KaggleStoreSales

Time series regression modeling on a dataset of supermarket sales across years, with the Darts library in Python.

Part 1: Predicting the total national sales, aggregated across all product categories & stores. Performing time decomposition & hybrid modeling, trying methods such as linear regression with custom features, AutoARIMA and random forest.

Markdown report, part 1

Part 2: Predicting the sales across all hierarchy levels (total sales, store totals and disaggregated series) and performing hierarchical reconciliation. Using Darts implementations of PyTorch global neural networks / deep learning models tailored for time series forecasting, such as D-Linear, LSTM and Temporal Fusion Transformer.

Markdown report, part 2

Kaggle competition submission: A simple submission using an AutoETS model for each of the 1782 disaggregated series resulted in a score of 0.42505 RMSLE, placing 61th out of 612 (top 10%) in the leaderboard at submission time (March 2023).

Kaggle competition notebook