| |
Dissertation
A modern economy experiences the new phase of the development. It is conditioned a few factors. At first, by introduction in science of modern mathematical methods. Secondly, appearance of the newest computer technologies, doing possible research of the difficult phenomena and processes.
The unpredictability of markets of equities, unexpected gallops of prices and incomprehensible trends, and also sudden falling is experienced an economy as heavy crises. Accordingly, that, who possesses the best methods of extraction of conformities to the law from noisy, chaotic time series – the least subject to influencing of similar crises in an economy, and, accordingly, can hope on a large income.
A prediction of market temporal rows is a necessary element of any investment activity. Self concept of investment – based on the idea of forecasting of the future.
Forecasting - one of claimed, but here one of the most intricate problems of intellectual analysis of data. The problems of prognostication are related to insufficient quality and amount of basic data, by the changes of environment a process flows in which, influence of human factors. A prognosis is always carried out with some error which depends on the utilized model of prognosis and plenitude of basic data. [1]
A task of forecasting of temporal rows was and remains actual, especially lately, when the powerful tools of collection and treatment of information are accessible to steel. Wide application for the decision of this task was got by neural networks. It is conditioned a presence in the financial time series of difficult conformities to the law, not discovered linear methods.
Initially neural networks targeted at recognition of structural patterns. Appearance, consisting of set of visual, semantic or other properties, is demonstrated in such tasks of network, and a network must recognize entrance appearance, as belonging to one or a few classes. [2]
In opposition it at prognostication of temporal rows appearances which change in time are processed. A prognosis depends, generally speaking, not only from current values but also from all of previous values of the forecast size. In such situation success of forecasting above all things depends on the method of forming of teaching selection. [3]
At such formulation a neural network is utilized as an universal tool of approximation of discrete function, teaching selection set sets. [3]
The basic problem of the use of neural networks is a necessity of forming of teaching selection, that not always possibly, for lack of formalized model of the probed object.
On this stage of development of knowledge’s plenty of methods, allowing to conduct an analysis, pre-treatment and forming of teaching selection accumulated about neural networks.
By another problem of application of neural networks to the task of forecasting of market of equities, consists in absence of method of selection of methods of pre-treatment of teaching selection, giving’s the best prognosis.
The decision of the problem indicated higher is possible with the use of evolutional algorithms. This idea actively develops as it applies to the questions of realization of the multivalent systems, adaptive conduct and artificial life, evolutional robotics.
By a research purpose conducted in this work there is development of subsystem, based on application of neural networks, allowing automatically to pick up the methods of pretreatment of teaching selection, with aim receipts of the most high-quality prognosis.
For achievement of the put purpose the successive decision of the followings tasks is needed:
global analysis of neuronet approach is in the task of forecasting, and also formulation on the basis of results of analysis of requirements to the applied neural network; global analysis of methods of pretreatment and forming of teaching selection; global analysis of evolutional algorithms, and choice of genetic statements as it applies to the decided task; approbation of the developed algorithms and analysis of the got results.
The supposed scientific novelty of the results got as a result of work will consist in the capacity of adaptation of subsystem for time series, due to possibility automatically to choose the method of pre-treatment and forming of teaching selection, for the receipt of the most high-quality prognosis.
research of modern methods of prognostication, based on neural networks; selection and analysis of methods of pre-treatment and forming of teaching selection, and also program development for their realization; development of subsystem allowing to make a market of equities forecast;
Features of application of neural networks
Features of application of neural networks which show their advantages as compared to other existent methods at the choice of model. [1]
Effectiveness at a decision normalized or badly formalized tasks. From well-known advantages of methods on the basis of neural networks it is necessary to select one most attractive is absence of necessity for the strict mathematical specification of model, that especially valuably at prognostication of badly formalized processes. It is known that most financial, business and other similar tasks formalize badly. Stability to the frequent changes of environment. Dignities of neural networks become noticeable, when the "rules of game change often": environment, which the forecast process, and also character of influence of influences, is in. Therefore, neural networks by the best appearance befit for the decisions of such tasks, as prognostication of fund market tendencies, characterized influencing of whole set of constantly changing factors. Effectiveness during work with the large volume of contradictory information. Neural networks will be preferably wherein there is very much a lot of the analyzed information conformities to the law are hidden in which. In this case different nonlinear co-operations are automatically taken into account also between influences. It is special important, in particular, for a preliminary analysis or selection of basic data, exposure of "falling out facts" or flagrant errors at making a decision. Effectiveness during work with incomplete information. The use of neural networks is expedient in tasks with incomplete or by "noisy" information, and also in tasks which intuitional decisions are characteristic for.
Limitations and failings, linked with the use of neural networks for forecasting [1]:
For effective prognostication, as a rule, some minimum of supervisions is needed (more than fifty and even to one hundred). However there is a lot of tasks, when such amount of statistical information is inaccessible. For example, at the production of seasonal commodity, statisticians of previous seasons it is not enough for a prognosis on a current season from the change of style of product, policy of sales and etc even at forecasting of requirements in a stable enough commodity on the basis of information about monthly sales it is impossible to accumulate statistics for period from 50 to 100 months. For seasonal processes this problem is yet more expressed: every season is one supervision actually. It should be noted that the satisfactory model of prognosis with the use of neuron network however can be built even in the conditions of shortage of information. Thus a model will be specified at a receipt in it fresh information. Other lack of models on the basis of neural networks are considerable temporal expenses for achievement of satisfactory result. This problem is not so substantial, if the small number of temporal sequences is probed, however much the usually forecasting system in area of management of operations includes to a few thousand temporal sequences from a few hundreds. We will mark that the overpriced expectations of effect from introduction of neuron networks in the row of financial structures in the USA and Great Britain did not justify oneself. So, one large investment bank on Ulsterite expended more than 1,0 million dollars in development of such system for optimization of financial operations, however, after some time forced was to go back to the old system. Principal reason of failure was become by insufficient as compared to expected level of the productivity, got as a result of introduction of the system. To teach and exploit a neural network for the decision of many tasks, as a rule, a not specialist can, but reliably to interpret results, and also specialists, having skills in the design of neuron networks, are numeral able to estimate meaningfulness of the got prognoses.
Method of forecasting of financial time series
on the basis of vehicle of neural networks.
Forecasting of time series is a calculation of size of its future values or descriptions, allowing to define this size, on the basis of analysis of the known values. A size subject a prognosis is named forecast a size.
It is assumed at prognostication, that the value of the forecast size depends on some factors, will name their determinatives, or signs. One of approaches to the task of prognostication based on supposition of dependence of the forecast size from the previous values of time series. The theoretical ground of such approach is a theorem of Takensa.[4]
The chart of decision of task of prognostication of temporal row can be presented as the following sequence of the stages:
Stage of preliminary transformations. Stage of structural synthesis of neural network. Parametric synthesis of neural network. Verification of error of prognosis on a control selection.
If a value of error of prognosis on the last stage will be in possible limits, a task can it will be considered decided, and the trained neural network can be utilized for the receipt of prognosis. It is otherwise necessary to repeat the first three stages. Application of neural network is shown on a picture 1.
 Picture 1 – Application of neural network.
Stage of preliminary transformations.
One of the most important stages in the decision of task of neuronet forecasting is forming of teaching selection. Exactly from composition, plenitude, qualities of teaching selection substantially depend time of teaching of neuron network and authenticity of the got information.
Necessary stages at forming of teaching selection:
- Encoding of input-output.
- Rate fixing of information.
- Preprocessing of data.
Encoding of input-output. In a difference from ordinary computers, able to process any character information, neuronets algorithms work only with numbers, because their work is based on the arithmetic operations of addition and increase. However much not every entrance or output variable in an initial kind can have numeral expression, consequently, all of similar variables must be coded – to translate in a numerical form, before to begin neuronet treatment actually.
Rate fixing and pre-processing of data. Entrance and output variables of neuronets can be quite different sizes. Thus, the results of neuronet design must not depend on units of measuring of these sizes. Namely, that a network interpreted their values uniformly, all of entrance and output variables must be resulted to single – to single – to the scale.
Let as a result of translation of all of necessary variables in a numerical form and subsequent rate fixing all of entrance and output variables are resulted to the single scale. Basic task of neuronet analysis – to find statistically reliable dependences between entrance and output variables. The unique information generator for conducting of analysis are examples from a teaching selection. What anymore than bats of information will bring every example – so much the better the present are utilized our order information.
Often, beforehand unknown, on how many one or another entrance variables are useful for the prediction of values of returns, there is temptation to increase the number of entry parameters, in a hope on that a network will define what from them most meaningful. However, complication of teaching of neural networks quickly increases with growth of number of entrances. Yet more important, that with the increase of amount of entrances exactness of predictions is worsened. Thus, the number of entrances must be hardly limited, therefore the choice of the most informing entrance variables is presented by the important stage of preparation of information for teaching of neural network.
Stage of structural synthesis of neural network.
This stage is implied by the choice of architecture of neurons and structure of connections between neurons.
Uniting in a network, neurons are formed by the systems treatments of information, which provide effective adaptation of model to the permanent changes from the side of external environment. In the process of functioning of network there is transformation of entrance vector of signals to weekend. The concrete type of transformation is determined as by architecture of neural network so by descriptions of neuron elements, management tools and synchronization of informative threads between neurons. The important factor of efficiency of network is establishment of optimum amount of neurons and types of connections between them.
Parametric synthesis of neural network.
On this stage, teaching of neural network is made.
In the task of prognostication neural networks are utilized with the controlled teaching, where a current return is constantly compared to the desired return. Libra at first is set by chance, but during next iteration corrected for achievement of accordance between the desired and current return. The created methods of teaching are aimed at minimization of current errors of all of neurons, created the continuous change of scales for achievement of acceptable exactness of network.
Teaching is considered complete under reaching by the neural network of certain level of exactness, when the desired values of returns are produced for the set sequence of entrance signals. After teaching of weight of connections fixed for further application. Some types of networks allow to utilize the continuous teaching for adaptation to the changing terms.
A task of forecasting of time series at the market of equities was and remains actual. Application of neuron networks to this task allows to attain enough good results. It is conditioned a presence in most financial time series of difficult conformities to the law, not discovered other, for example by linear methods.
As well as any other type of analysis, an analysis with the use of neural networks in the prediction of financial time series requires the enough difficult, careful pre-processing of data. One of the most important stages in the decision of task of neuronets forecasting is forming of teaching selection. Exactly from composition, plenitude, qualities of teaching selection authenticity of the got information depends substantially.
This master's degree work is devoted the use of neural networks, for the decision of task of forecasting of financial time series and to development of programmatic tools for realization of this forecasting.
Круг П.Г. Нейронные сети и нейрокомпьютеры: Учебное пособие по курсу «Микропроцессоры». - М.: Издательство МЭИ, 2002. – 176с. Тарасенко Р.А., Крисилов В.А. Предварительная оценка качества обучающей выборки для нейронных сетей в задачах прогнозирования временных рядов \\ Труды Одесского политехнического университета – 2001 Вып.1 (13) – с. 90-93 Тарасенко Р.А., Крисилов В.А. Выбор размера описания ситуации при формировании обучающей выборки для нейронных сетей в задачах прогнозирования временных рядов. \\ Труды Одесского политехнического университета – 2001 Вып.1 Ежов А.А., Шумский С.А. Нейрокомпьютинг и его применения в экономике и бизнесе. – М.: МИФИ, 1998. – 222с. Тарасенко Р.А. Метод анализа и повышения качества обучающих выборок нейронных сетей для прогнозирования временных рядов. // Диссертация на соискание ученой степени кандидата техническихнаук.-ОНПУ-2002. Уоссермен Ф. Нейрокомпьютерная техника. М.: Мир, 1992.
|
|
