A Machine-Learning Ensemble Method for Temporal-aware Sales Forecasting

Machine-learning-driven sales forecasting is gaining momentum in inventory management based on customer shopping behaviors. However, it is difficult to build a one-size-fits-all sales forecasting model for general purpose use due to the volatility of product turnaround and complex consumer behaviors. Traditionally, single forecasting models such as ARIMA modeling are used to predict sales. However, existing models have little real- world applicability given the apparent forecast disparity between experimental and real-world predictions due to model overfitting. This project proposes an ensemble Machine-learning method using weighted XGBoost, ARIMA, and Holt-Winters for sales forecasting of a grocery chain to address model overfitting issues. The intuition of the work is to combine several ML models to reduce variance (bagging), bias (boosting), and improve predictions (stacking). The contribution of this work includes: (1) implementation of exploratory modeling for in-depth exploratory data analysis (EDA) in grocery datasets and (2) feature engineering for the weighted integration of designed forecasting models using a lag and sliding window techniques. The preliminary result of the research places the model developed at the 70
^th
percentile of the Corporacion Favorita Sales Forecasting Competition in Kaggle.