/M5-Forecasting-Accuracy

102nd place (silver medal).

Primary LanguagePython

M5-Forecasting-Accuracy

This repository is my solition for kaggle M5 Forecasting Accuracy. See competition website for the details.

Result

  • Local cv score: 0.567146
  • Private LB score: 0.60101

Features

  • Basic statistics features: 7~28days rolling statistics.
  • Lag features: 0~28days lag.
  • Price features: rolling stat, momentum etc.
  • Calendar features: seasonality, event flag, holiday flag etc.
  • Other features: days from release, zero sales ratio (7~28days).

Cross Validation

  • Custom timeseries split (3fold): last 2 month (fold1,2) + 1 year before (fold3)
  • Group k-fold (5fold): group = year + month

Models

LightGBM with tweedie loss. Blending 4 models with different cv strategies (TS or group) and different data split (by category or by week or no split).

name cv split fold1 fold2 fold3 avg weight private
Weekly Group Group Week 0.60118 0.50369 0.65736 0.58741 0.4175 0.57412
Weekly TS TS Week 0.61178 0.53015 0.68283 0.60825 0.2320 0.60082
Category Group Group Category 0.62392 0.48934 0.69283 0.60203 0.2052 0.54479
No split Group Group - 0.60705 0.49313 0.66869 0.58963 0.1732 0.58869
Blend - - 0.54787 0.51075 0.64282 0.56715 - 0.60101