As mentioned earlier, an artificial neural network is widely used in various fields of science, including in the financial sector. Consider examples of the use of neural networks in managing various risks:
- methods for assessing investment risks. Here it is necessary to analyze the organizational and economic stability of organizations and the likelihood of forecasting bankruptcies;
- methods for assessing production risks. In this category, you can consider forecasting the required supply of raw materials, as well as optimal production planning;
- assessment of currency risks. An important issue here is forecasting quotes in the foreign exchange market;
- assessment of trade risks. It is necessary to conduct a study of the demand factor, as well as forecasting and price analysis;
- credit risk assessment. In this case, it is worth considering the forecast of credit efficiency, analysis of the borrower's creditworthiness [2, 32].
2.3.1 Methods for assessing investment risks
Methods for assessing investment risks, namely forecasting bankruptcy of an enterprise or organization, were first used in the sixties of the last century. Then such methods as methods of one-dimensional statistics, multidimensional Altman analysis [33], multidimensional discriminant analysis, recursive separating algorithm and so on [34] found application.
It is worth noting that the use of traditional methods of mathematics has its drawbacks associated with the definition and evaluation of the corresponding model. Consider exactly what conditions complicate this choice and evaluation of the best model. First, the complexity of the models used is limited by the evaluation methods used. Which model will be applied often depends on the computational complexity of the algorithm, and not the accuracy of the model.
Secondly, standard estimation methods work when the conditions for the normal distribution of the totality of the source data are met. According to the analysis of literary sources, we can conclude that the financial ratios used in the prediction of insolvency are not distributed according to the Gauss law.
All of the above allows you to take a fresh look at the use of artificial neural networks in risk analysis. The main advantages of intelligent forecasting systems: the ability to obtain the correct solution to the problem in the presence of incomplete and distorted data after setting the network parameters; the ability to take into account a large number of additional factors affecting the quality of forecasting; they are resistant to interference, have high speed.
In matters of building a neural network, it is necessary to develop its topology, as well as determine the learning mechanism and testing procedure. A wide database is of great importance in obtaining the right conclusions. To train the network, you need input data, that is, data with reliable financial statements and coefficients calculated on its basis.
The three-layer perceptron models and backpropagation algorithm have been widely used as a training model [35]. In this case, all network elements build a weighted sum of their inputs, adjusted as a term. Then they pass this activation value through the transfer function, obtaining the output value of this element. These elements are organized in a layered topology where the signal is transmitted directly.
This network can be easily interpreted as an input-output model in which weights and threshold values are free parameters of the model. It is worth noting that the network can simulate a function of almost any degree of complexity.
The number of layers and the number of elements in each layer determine the complexity of the function. When constructing multilayer neural networks, determining the number of intermediate layers and the number of elements in them is one of the important issues. The number of input and output elements is determined by the conditions of the problem.
The literature [3] gives an example of a backpropagation algorithm. This algorithm has its own characteristics. In the backpropagation algorithm, the error surface gradient vector is calculated, which indicates the direction of the shortest descent along the surface from a given point. This method reduces the error. However, difficulties may lie in determining the length of steps [2, 36].
2.3.2 Credit risk assessment
As noted above, credit risks are one of the main risks in the banking sector. Assessing the creditworthiness of a client is precisely the area where neural network applications were the first to find applications. It is worth noting that the use of this method showed a fairly good effect. In this case, when assessing the solvency of customers, the probability of own losses from an untimely refund was determined. When training a neural network, baseline data on previously approved credit histories is of great importance. Having a data on a client base, a neural network will predict with a high degree of accuracy the degree of solvency of the borrower. In practice, each bank has its own valuation methods and its client base.
In this connection, the model based on the neural network for each specific case will be different. When developing a model based on an artificial neural network, it is worth considering a sufficient number of factors.
Often, neural network forecasts are combined with expert assessments, which are represented by a system of requirements presented by a bank to potential borrowers. If the forecast is about eighty or ninety percent, then it is considered successful. Based on the analysis of the application of neural networks in practice, we can conclude that these networks allow us to determine more than ninety percent of potential defaulters
Today, more and more enterprises are intrigued by improving the forecast quality. The use of neural networks to predict the results of loans determines the possibilities of lending to enterprises and the advisability of providing loans and loans without collateral based on an analysis of additional information about the consumer of loans. In this case, a risk assessment is performed based on the construction of a nonlinear model [37].
However, like any other model, models based on a neural network have both advantages and disadvantages. As already noted above, the use of neural networks allows the most efficient way to analyze the financial condition of the company in order to timely identify and eliminate financial risks. The main feature of neural networks is that, unlike expert systems, they, in principle, do not need a previously known model, but build it themselves only on the basis of the presented information.
Consider the most important advantages of neural networks:
- the ability to successfully obtain a result even with incomplete, distorted input information;
- the ability to learn from a variety of examples in those cases where the laws of the development of the situation and the dependence function between the input and output data are unknown;
- The use of a trained neural network is available to any users;
- error tolerance: performance remains when a significant number of neurons are damaged;
- the ability of the network to recognize patterns even with strong interference and distortion.
Neural has found particular use in difficult formalization of the decision process or impossible. Neural network technologies are quite a powerful modeling tool, since they are nonlinear in nature. That is, they have found widespread use in an unknown form of communication between input and output. It should be borne in mind that there is a relationship between input and output parameters. The peculiarity of training a neural network is that the dependence itself will be deduced. This is the reason for the widespread use of artificial neural network algorithms in various industries, including the economy.
Today it is worth noting the widespread use of these networks to solve problems in compiling algorithms for determining the analytical description of the dependencies of the operation of economic objects. It is worth noting that the use of neural network techniques is aimed at solving some of the problems of economic and statistical modeling, as it also improves the adequacy of mathematical models, thus bringing them closer to economic reality [38, 39].
It is well known that having some kind of real or physical model or process, it is difficult to implement its identical one hundred percent mathematical model, taking into account all possible factors. It is worth remembering that when compiling the mathematical model, certain assumptions are used that are accepted for each specific case. Note that economic, financial and social systems are inherently very complex. And in this case, neural network technologies come to the rescue, which can be used in solving many poorly formalized tasks. This may include analysis of financial and banking activities, stock and foreign exchange markets, which are associated with high risk patterns of behavior of borrowers and so on.
As noted above, the use of neural network technologies has found wide application in solving problems in the financial industry.
These problems include: insurance activities of banks; bankruptcy forecasting; application of neural networks to tasks of exchange activity; forecasting the economic efficiency of financing innovative projects; prediction of loan results; customer solvency assessment and many others [37].
But, like any thing in neural network technologies, they also have their drawbacks. It is worth noting that this technology does not have any great experience in application, and in this regard, the question arises of the reliability of the use of such applications. It is worth noting that artificial neural networks can be inaccurate both with a faulty computer and with a working one.
Here the conclusion suggests itself that when solving important problems and issues it is advisable to use these networks not as the main method, but as an additional one. Or in those cases when the solution of certain issues is typically not enough. The next negative point is the fact that neural networks are unable to explain how to solve problems. The learning process itself is produced by a program, for example, a matlab, giving out some final result. It is quite difficult to understand exactly how the presentation of learning outcomes occurs and this leads to the difficulty of analyzing the result [37].