資訊管理學報
柯博昌;田育任;
頁: 203-226
日期: 2008/07
摘要: 類神經網路近幾年被許多學者證明能有效率應用於求解大規模非線性複雜問題，然而應用於有限資源配置(limited resource allocation)問題上並不多見。投資組合資金分配是典型的資源配置問題，傳統投資組合的資金配置方式建立在假設及限制條件下，不符合現實投資環境；被廣泛應用於資金配置問題的遺傳演算法，存在著需要將權重正規化而產生不合理解的問題。現有的類神經網路模型應用於投資組合資產配置決策上，無法最佳化輸出層神經元做為個別資產資金配置比例以及其比例總合為100%。若強制使用正規化方法以符合配置比例總和符合等於1，容易因為某一比例改變，而造成整體配置的相對比例隨之改變，影響訓練結果的振盪與不穩。為解決上述方法應用在資金配置問題，本研究提出配置型類神經網路模型，在無任何假設及限制條件下，求解投資者面臨不確定因素影響及有限自有資金情況下之投資組合個別資產資金配置比例，並且保證配置比例總合為100%。並且考量個別資產預期報酬與風險、個別資產間報酬相互影響關係及投資者本身的風險趨避程度(risk averter)。實證結果發現：配置型類神經網路應用於投資組合資產配置比例問題上，不但能使類神經網路於輸出層之投資組合資產配置比例總合為100%，且能最佳化個別資產配置比例；此外，配置型類神經網路的投資報酬率優於遺傳演算法與台灣加權股價指數。
關鍵字: 類神經網路;有限資源配置;投資組合;資金分配;遺傳演算法;
Abstract: Portfolio management is one of limited resource allocation problem. The primary goal of portfolio is to optimally allocate investor's asset by considering the trade-off between risk and return. The well-known Markowitz mean variance optimization is a quantitative tool which makes the investment in a given number of different assets and limits the amount of capital to be invested in each asset. The artificial neural network (ANN) proposed a nonparametric efficiency measurement approach with nonlinear capability to solve large-scale complex problem effectively. However, the traditional ANN model cannot guarantee that the summation of produced investment weights always preserves 100% in the output layer. This article introduces an allocated neural network model to optimize the investment weight of portfolio, which will dynamically adjust the investment weight as a basis of 100% of summing all of asset weights in the portfolio. In addition, Genetic Algorithm (GA) is another common evolutionary computation technique used to optimize portfolio management. The experimental results demonstrate the feasibility of optimal investment weights and superiority of ROI compared with GA and benchmark TSE (Taiwan Stock Exchange).
Keywords: Neural network;limited resource allocation;portfolio;asset allocation;genetic algorithm;
瀏覽次數: 11749 下載次數: 154
引用 導入Endnote
頁: 203-226
日期: 2008/07
摘要: 類神經網路近幾年被許多學者證明能有效率應用於求解大規模非線性複雜問題，然而應用於有限資源配置(limited resource allocation)問題上並不多見。投資組合資金分配是典型的資源配置問題，傳統投資組合的資金配置方式建立在假設及限制條件下，不符合現實投資環境；被廣泛應用於資金配置問題的遺傳演算法，存在著需要將權重正規化而產生不合理解的問題。現有的類神經網路模型應用於投資組合資產配置決策上，無法最佳化輸出層神經元做為個別資產資金配置比例以及其比例總合為100%。若強制使用正規化方法以符合配置比例總和符合等於1，容易因為某一比例改變，而造成整體配置的相對比例隨之改變，影響訓練結果的振盪與不穩。為解決上述方法應用在資金配置問題，本研究提出配置型類神經網路模型，在無任何假設及限制條件下，求解投資者面臨不確定因素影響及有限自有資金情況下之投資組合個別資產資金配置比例，並且保證配置比例總合為100%。並且考量個別資產預期報酬與風險、個別資產間報酬相互影響關係及投資者本身的風險趨避程度(risk averter)。實證結果發現：配置型類神經網路應用於投資組合資產配置比例問題上，不但能使類神經網路於輸出層之投資組合資產配置比例總合為100%，且能最佳化個別資產配置比例；此外，配置型類神經網路的投資報酬率優於遺傳演算法與台灣加權股價指數。
關鍵字: 類神經網路;有限資源配置;投資組合;資金分配;遺傳演算法;
A Novel Neural Network Model for Portfolio Optimization
Abstract: Portfolio management is one of limited resource allocation problem. The primary goal of portfolio is to optimally allocate investor's asset by considering the trade-off between risk and return. The well-known Markowitz mean variance optimization is a quantitative tool which makes the investment in a given number of different assets and limits the amount of capital to be invested in each asset. The artificial neural network (ANN) proposed a nonparametric efficiency measurement approach with nonlinear capability to solve large-scale complex problem effectively. However, the traditional ANN model cannot guarantee that the summation of produced investment weights always preserves 100% in the output layer. This article introduces an allocated neural network model to optimize the investment weight of portfolio, which will dynamically adjust the investment weight as a basis of 100% of summing all of asset weights in the portfolio. In addition, Genetic Algorithm (GA) is another common evolutionary computation technique used to optimize portfolio management. The experimental results demonstrate the feasibility of optimal investment weights and superiority of ROI compared with GA and benchmark TSE (Taiwan Stock Exchange).
Keywords: Neural network;limited resource allocation;portfolio;asset allocation;genetic algorithm;
瀏覽次數: 11749 下載次數: 154
引用 導入Endnote