類神經網路在預測台灣貨幣市場利率上的應用

Abstract

本研究擬建立一套利率預測模式，以提供銀行經營者在制定利率政策時之 參考。此利率預測模式將應用類神經網路於學習上的特性來建立。在利率 預測模式輸入資料的部份，首先是以總體經濟理論配合計量經濟模式整理 出各種影響利率的經濟變數。再從務實的角度，將這些經濟變數以問卷的 方式，發予臺灣金融界的專家，評估各種影響利率的經濟因素及其影響的 程度。經問卷統計的結果分析後，篩選出影響利率的經濟因素。同時蒐集 各經濟因素相對的經濟指標，以作為利率預測模式學習預測的樣本資料。 而這些樣本資料則利用再處理的過程，以產生三種不同的數據型式，有原 始數據ˋ每月數據變化量ˋ三個月數據移動平均值等。將這些處理過的數 據代入利率預測模式中進行學習。利率預測模式的學習效果是以輸出值及 目標向量的平均絕對誤差值作為評估的依據，透過不同數據型式的輸入模 式將產生不同的學習效果。利率預測模式若再經由不斷的模式設定改變， 其預測效果可經由此一訓練程序達到最佳狀況。在利率預測模式設計的部 份，則採用監督式的類神經網路模式，以逆向傳遞演算法( Backpropagation) 為利率預測模式的主體。經由模式實驗的結果，其最 佳預測效果平均絕對誤差值達5%以內。 One mode is established in the forecasting of the interest rate during this present research effort for the sake of provid- ing a reference for the banking administrator in determining the interest-rate policy. This mode in the forecasting of the inter- est-rate would utilize the characteristic of Neural Network appication in learning. The input data section is involved with the mode in the for- ecasting of the interest rate. The Macro-economic theory a acco- mpanied with the econometric models are first used for arrang- ing every of economic variable which would affect the interest rate. From a practical viewpoint, these economic variables are next converted into request form and sent to the banking expert for evaluation of every kind of economic factor which would aff- ect the interest rate and their affecting degree. Once the stat- istical result from the request form have been analyzed, those factors which would affect the interest rate are next identif- ied.The related economic index for each economic factor is simu- ltaneously collected and function as the sampling data concerned with learning in the forecasting of the interest-rate.These man- aged data are inputted into the forecasting of the interest rate during the processing of learning. The learning efficiency about the mode in the forecasting of the interest-rate is evaluated by the averaged ABS's error between the output value and object vec- tor.The effect in the forecasting would attain the optimum cond- tion through means of this training process, that is if the mode is continuously changed in the forecasting of the interest-rate. The design section about the mode in the forecast of the int- erest rate utilize the Neural Network.The primary section of the Neural network utilize Backpropagation. Their averaged ABS' error for the result of the optimum forecasting is to attain below 5%.