M5 Forecasting - Uncertainty

Kaggle Competition link

https://www.kaggle.com/c/m5-forecasting-uncertainty

Submission Details

The submission had a Weighted Scaled Pinball Loss score of 0.25547 on the private leader board and 0.06691 on the public leader board

Steps to run project code

Packages to be installed

Following packages specified in requirements.txt file need to be installed - keras, lightgbm, numpy, pandas, psutil, scikit-learn, scipy, tensorflow-gpu and tqdm

Folders needed to run code

Create folder named 'data' in the project directory and create sub-folders named - 'lgbm_datasets', 'models', 'processed_data', 'raw_data' and 'submissions' within it

Download Competition Data

Download the competition data files from Kaggle Competition Data Link and place them in the 'raw_data' folder

Also download the 'sample_submission.csv' file from Kaggle Competition Data Link, rename it to 'sample_submission_accuracy.csv' and place it in the 'raw_data' folder

Running Code

Run the __main__.py file. It is advised to run the code in pieces. For reference, it took me about 3 days on my HP Spectre i5

Model

• Initially 8 separate LGBM models are built for each store to derive point forecasts
• Model predictions of these 8 models are concatenated to form the LGBM point forecast
• 3 different Keras deep learning models are built with different embeddings to derive point forecasts
• Arithmetic average of these 3 keras models is used to create a deep learning point forecast
• A weighted average of the concatenated LGBM model and the averaged keras models in 1:3 ratio constitutes the final point forecast
• Finally uncertainty predictions are made using this this average point forecast