Retail Demand Forecasting: a Comparison Between Deep Neural Network and Gradient Boosting Method for Univariate Time Series

The traditional retailer had to deal with stocking of inventory and merchandising based on raw estimates of his daily or weekly product sales to make available the right product at the right time for his customers. The modern retailer has to deal with same store operations such as inventory planning and merchandising but with several forecasting techniques at hand such as ARIMA, Holt-Winters, Exponential Smoothing, Neural Networks and Gradient Boosting Methods, he is efficiently and more accurately able to predict store-product sales to replenish inventory with minimum capital requirement. This paper aims at implementation and comparison of two popular and intriguing machine learning techniques, Deep Neural Network (DNN) and Gradient Boosting Method (GBM) for univariate time series sales data at store-day level of a German retail giant. Several studies have leveraged DNN and GBM techniques for forecasting but on a multivariate time series data, that is predicting with the help of causal factors (variables). The paper discusses the feasibility and application of both these techniques on a univariate time series because not every retailer may have access to adequate supply chain metrics to create a multivariate model. The data used for analysis has been picked up from Kaggle. Performance measures used are mean absolute error (MAE) and root mean square error (RMSE). The DNN and GBM models have been developed in python using Keras and Ensemble packages.