Forecasting Drinking Water Sales Values with Artificial Neural Networks: A Comparative Analysis with ARIMA and Winters Methods

This study aimed to forecast drinking water sales accurately for a water company dealer using the artificial neural networks method. The data used in this study is the total monthly sales number of dispenser-size water bottles of a water company's dealer in Bursa. The data consists of 85 months, from May 2017 to May 2024. In this context, an artificial neural network model was developed, and the estimations' performance was quite good. The histogram of estimation errors and normality tests showed a normal distribution. The findings show that the network can generalize. Besides this, visualizing the actual and estimated values showed that they follow the same patterns. As a result, it was concluded that monthly sales can be forecasted with the model developed using the threshold values and weights obtained from the trained network. Long-term forecasts were made and interpreted for the water company dealer using the developed model. Finally, the proposed artificial neural network was validated by comparing it with the average absolute percentage error values of alternative models, seasonal autoregressive integrated moving average, and seasonal exponential smoothing models.