Current Competition:

• Predict futures sales
• Timeline: 2 months（01/20/2021）
• Current progress(leaderboard): 0.986 rmse.

DOING

• ETA: 12/20/20
• Re-implement the “Future Sale 3” notebook
  • Future Sales 3
  • Things to do:
    1. Data Cleanning
       • Same shop but have different id (USED)
       • OPTIONAL: Mis-spell in shop name and item category
       • OPTIONAL: Item name cleaning
    2. Features Engineering
       • Extract city/category from shop_name
       • Lag item_cnt_month(mean)
         1. month/shop (USED)
         2. month/item (USED)
         3. month/shop/item (USED)
         4. month/city (USED)
         5. month/city/item (USED)
         6. month/shop/type (USED)
       • price
         1. average price for item
         2. lag price for month/item
         3. delta … lag_price/avg_price (USED)
       • revenue
         1. lag total renveue for month/shop
         2. delta … lag_revenue/avg_revenue (USED)
       • Date:
         1. Month (USED)
         2. Days (USED)
         3. Months of first item/shop (USED)
         4. Months of first item (USED)
    3. Basic feature:
       • Item_id (USED)
       • Shop_id (USED)
       • Date_block_num (USED) Note: Switching to google colab from kaggle notebook for more consistent performance.

Result:

• Current baseline model: ?(public score)
  • model: single xgb with no hypertuning … 0.911 pls
  • completed: 12/18/2020

Next:

• Re-implement the “top 2…” notebook
  • Two step model(classification –> regression) … no improvement
  • Error analysis
  • Ensemble

---