According to the SibFU researchers, the new neural network will be the first "investor assistant" capable of making decisions under the conditions of real market dynamics. Their work has been published in the journal “Applied Mathematics and Computation”.
Linear programming is an optimization tool for mathematical models used in a variety of fields. In finance, for example, the goal of optimization is maximum profit or minimum cost, the scientists explained.
For more than a decade, neural networks have been actively used to solve real-time linear programming problems with changing conditions, the so-called dynamic problems. At the same time, interest in financial optimization with the help of neural networks is also gaining momentum in the world, but no tool for solving dynamic problems has been created for this sphere, experts noted.
Scientists from the Siberian Federal University (SibFU) have improved quadratic and linear programming algorithms based on a neural network approach and fuzzy logic systems to solve dynamic problems, including financial management.
"We have proposed an improved version of the well-known LVI-PDNN neural network method, and specifically, for the first time in the world, its application to dynamic financial problems. With our development, investors will be able to make more accurate decisions. We started with the problem of minimum cost portfolio insurance (MPI)," said SibFU Chief Scientific Officer Predrag Stanimirovic.
According to the creators, the key feature of the new system is a fuzzy logic controller implemented in the LVI-PDNN structure, which operates with degrees of truth instead of the classical real dilemma. This reportedly increases the adaptability of the system in solving dynamic problems.
"In order for a neural network to be able to recognize real-world objects or situations, it must be trained using various optimization algorithms. We are now developing a new class of zeroing neural networks (ZNN) that are capable of solving optimization problems on their own. Such neural network optimizers can be implemented in hardware, in the form of chips, which will make them extremely fast in the future," Professor Alena Stupina, head of the SibFU Department of Economic and Information Technology Management, explained.
The study involved experts from the National and Kapodistrian University of Athens (Greece), University of Nis (Serbia), Swansea University (UK) and Jiangnan University (PRC).
The research team's next goal is to train neural networks to solve other financial problems so that they can fully predict the behavior of investment portfolios.