Kaggle competition: https://www.kaggle.com/c/m5-forecasting-accuracy
Welcome to my first kaggle competition! It is a top 7% solution.
Author: Yuman Lin and Shengjian Chen
This is one of the two complementary competitions that together comprise the M5 forecasting challenge. The main goal of this M5 competition is to estimate the unit sales of Walmart retail goods.
Data
https://www.kaggle.com/c/m5-forecasting-accuracy/data
calendar.csv - Contains information about the dates on which the products are sold.
sales_train_validation.csv - Contains the historical daily unit sales data per product and store [d_1 - d_1913]
sample_submission.csv - The correct format for submissions.
sell_prices.csv - Contains information about the price of the products sold per store and date.
sales_train_evaluation.csv - Includes sales [d_1 - d_1941] (labels used for the Public leaderboard)
Model
Core algorithm: lightgbm
Two ways:
(1) different models according to different state_id (total 3 models, CA, TX and WI)
(2) different models according to different store_id (total 10 models, CA_1, CA_2, CA_3, CA_4, TX_1, TX_2, TX_3, WI_1, WI_2, WI_3)
I use the second way as my final submission because it performs better in my validation set.
With the prediction of model 2, I apply an interesting trick (the code: M5 - Dark Witch Time by Store). https://www.kaggle.com/marutama/m5-dark-witch-time-by-store
It is a controversial notebook and even some comments say that the notebook distoried the public leaderboard. However, I did see a significant improvement on my prediction.
I only upload some of the notebooks but others are similar. You only need to change the store_id or state_id to train your data.
Evaluation
Weighted Root Mean Squared Scaled Error (RMSSE)
https://www.kaggle.com/c/m5-forecasting-accuracy/discussion/133834
Submission
The submission_v100.csv contains my final result. The RMSSE is 0.66751 (private leaderboard).
Last but not least, the RMSSE of my validation set is less than 0.5. Extremely overfitting! I believe most of the submissions suffer from the same problem: after all, the first prize is 0.52 but many open jupyter notebooks are less than 0.5...