| 摘要: | 在服飾產業的快時尚潮流下，終端客戶的需求具有相當大的不確定性，而紡織產業鏈屬中下游的成衣製造業之商業模式大都為B2B，客戶通常是國際時尚品牌零售商如GAP、H&M、ZARA等，零售商為了降低成本與風險，訂單大都是以少量多樣交期短的型式，成衣製造業需配合客戶進行快速回應的協同管理來因應，然而在供應鏈的長鞭效應下，庫存成本的高低是直接影響利潤高低的關鍵因素之一，因此需要對市場需求更加敏銳，銷售預測的準確與否將是改善獲利及快速回應的關鍵，其並可優化生產和需求計劃，使得供應鏈的運作更有效率及快速，並減少浪費與節省成本。
本研究以國內上市的成衣製造的供應商為例，其銷售預測是仰賴業務人員以主觀經驗、過往客戶銷售相關數據與訂單透明度等資訊，整合進行人工的銷售預測，然而對銷售歷史大數據資料並無良好的大數據分析工具來協助銷售預測。研究證實使用適量的銷售歷史資料應用於機學習回歸模型確實能改善銷售預測的錯誤率。本研究進行初步的機器模型預測績效比較，實驗結果顯示，本研究個案的各資料集的預測績效以Ridge regression預測績效較佳，不論是在預設參數或是相對最佳參數組合，相對地Linear regression與SVR的預測績效則較差。在以預設參數執行預測的績效領先群模型中，進一步使用Grid search找出相對之最佳參數組合來執行預測評估，由於最佳參數組合是以MAE為評估指標，因此各模型MAE值皆比預設參數好，RMSE值亦多有改善，然而少數模型的預測績效並未改善或改善程度有限或無太大差異，例如GBR與DTR模型，顯然該模型的穩定性不適合於本研究個案的資料集，並且搜尋模型最佳參數在本個案的研究中相當的耗時。另外在模型訓練資料增加成為兩年度來預測下一年度時，其預測錯誤率會上升，例外情況是在銷售數量預測上，僅有ANN的預測績效更佳。
;Under the trend of fast fashion in the apparel industry, the needs of end customers are often quite uncertain. The business model of the garment manufacturing industry in the middle and lower reaches of the textile industry chain is mostly B2B. The customers are usually international fashion brand retailers such as GAP, H&M,ZARA, etc. In order to reduce costs and risks, retailers mostly use a small number of different orders with short lead times. The garment manufacturing industry needs to cooperate with customers to respond quickly and collaboratively. However, under the bullwhip effect of the supply chain, inventory Cost is one of key factors that affect profit directly. Therefore, it is necessary to be more sensitive to market demand. The accuracy of sales forecasts will be the key to improving profitability and rapid response. Production and demand planning can be optimized to make the operation of the supply chain more efficient. It is efficient and fast, and reduces waste and saves costs.
This study takes a domestically listed garment manufacturing supplier as an example. Its sales forecast relies on the integration of manual sales forecasts by business personnel based on subjective experience, past customer sales-related
data and order transparency and other information. However, the sales history is big data. There are no good big data analysis tools to assist in forecasting. This study conducted a preliminary comparison of machine model prediction performance. The experimental results show that the prediction performance of each data set in this case study,ridge regression model is better than others. Whether it is in the default parameters or the relatively best combination of parameters. Linear regression and SVR have poor predictive performance. In the performance leading group models with default parameters,that further execute prediction after using Grid search to find the relative best hyper parameters , due to the best parameter combination uses MAE as the evaluation index, the MAE value of each model is better than expected.Many models’ RMSE value has also improved. However, the prediction performance of a few models has not improved or the degree of improvement is limited or not much different, such as the GBR and DTR models. Obviously, the stability of these models is not suitable for the data set of this research case. And searching for the best hyper parameters of the model is quite time-consuming in this case study. In addition, when the model training data is increased into two years to predict the next year, its prediction error rate will increase. The exception is that only ANN′s prediction performance is better. |
