本文實證分析臺灣短期利率變動極端行為在開放中、開放後的差異，並建議最適合臺灣商業本票利率的風險值模型。我們應用極值理論模型估計商業本票利率變動分配的尾部，並比較開放中、開放後兩個子樣本期間內，極值模型於計算風險值之實際表現。本文亦建議一種條件極值非參數估計法，此法結合GARCH模型與非參數法，用以估計異質數列的條件機率分配之尾部。實證結果顯示，臺灣商業本票利率的變動分配具有厚尾與不對稱現象，表示常態分配的假設對於風險值的計算並不適當；正式的統計檢定結果顯示，在開放中，條件利率變動分配之右尾較左尾為厚；然兩尾的厚尾程度在開放後並無顯著差異，此外，有更強的證據支持條件利率變動分配的左尾曾發生結構性改變，而右尾發生結構性改變的可能性較低；回溯測試的結果指出，應用非參數法時不能忽略利率資料的相依性與條件異質性；相較於其它模型，結合GARCH模型與非參數法的方式在風險值的預測績效最佳，且表現最為穩健，較不受開放政策的影響；就各信賴水準下的模型表現而言，極值模型在預測高信賴水準（如99％，99.5％）的風險值上，確實有不錯的能力，凸顯其在風險管理領域的重要角色。

We empirically analyze the extreme behavior of short term interest rate changes of Taiwan and suggest the proper models for VaR (Value at risk) estimation. By applying extreme value theory (EVT), we characterize the tails of distributions of changes of Taiwan Commercial Paper rate and evaluate VaR forecasting performances of EVT models. Comparisons are made between a subperiod, from 1992 through 1998, in which the deregulation policies been implemented with a subperiod followed by this. We propose a method that combined GARCH and non-parametric EVT model for estimating the tail index of conditional distribution of heteroscedastic financial time series. The empirical results show that the distributions of changes of rate are fat-tailed and asymmetric, indicating the normality assumption for VaR calculation is inappropriate. According to the results of formal statistical tests, we find that the right tail of conditional distribution is statistically fatter than the left one during deregulation. However, the tail fatness of two tails is insignificantly different after deregulation. In addition, the evidences of structure change after 1998 in left tails are stronger whereas those of right tails are weaker. In backtesting, the conditional non-parametric EVT model outperforms the other. Besides, its forecasting performances show robustness both during and post deregulation as well. These results imply that the dependence and the stochastic nature of the volatilities of time series should be accounted for when applying non-parametric EVT. Moreover, the empirical results also show that conditional EVT models provide more reliable VaR forecasting at a higher confidence level (eg. 99%, 99.5%).