Adaptive Generalized Non-linear Grey Bernoulli Model with applications to sales forecasting: A Case of Punch Wholesaler
| Autor: | CHEN, KUO-WEI, 陳國瑋 |
|---|---|
| Rok vydání: | 2017 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 105 Non-linear grey Bernoulli model (NBGM(1,1)) is a grey forecasting model proposed in the recent decade. The model is developed in order to improve the nonlinear adaptability of GM(1,1).Some researches provide ideas to improve its prediction capability, especially from the optimization of parameter aspects of power exponent, smoothing factor of background value, selection of the initial condition, or scaling factor of the residual modification. According to previous studies, most research has optimized only one or two parameters, and there are no studies to consider the above four. This study proposes a generalized NGBM(1,1) that takes above four parameters into consideration. As parameters are set to be the default values, the generalized NGBM(1,1) will degrade to a corresponding standard grey prediction model, totally 16 types of forms can be yielded. However, the process of manually adjusting the parameters is very boring, and it is difficult to confirm its correctness. Therefore, this study introduces the genetic algorithm (GA) to find the best parameters in an automated manner. In addition, this study finds that there may be a dependency relationship between the choice of model type and the structure of data sequence fluctuation. Not every data sequence structure is suitable for prediction through the generalized NGBM(1,1). The more the parameter is released, the more likely the in-sample is over-fitting and the prediction accuracy of the out-sample is extremely poor. In order to solve this problem, this study further proposes the model of adaptive generalized NGBM(1,1). First of all, the data sequence is divided into training data and testing data. The training data are used to establish a large number of data sequences. On the one hand the adaptive resonance theory 2 (ART2) clusters the typical data sequence structure; On the other hand, the generalized NGBM(1,1) associates with GA finds a suitable model for each sequence. Finally, we try to map the relationship between the different data sequence structure and the 16 types of prediction models. In testing stage, each data sequence is allocated to the most similar cluster throughout ART2, mapped to the corresponding suitable prediction models and then makes a one-step-ahead prediction. An empirical analysis on 10-year weekly sales data of a punch retailer in Taichung is conducted. Our proposed method is compared with some traditional prediction models. Experimental results show that the proposed method can provide higher prediction accuracy, helping the retailer to make appropriate ordering strategies. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|