kiccho1101 / kaggle_m5_forecasting

My 207th🥈 solution for Kaggle M5 Forecasting Accuracy Competition

kaggle_m5_forecasting

My 207th🥈solution for Kaggle M5 Forecasting Competition (https://www.kaggle.com/c/m5-forecasting-accuracy/overview)

The detail of the solution: https://github.com/kiccho1101/paper/blob/master/Kaggle/M5.md

Features

• Rolling mean, std
• Rolling grouped mean, std
• Shift
• Discount rate
• Event strength
• catch22 features
• Target encoding features
• etc...

PB LeaderBoard History

2020-04-02 0.64561 (first submission)

2020-04-04 0.63581

2020-04-04 0.55002 (thanks to dark magic https://www.kaggle.com/kyakovlev/m5-dark-magic)

2020-04-05 0.53538

2020-04-10 0.51792

2020-04-11 0.50514

2020-04-11 0.48833 (thanks to iterative prediction https://www.kaggle.com/kneroma/m5-first-public-notebook-under-0-50)

2020-04-12 0.48273

2020-04-23 0.47101

2020-04-24 0.46930

Ponchi

Step1. Set up environments with pipenv

pipenv install --dev --skip-lock
sh install_lightgbm_2.3.2.sh

Step2. Download data

Unzip it and put the csv files in ./kaggle_m5_forecasting/ directory.

Step3. Start up luigi / mlflow server (in other terminal windows)

luigid  # localhost:8082

mlflow ui  # localhost:5000

Step4. Run the cross validation

pipenv run python main.py m5.LGBMCrossValidation