Time Series Forecasting Using Neural Networks Minimizing Food Waste By Forecasting Demand in Retail Sales

Abstract

A third of the food produced for human consumption is wasted annually, amounting to 1.3 billion tons of food waste per year [10]. Minimizing these enormous quantities of waste would not only be beneficial for the planet but also help feed an ever increasing human population. One way of minimizing this waste is by helping retail sellers better plan their logistical operations by accurately predicting the demand of goods using forecasting models. The procedure in time series forecasting has traditionally been to use statistical models such as autoregressive integrated moving average (ARIMA) and exponential smoothing methods. These methods have been shown to be limited in their predictive capabilities as the sizes of data sets and the number of variables increases. In the last decade new machine learning algorithms have been used extensively in various fields and has opened up the door for utilization of models based on neural networks in time series forecasting. A subset of these new machine learning algorithms, such as transformer based models and recurrent neural networks, have been proven to be especially suitable for temporal data. In this thesis we investigate two models, Temporal Fusion Transformers (TFTs) and Deep Temporal Convolutional Networks (DeepTCNs), showcasing their abilities to generate accurate forecasts of retail sales on a real-world data set. We demonstrate, using several metrics, their ability to outperform baseline models.