Sales Forecasting Using High-Performance Computing

Naga Lakshmi Subha Pavan Kumar Ghantasala

Abstract: “Sales Forecasting Using High-Performance Computing” paper is intended to improve inventory management and assortment planning of a national retailer through reliable and efficient sales forecasting. It is focused on building upon the current “bottoms-up approach” model used by the company, by creating a robust regression model with optimal interaction terms, feature engineering, and hyper-parameter tuning using high-performance computing. Exploratory data analysis was performed on three categories of data- batteries, brakes, and filters, which provided information about demographics of customers, the geography of the store regions, gross revenue, and quantity sold for the past two years at a unique SKU-Store combination. Data aggregation and other transformations were done as per business requirements and for ease of modeling. Advanced programming was performed using Bell Cluster and Dask machine learning libraries. Several feature engineering experiments were carried out to create 2nd-degree interaction terms, PCA analysis, etc. The response variable was predicted by building several models such as linear regression, lasso regression, and OLS. The final regression model selected has the highest Adjusted R-square and lease root mean square error, which were the two metrics defined for model selection. This model had the highest interpretability and lowest run time.