Forcasting Butadiene Price by Artificial Neural Networks

[[abstract]]丁二烯為製造合成橡膠之主要原料，新興工業國家經濟快速成長下需求大增，若生產工廠故障或國際性經濟及政治事件，都將影響供需狀態，導致價格波動，相關業者面臨沈重的採購壓力。若能準確預測丁二烯價格，先行掌握原料行情趨勢，將可有效降低採購成本，增進整體經營績效。 本研究應用倒傳遞類神經網路(Back-Propagation Neural Networks BPNN)預測每周及每月之丁二烯價格，可使業者實際應用於採購工作中，了解行情趨勢，進行最佳採購規畫。模型輸入變數包含生產及應用面因素、相關聯產品因素及其他相關經濟性因素，於執行完結合不同網路架構及學習率的測試後，即提出一較佳 BPNN模型進行後續實驗，並進一步加入時間序列變數以調整BPNN模型輸入的結果。 經實證研究顯示，最佳週網路架構設定值為：一層隱藏層、十八個隱藏層處理單元數、十五個輸入變數及一個輸出變數。最佳月網路架構設定值為：一層隱藏層、六個隱藏層處理單元數、十四個輸入變數及一個輸出變數。天然橡膠、乙烯及粗碳四價格同為影響週及月價格預測模型之重要價格因素。另為測試丁二烯價格預測準確度，以連續20期為一驗證期間，得到週模型最佳預測準確率為93%，月模型最佳預測準確率為91%。這證明本研究所建構的類神經網路模型可提供企業於採購決策時之有用參考。[[abstract]]Butadiene is one of the key materials for producing synthetic rubber and its global demand increases rapidly as the economy grows quickly in the developing countries Its price fluctuates when the supply and demand are unbalanced resulted from the manufacturing plant shut down or the international economic or political factors The synthetic rubber makers face the pressure of procurement cost at all times If the price trend can be forecasted more accurately the makers can reduce the cost and improve their operation performance This study is to apply the back-propagation artificial neural networks (BPNN) to forecast both weekly and monthly butadiene prices that practitioners can implement in their daily work to understand the price trend and to precede an efficient procurement project We include up-stream and down-stream product prices related product prices and other economic indices as the inputs to the BPNN After performing different combinations of network structure and learning rate we propose a reasonably good BPNN model for the experiment We further experiment with time series variables to adjust the inputs of the BPNN model The resultant network structure includes one hidden layer with 18 nodes 15 input nodes and one output node for weekly model; one hidden layer with 6 nodes 14 input nodes and one output node for monthly model Experiments also show that the prices of natural rubber ethylene and crude C4 are the common variables affecting the weekly and monthly butadiene prices We test the accuracy of butadiene price forecasting for a period of consecutive 20 terms and find a best accuracy of 93% for weekly model and 91% for monthly model This demonstrates that the BPNN butadiene price forecasting model can assist companies in their procurement decision