機構投資者在現今全球的金融市場中佔有舉足輕重的地位，但是在財務理論的領域裡，他們卻是被極度忽略的一群。本文的第一個部分（見第二章）建構在傳統的期望效用理論之下，進而推導出機構投資者的最適動態資產配置模型。研究發現機構投資者的最適動態資產配置乃是由標竿避險元素與跨期-規模避險元素所共同組成。標竿避險元素述說了機構投資者跟隨標竿投資組合的現象，而跨期-規模避險元素除了為資產配置迷思提供了一個可能的解決之道外，更指出機構投資者會隨著所管理的資產增加而趨於保守。再者，近年來傳統的期望效用理論履遭學者們的質疑，許多實證結果均顯示展望理論更能貼切描述人們的行為模式。本文的第二個部分（見第三章）假設機構投資者的行為模式符合展望理論的公理與假說，進而推導出機構投資者的動態資產配置模型。研究發現當機構投資人處於獲利的狀態之下時，其最適動態資產配置和第二章所得到的結果完全相同，但是，當機構投資人處於損失的狀態下時，他會變得比較積極，持有的風險性資產會大於處於獲利狀態之下時所做的決策。雖然行為財務學已行之有年，但是大家對於損失趨避係數對資產配置所造成的影響所卻知極為有限，本文在此提供了一個參考的模型。本研究發現，損失趨避係數對動態資產配置的影響力會被風險趨避係數、個別投資人對機構投資者績效的敏感度以及機構投資者本身所收取的管理費所抵消掉。此外，近年來金融市場巨幅震盪的現象履見不鮮，本文的最後一個部份（見第四章）假設機構投資者的行為模式符合展望理論的公理與假說，進而在跳躍模式下推導出機構投資者的動態資產配置模型。研究發現在跳躍模式下機構投資者的最適動態資產配置乃是由標竿避險元素、跨期-規模避險元素與跳躍避險元素所共同組成。這個新的元素－「跳躍避險元素」，用以描述機構投資者在面對 跳躍模式所帶來的不同衝擊時所產生的不同回應。本研究發現即使面對相同的投資環境，機構投資者仍然會因為本身所處的狀態不同而有不一樣的投資決策，這個結果迥異於傳統的理論模型，是一個相當有趣且值得進一步研究的議題。此外，本研究還發現損失趨避係數在不同的狀況之下會分別發揮不同的影響力，對損失趨避係數在財務理論上的意義提供了另一個新的視野。 / Institutional investors do matter in financial market, but most of the studies on institutional investors have not determined holdings of different assets by institutional investors. Institutional investors who receive payments and deposits from their customers but they are also subject to withdrawals from them. Compared with individual investors, institutional investors do bear the extra risk that evokes from individual investors. Appling dynamic programming approach, we derive the optimal dynamic asset allocation of institutional investors. In chapter 2, we find that the optimal dynamic asset allocation of the institutional investor with exponential utility function contains two components: the benchmark hedge component and the intertemporal-size hedge component. The benchmark hedge component indicates that the institutional investor takes care of the volatility of benchmark portfolio. The intertemporal-size hedge component provides a possible solution to asset allocation puzzle and depicts that the position of risky assets held by the institutional investor is inversively proportional with its total net managed assets. In chapter 3, we take operating cost into account and find that the optimal dynamic asset allocation of the institutional investor with revised value function will hold more risky assets when she is facing losses, and the sensitivity of loss aversion to dynamic asset allocation strategy is inversively proportional with the absolute risk aversion coefficient, the sensitivity of flow to performance, and the management fee charged by the institutional investor. In chapter 4, we consider both the operating cost and the risk of a sudden large shock to security price into account and find that the optimal dynamic asset allocation of the institutional investor has a further component than that in chapter 3. The further component is labeled "jumps hedge component". Besides, the optimal dynamic asset allocation is divided into four situations that figure the institutional investor with different status quo will make different investment decision. It is a very surprisingly result. Furthermore, we find a very interesting phenomenon that the loss aversion coefficient plays different roles in different situations.

增強式學習（Reinforcement Learning）透過與環境不斷的互動來學習，以達到極大化每一期報酬的總和的目標，廣泛被運用於多期的決策過程。基於這些特性，增強式學習可以應用於建立需不斷動態調整投資組合配置比例的動態資產配置策略。 本研究應用Deep Q-Learning演算法建立動態資產配置策略，研究如何在每期不同的環境狀態之下，找出最佳的配置權重。採用2007年7月2日至2017年6月30日的美國中大型股的股票ETF及投資等級的債券ETF建立投資組合，以其日報酬率資料進行訓練，並與買進持有策略及固定比例投資策略比較績效，檢視深度增強式學習在動態資產配置適用性。 / Reinforcement learning learns by interacting with the environment continuously, in order to achieve the target of maximizing the sum of each return. It has been used to solve multi-period decision making problem broadly. Because of these characteristics, reinforcement learning can be applied to build the strategies of dynamic asset allocation which keep reallocating the mix of portfolio consistently. In this study, we apply deep Q-Learning algorithm to build the strategies of dynamic asset allocation. Studying how to find the optimal weights in the different environment. We use Large-Cap, Mid-Cap ETFs and investment-grade bond ETFs in the U.S. to build up the portfolio. We train the model with the data of daily return, and then we measure its performance by comparing with buy-and-hold and constant-mix strategy to check the fitness of deep Q-Learning.