Μ5 competition - time series forecasting by implementing EDA, feature engineering and modelling with statistical and machine learning methods

Abstract

The purpose of this dissertation was to analyse the daily sales for several items sold in different Walmart stores in different US cities within 5 years and proceed to forecast the sales for 27 days, based on the historic sales data. Our methodology was to first conduct an Exploratory Data Analysis (EDA) on the provided datasets in order to better understand and gain insights from our datasets. Then we proceeded with the Feature Engineering (FE) in order to prepare and optimize our data for use in the Modelling phase. Several different Modelling methods were examined and evaluated and the best was selected based on their comparative RMSE. It was concluded that LightGBM was the best model, which accomplished the highest accuracy in predicting the sales values for the items in the defined period.