Search

Portfolio optimization plays a significant role in financial decisions. In finance terms, a collection of all stocks or assets held by a public or private institute is known as a portfolio. The portfolio selection problem refers to the optimal distribution of budget on the available stocks such that the expected mean-return is maximized (profit), and the risk is minimized. The factor to measure risk is the variance of the portfolio return, smaller the variance lower will be the risk. This approach was introduced few decades ago by Markowitz's modern portfolio theory. The Markowitz mean-variance portfolio selection (MVPS) is widely considered as a significant investment method. The use of quadratic programming (QP) techniques is one option for addressing the static MVPS problem. The continuous-time MVPS (CTMVPS) problem is defined and studied as a time-varying (TV) quadratic programming (TVQP) problem in this article. Using real-world datasets, the CTMVPS problem is approached by two different TVQP neural network solvers. These solvers are based on the Lagrange multiplier method and are called the zeroing neural network (ZNN) and the linearvariational-inequality primal-dual neural network (LVI-PDNN). The experiment findings illustrate and compare the performances of the ZNN and LVI-PDNN solvers in two various portfolio configurations. According to the numerical experiments, when the portfolio has small dimensions, the ZNN solver outperforms the LVI-PDNN solver in terms of accuracy, however when the portfolio has large dimensions, the contrary is happening. As a result, the efficiency of NN solvers is primarily determined by the portfolio dimensions. Our approach is also verified by these experiments as an excellent alternative to conventional MATLAB methods, i.e., the MATLAB's QP solver "quadprog". To the best of our knowledge, this is an innovative approach that incorporates robust neural network techniques to provide an online, thus more realistic, solution to the CTMVPS problem. In this way, we present an online solution to a TV financial problem while eliminating static method limitations. It is worth mentioning that an online solution consists of consecutive solutions with the feature that the previous solution is used as an initial input instead of a random input at each solution's iterative process. Hence, the online solution of a time-varying financial problem is a great technical analysis tool as well as an important financial analysis tool. In addition, to promote and contend the outcomes of this research, we created a MATLAB repository for the interested user, that are publicly accessible on GitHub. [ABSTRACT FROM AUTHOR]

Since the turn of the century, banks and other financial institutions have been granting loans. Given that credit risk emerges mostly when borrowers are unable or unwilling to pay, rigorous background screening of a customer prior to approval of a loan failing is an essential necessity to sustain oneself in such a business. It is worth noting that the amount of non-performing loans in the economy is significant because these loans weigh on bank profits and use valuable resources, limiting banks' ability to grant new loans. Problems in the banking sector can swiftly spread to other sections of the economy, jeopardizing employment and economic growth. As a result, there is an urgent need to develop better models for determining whether or not to grant a loan. This paper introduces a 3-layer feed-forward neural network model for binary classification problems that is trained using a novel bio-inspired weights-andstructure-determination (WASD) algorithm dubbed BI-WASD. The BI-WASD algorithm integrates a meta-heuristic optimization algorithm called beetle antennae search (BAS) for discovering the optimal weights and structure of the BI-WASD based neural network (BI-WASDNN). It is worth mentioning that WASD-based neural networks, i.e. neural networks trained by a WASD algorithm, are known to resolve the shortcomings of traditional back-propagation neural networks such as slow training speed and local minimum. This is due to the WASD algorithm's use of the weights-direct-determination (WDD) method to directly determine the optimal linking weights between the hidden and output layers while also acquiring the optimal number of hidden-layer neurons. Because BAS can solve combinatorial optimization problems, the BI-WASD algorithm outperforms traditional WASD algorithms when it comes to obtaining the optimal weights for the neural network. It is worth mentioning that the BI-WASD algorithm not only obtains the optimal hidden-layer weights directly, but also determines the optimal BI-WASDNN structure automatically and efficiently. Another improvement over traditional WASD algorithms is that the BI-WASD algorithm uses cross-validation to address bias and avoid being stuck in local optima during the training process. Applications on two loan approval classification datasets validate our BI-WASDNN model in order to demonstrate its outstanding learning and predicting performance. Since these applications use real-world datasets that include strings and missing values, an algorithmic method for preparing data is also suggested to make them manageable from the BI-WASDNN. A comparison of the BI-WASDNN model to five other highperforming neural network models is included, as well as a MATLAB package that is publicly available through GitHub to support and promote the findings of this research. [ABSTRACT FROM AUTHOR]