Developing defuzzifying method of fuzzy time-variant series for forecasting product demand

Abstract

This thesis is about proposing a method of defuzzifying fuzzy time-variant series on forecasting demand by developing time-variant model considering Mean Absolute Error (MAE) and trend of primary forecasting (MAE&Trend). The previous method of defuzzifying was based on Artificial Neural Network (ANN). Defuzzifying by ANN needs model identification which is time consuming. To justify the robustness of the proposed method, it is compared with Song‘s and Chen‘s fuzzy time-invariant methods as well as Autoregressive Integrated Moving Average (ARIMA) time series method. Advantage of fuzzy method is its robustness in covering large range of data including linguistic data as well as numerical and its ability to forecast by small amount of data. Accuracy of the methods is achieved by comparing Mean Absolute Percentage Error (MAPE). The MAPE results are modified by omitting outranged data to have more logical evaluations. Based on the evaluations, the proposed model of MAE&Trend shows better accuracy in forecasting demand.