: 0.6lgbm +0.4(lstm+cnn epoch3) For lightgbm, according to data provided by the organizer, we custom lag-7 and lag-28 and their mean feature which is pretty helpful for the prediction. For the NN model, we use layers combination of LSTM for remember history features, and 1-D CNN to capture neibougher features.
Result: Silver Medal, top 2% ranking 93/5558 in final private leaderboard.
General exploratory data analysis for the time series data among 5 years.
Random id choosed time series data plot.
Aggregate Sales plot.
More detailed EDA credit to [Kaggle Competition Notebook] (https://www.kaggle.com/headsortails/back-to-predict-the-future-interactive-m5-eda)
Run python lgbmodel.py
Will get the result in submission_lgbm.csv
Results of feature importance.
First preprocess train-test data, run python make_train_test_data.py
After get the data prepared, run python m5_full_train.py
Will get the result in submission_lstmcnn.csv
Run python test2.py
Will get final submission which we submit in the competition.