Products Sales Forecasting Using Statistical And Machine-Learning Models – A Case Study رسالة ماجستير

Abstract

Sales forecasting is considered a pivotal tool to manage businesses from various disciplines, and a foundation to build an effective planning process in the company. Business owners’ priority appears mainly in making accurate sales estimates to limit the challenges of underestimating or overestimating sales that affect their business costs and operations. In this thesis, two statistical models were applied, which are Autoregressive Integrated Moving Average (ARIMA) and Seasonal Autoregressive Integrated Moving Average (SARIMA), and four Neural Networks (NNs) which are Recurrent Neural Networks (RNNs), Long-Short Term Model (LSTM) Multi-Layer Perceptron Neural Networks (MLPNNs) and Radial Basis Function Neural Networks (RBFNNs). A statistical model has been combined with each neural network model to build four hybrid models. This study examines the efficiency in predicting sales and capturing patterns for five products as pure models and as hybrid models. Two scenarios were followed to apply these models. The first scenario was a combination of sales for five products, and the second scenario was based on each product level (individually). Models performance evaluation were used (MSE, RMSE, and MAE). Final results have shown that forecasting sales individually for each product is better than forecasting sales for all products as a combination. Results have shown that hybrid models of ARIMA-MLPNNs significantly improve prediction accuracy compared to individual statistical models, four neural networks models, and other hybrid models for combined products sales. The ARIMA MLPNNs hybrid model has achieved an RMSE of 131.64 followed by the ARIMA-LSTM demonstrated an RMSE of 447.68, which has achieved better performance than the individual statistical model of SARIMA, four neural networks and other hybrid models. For individual product sales, the ARIMA-MLPNNs model has achieved RMSE of 31.13 for dairies, 19.54 for ice-cream, 51.74 for drinks, 60.99 for snacks and chips, and 74.21 for cleaning materials, while ARIMA-LSTM demonstrated better performance than individual statistical model SARIMA, four neural networks and other hybrid models with an RMSE of 80.54 for dairies, 50.64 for ice-cream, 169.13 for drinks, 188.03 for snacks and chips and 167.93 for cleaning materials. These findings suggest that hybrid models can provide more accurate predictions for products sales forecasting