La première partie de cette thèse introduit une nouvelle méthodologie pour la réalisation d’exercices de stress-tests. Notre approche permet de considérer des scénarios de stress beaucoup plus riches qu’en pratique, qui évaluent l’impact d’une modification de la distribution statistique des facteurs influençant les prix d’actifs, pas uniquement les conséquences d’une réalisation particulière de ces facteurs, et prennent en compte la réaction du gestionnaire de portefeuille au choc. La deuxième partie de la thèse est consacrée à la valorisation des obligations à maturité très longues (supérieure à 10 ans). La modélisation de la volatilité des taux de très long terme est un défi, notamment du fait des contraintes posées par l’absence d’opportunités d’arbitrage, et la plupart des modèles de taux d’intérêt en absence d’opportunités d’arbitrage impliquent un taux limite (de maturité infinie) constant. Le deuxième chapitre étudie la compatibilité du facteur "niveau", dont les variations ont un impact uniforme sur l’ensemble des taux modélisés, a fortiori les plus longs, avec l’absence d’opportunités d’arbitrage. Nous introduisons dans le troisième chapitre une nouvelle classe de modèle de taux d’intérêt, sans opportunités d’arbitrage, où le taux limite est stochastique, dont nous présentons les propriétés empiriques sur une base de données de prix d’obligations du Trésor américain. / In the first part of this thesis, we introduce a new methodology for stress-test exercises. Our approach allows to consider richer stress-test exercises, which assess the impact of a modification of the whole distribution of asset prices’ factors, rather than focusing as the common practices on a single realization of these factors, and take into account the potential reaction to the shock of the portfolio manager.
 The second part of the thesis is devoted to the pricing of bonds with very long-term time-to-maturity (more than ten years). Modeling the volatility of very long-term rates is a challenge, due to the constraints put by no-arbitrage assumption. As a consequence, most of the no-arbitrage term structure models assume a constant limiting rate (of infinite maturity). The second chapter investigates the compatibility of the so-called "level" factor, whose variations have a uniform impact on the modeled yield curve, with the no-arbitrage assumptions. We introduce in the third chapter a new class of arbitrage-free term structure factor models, which allows the limiting rate to be stochastic, and present its empirical properties on a dataset of US T-Bonds.

台灣早期有關系統風險（β）的研究皆指出β不能解釋台灣股票報酬的變異，故控尋更能解釋股票報酬的風險因素為本文的主要目的之一。 本研究分析民國71年7月至85年5月股票上市公司資料（排除金融、保險、及變更交易方式的公司）。因民國79年股價指數從2月的最高點12,495急遽下滑至10月的2,560，故分析上將79年度予以排除。在71年7月至78年12月的時段中，整體市場因素（RM-RF）不能解釋股票報酬的變異。此點發現與台灣早期研究的結論一致。其他變數顯著者僅有與規模有關的因素（SZSMB），或與負債比率有關的因素（DEHML），其中以 SZSMB的解釋能力最強。在民國80年1月至85年5月的時段中，所有模式中整體市場因素（ RM-RF）的係數皆顯著，並且是所有因素中最顯著者。這點發現與前時段（71年7月至78年12月）的結果有很大的不同。其他的變數顯著者，有代表成長機會的BMHML（與淨值市價比有關的因素）、EPHML（與益本比有關的因素）、或CPHML（與營運現金市價比有關的因素），及代表利率結構有關的風險因素TERM（與利率期間結構有關的風險溢酬）、或DFT（與利率違約風險有關的風險溢酬）。其中以（RM-RF）、EPHML、CPHML及TERM的風險組合最能解釋股票報酬的變異。 應用更完整的股票報酬解釋變數，探討股票報酬與盈餘間的關係，亦為本文主要目的之一。經分析以(1)各時段最能解釋股票報酬的因素組合為基礎，計算異常報酬；(2)單獨的以整體市場因素（RM-RF）為基礎計算異常報酬，然後再分別估出盈餘反應比較係數（ERC）比較之。結果顯示，以各時段最能顯著解釋股票報酬的因素組合為基礎的ERC為正的顯著，且其ERC大於只以整體市場因素（RM-RF）為基礎所算出的ERC。 另外，關於盈餘品質假說之測試，經以公司規模大小為虛擬變數放入迴歸式中，結果顯示，代表大公司的虛擬變數之係數時而為正，時而為負，且都不顯著，故盈餘品質假說未獲得支持。 再者，關於成長機會與ERC關係之測試，經以公司成長機會大小為虛擬變數放迴歸式中，結果顯示，代表成長機會的虛擬變數之系數時而為正，時而為負，且大都不顯著，故成長機會大的公司之ERC大於成長機會小的公司之ERC的假說，未獲得實證的支持。 / Earlier studies (Chen 1990; Chiu 1990; and Wang 1992) found that systematic risk (β) could not explain the variance of stock returns in Taiwan. The findings were inconsistent with the Capital Asset Pricing Model (CAPM). One of the major purposes of this paper is to examine the factors that have higher explanatory power of stock returns. To test the hypotheses, this study uses the data of Taiwanese listed companies covering the period from July 1982 to may 1996. The 1990 data are excluded because the stock market index climbed to a record high of 12,495 in February 1990 and then fell sharply to allow level of 2,560 in October 1990. The "crash" might cause structural changes in stock market, so the analyses are conducted separately for the periods before and after the crash, namely the prior-crash period (from July 1982 to December 1989) and the post-crash period (from January 1991 to May 1996). The empirical results show that for the prior-crash period the overall market factor (market returns minus risk free rate, RM-RF) can not explain the variance of stock returns. The findings are consistent with those of previous studies. However, we find that the factor-related to size (SZSMB) and the factor related to debt/equity ratio (DEHML) have significant association with stock returns. Furthermore, SZSMB has higher explanatory power. In contrast, the overall market factor is the most significant factor for the post-crash period. Other factors that are significant consisted of (1) proxies for growth opportunities, including book-to-market equity (BMHML), earnings/price ratio (EPHML), and cash flow/price ratio (CPHML), and (2) the factors related to interest structure, including term structure (TERM) and default risk (DFT). Among these factors, the set of RM-RF, EPHML, CPHML, and TERM explains the variance of stock returns most. Another purpose of this paper is to use the aforementioned findings to study the relationship between stock returns and earnings. The results show that the earnings response coefficients based on the most explanatory factor portfolio of each period are positive and significant, and are greater than those based on the traditional systematic risk (β). The tests for earnings quality hypothesis indicate that the coefficients of the dummy variable proxies for big companies are insignificant. The earnings quality hypothesis is not supported. The tests regarding the relationship between growth opportunities and earnings response coefficients show that the coefficients of the dummy variable proxies for high growth companies are unstable. The hypothesis that the earnings response coefficients of high growth companies are greater than those of low growth companies is not supported by empirical evidence.