Sales forecasting using long short term memory networks

Abstract

The Annapolis Cider Company would like to estimate their future hourly and daily sales based on their historical sales, date and time, weather conditions and knowledge of special events in Annapolis Valley. This study uses machine learning methods to predict hourly and daily sales for up to two days in advance. The Annapolis Cider Company has provided sales data stored in their POS system for up to four years. Both observed and forecast weather data have been retrieved from the Dark Sky website. Special event data has been extracted from various online calendars from the web. This study focues on the performance of LSTM neural networks compared to ARIMA, persistence and Avg6Week models. Mean absolute error (MAE) and mean absolute percentage error (MAPE) are used to evaluate the performance of each model. The LSTM daily prediction model provides the lowest test MAE of $597.93 and test MAPE of 19.7% on daily basis. It decreases the test error by 22.6% compared with a persistence daily model. The LSTM hourly prediction model has the lowest test MAE of $51.59 and MAPE of 40.8% on an hourly basis, which is 17.2% lower than the persistence hourly model.