|Abstract: ||本研究應用類神經網路以預測我國公債市場上之未來利率與未來的利率期限結構,故內容可分為兩部分:一是探討我國公債利率期限結構中的資訊內涵問題,並將類神經網路模型與時間數列模型作一比較;二是探討我們是否可以使用多輸出類神經網路架構同時預測不同期限的利率。 我們以Swanson & White的模型選擇標準來評估利率期限結構中的資訊內涵,結果發現雖然帶漂移項的隨機走路模型可為常用的Schwarz資訊與一期預測均方差選擇標準選出而成為最佳預測模型,但是若以混淆指數來看,利率期限結構中所含的遠期利率資訊仍然對未來的現貨利率變動方向預測是有所幫助的。不過,我們的估計結果卻也顯示遠期利率的變化率卻無助於未來現貨利率的變化方向。 至於倒傳遞與輻射基底的類神經網路在混淆指數的表現上有百分之六十以上的成功預測率,似乎意味著類神經網路對預測未來利率方向的能力上應仍是優於隨機走路模型。再者,估計結果也顯示預測愈遙遠的現貨利率,類神經網路愈能抓住殘差項的非線性特質而有助於未來利率走向的預測。不過,類神經網路對未來利率水準的預測能力卻是很差的。 最後,當我們以類神經往網路同時估計不同期限現貨利率的預測結果,雖然模型的平均預測誤差值結果尚稱良好,但是預測誤差的標準差仍嫌過高。此外,以類神經網路同時預測未來現貨利率的方向都不較隨機猜測好到那裡。因此,僅是直覺地使用過去的利率作為預測未來利率的架構顯然是不夠的。|
This research project applies the neural network to predict the future direction of interest rates and the future shape of the term structure of interest rates. It has two parts-one is to investigate the information content of the term structure of interest rates and then to compare the results of the neural model with the time series model, the other is to investigate whether we can use neural model with multiple output to predict various interest rates of the term structure simultaneously. Applying the model selection criteria utilized by Swanson & White (1994), we found that although the random walk with a drift is the best model by Schwarz information criterion and the one-step forecast mean squared error criterion, but in terms of confusion index, the future rates implied by the term structure of interest rates are still useful in forecasting the direction of the future spot rate. Our results, however, also indicate that the change in the future rate does not help to predict the change of the direction of the spot rate. In terms of the backpropagation and radial basis neural network, we have over 60% successful prediction rate. It seems neural network model is better than the random walk model in forecasting the future spot rate. Furthermore, the more distant future of the spot rate is, the better the neural network model is in grasping the nonlinearity of the residuals, and the more helpful the neural network model is in predicting the direction of the future spot rate. However, neural network model is not too good in predicting the level of the future spot rate. Finally, when we use the neural network model to predict various interest rates simultaneously, we have a fair mean forecasting error, but the standard deviation of the forecast error is still too high. Furthermore, the prediction of the direction of the future spot rate is no better than a random forecast. It is, therefore, not enough simply using intuitively appealing structure with previously observed term structure of interest rates to predict the future interest rates both in directions and levels.