Forecasting of future yarn demand in a sewing thread company by using statistical and neural network based methods - a case study

Abstract

Demand forecasting plays a critical role in every business, especially in the manufacturing industry. Most of the operational decisions in the manufacturing industry are based on some kind of forecast for future demand. As a result, manufacturing companies pay high attention to the demand forecasting process, and this study has devoted attention to this particular issue. The objectives of this research are to evaluate the demand forecasting methods, analyze the sales data of the company using several forecasting techniques, and lastly, propose the most suitable forecasting method to the sewing thread company. The forecasting methods used in the analysis involve the time series forecasting method and the neural network-based forecasting method. The forecasting methods are assessed by using error measurement tools such as mean absolute deviation (MAD), mean squared error (MSE) and mean absolute percentage error (MAPE). These forecast error analyses are used to monitor the forecast results of various methods. The result of this study shows that an artificial neural network is chosen as the most suitable forecasting method as it produces the most accurate result with the least forecast error. The ANN and ANFIS model were constructed by considering the seasonal factor of the current quarter, the actual sell of the current quarter, the average forecasted demand for the last two quarters, and the last year's respective quarter demand as input parameters, whereas the forecast was the only output parameter. The system was trained with 60% data for constructing the ANN model, whereas testing and validation were done with 40% data. To train the network, feed-forward backpropagation with the Levenberg-Marquardt learning algorithm was used. The coefficient of determination (R2) was found to be 0.84 and 0.65 for the ANN and ANFIS models, respectively. Both prediction models exhibited excellent mean absolute error percentages (MAPE) 0.000005% for the ANN model and 19.5% for the ANFIS model. Furthermore, an outstanding root-mean-square error (RMSE) of 0.242 and 11,34,629 for the ANN and ANFIS models were observed. These results suggest an excellent performance of the developed models in predicting yarn demand in the sewing thread industry.