時間序列趨勢預測機制之設計及其於原物料價格預測之應用

Abstract

原物料採購對於許多企業為一重要議題，然而原物料價格常有波動，故準確地預測原物料價格趨勢對於許多企業在提升競爭力上愈趨重要。目前關於原物料價格預測之研究，多專注於從過去某個時間區間裡的資料去建構特徵及趨勢模型，然而其並未完整考慮各種可能影響預測結果之因素，如連續趨勢之間的變化、採用特徵與否之分類學習方法以及各種相關參數等。有鑑於此，在本篇論文當中，我們提出了一個時間序列趨勢預測機制並將其運用於原物料價格趨勢預測。此機制中包含了以特徵為基礎的分類方法以及以非特徵為基礎的分類方法。在以特徵為基礎的分類方法中，我們在非循序領域下以及循序領域下建構特徵，再將建構後的特徵輸入以特徵為基礎的分類器去做預測。而在以非特徵為基礎的分類方法中，我們採用深度學習方法來自動建構特徵及預測模型。經由一系列使用真實資料集的實驗評估，可觀察到我們所提出來的分析架構能夠有效的預測原物料價格趨勢。

The material procurement is an important issue for many corporations. However, the material price fluctuates frequently. Therefore, precise forecast on the trends of the raw-material prices becomes increasingly important for many companies to enhance their competitiveness. Most of existing related researches focused on constructing features and models from trends at past time intervals. However, they did not fully consider rich factors that may affect the forecasting results, such as the varieties between the trends, the way of using features in classifications and a variety of related parameters. In this thesis, we propose an effective framework for forecasting the trends of the material price. The proposed framework include Feature-Based Classification and Non-Feature-Based Classification. In Feature-Based Classification, we construct features in both of the non-sequence domain and the sequence domain. These features are then fed into Feature-Based Classifier to do forecast. In Non-Feature-Based Classification, we use the deep learning method to automatically construct features and prediction models. We conduct a series of experiments on real data to evaluate the performance of our proposed framework. The experimental results show that our proposed framework are effective for forecasting the trends of the material prices.