songxxiao / m5_compete

Sliver Solution (Top 2%) for Kaggle M5 Forecasting competition / Kaggle M5沃尔玛销量时间序列预测竞赛 银牌方案(Top 2%)

Sliver Solution (Top 2%) for Kaggle M5 Forecasting competitions

Kaggle M5沃尔玛销量时间序列预测竞赛 银牌方案(Top 2%)

Xiao Song 宋骁

https://xsong.ltd/
Kaggle profile

M5 Forecasting – Accuracy – Top 103rd rank 2% (solo silver medal)
M5 Forecasting – Uncertainty – Top 18th rank 2% (solo silver medal)

Kaggle M5沃尔玛时间序列预测竞赛Baseline分享
Kaggle M5 沃尔玛销量时间序列预测 竞赛总结

Summary

Accuracy-Competition

• FE: adding a weekends feature.

• I used 4 single models:

  • time series split 3 fold cv - rmse loss function

  • time series split 3 fold cv - poisson loss function

  • time series split 3 fold cv - self-defined loss function

  • rolling prediction

• Then I calculate all prediction's mean value, the I recode very tiny prediction to 0. Finally, I use a magic multiplier close to 1.

Uncertainty-Competition

• Use public kernel released by KrisztianSz to convert point prediction to uncertainty. I use accuracy final submission as input and get a single submission 1.

• Use a public kernel released by Ulrich GOUE as a single submission 2.

• Finally, I did an averaging from above 2 submission files and convert all negative prediction to 0.

Happy Kaggling!