D.V. Yashin – Post-graduate Student, Ulyanovsk State Technical University
E.N. Egov – Assistant, Ulyanovsk State Technical University
There are many methods for predicting time series. Each method has its advantages and disadvantages. At the same time, the problem of choosing the best forecasting method for a particular time series is of considerable complexity. Combined hybrid prediction models allow you to obtain a final forecast based on the results of several methods. This can significantly reduce the risk of choosing a non-optimal method.
When using a combined hybrid model, the problem of choosing the best method is transformed into the problem of selecting a subset of the best methods from the base set of methods. This problem can be solved by an expert or by dividing a time series into a training and control part with subsequent forecasting of control values on the training data and estimation of the prediction error.
Obviously, the effectiveness of a particular method depends to a large extent on the characteristics of the predicted time series. This statement underlies the method proposed in this paper for selecting individual prediction models from the base set. In this approach, machine learning methods are used, namely a specially developed neural network that selects methods according to the values of the metrics corresponding to the characteristics of the time series that are relevant for the subject area.
In this paper, we consider the main characteristics of the time series that are significant for solving the forecasting problem. Also, the criteria used to evaluate these characteristics are considered. Formulas for assessing the degree of expression of characteristics ac-cording to the criteria are proposed.
The configuration of a neural network for selecting prediction methods from a base set of a combined model is proposed. The input values of the neural network are time series metrics. Each method from the base set corresponds to one of the output layer's neurons. The neural network learns to calculate the estimated value of the prediction error for each method from the base set. The software implementation of this network uses the language R with the built-in package «neuralnet».
In determining the optimal set of time series metrics, the greatest difficulty is the balance between the full coverage of the characteristics of the time series and the requirement of a low degree of correlation between the characteristics for the correct functioning of the neural network. The criteria corresponding to the same characteristic of the time series are often based on different regularities. Therefore, the values of the different criteria for one characteristic do not always have a high degree of correlation between themselves. In this paper, a series of computational experiments was performed on the basis of which an optimal set of metrics was determined.
The effectiveness of the proposed solution was tested on a time series from the competition «Computational Intelligence in Forecasting» (CIF) 2015−2016 . Using only the prediction methods chosen by the neural network in calculating the aggregated forecast allowed to reduce the total error from 13.131% to 9.234%.
- Computational Intelligence in Forecasting (CIF). URL = http://irafm.osu.cz/cif/main.php (ot 26.04.2018).
- Yarushkina N.G., Filippov A.A., Romanov A.A., Moshkin V.S., Egov E.N. Developing a system for time series data mining on the basis of F transform and the domain-specific ontology // Joint 17th world congress of international fuzzy systems association and 9th international conference on soft computing and intelligent systems (IFSA-SCIS). 2017. Institute of Electrical and Electronics Engineers Inc. 2071.
- Afanas’ev V.N., Yuzbashev M.M. Analiz vremenny’x ryadov i prognozirovanie. M.: Finansy’ i statistika. 2001. 228 s.
- Vasil’ev A.A. Genezis gibridny’x modelej prognozirovaniya na osnove ob’‘edineniya prognozov // Vestnik TvGU. Ser. E’konomika i upravlenie. 2014. № 23. S. 316−331.
- Gorelik N.A., Frenkel’ A.A. Statisticheskie problemy’ e’konomicheskogo prognozirovaniya // Statisticheskie metody’ analiza e’konomicheskoj dinamiki: Uch. zap. po statistike. M.: Nauka. 1983. T. 46. S. 9−48.
- Kobzar’ A.I. Prikladnaya matematicheskaya statistika. Dlya inzhenerov i nauchny’x rabotnikov. M.: Fizmatlit. 2006. 816 s.
- Lukashin Yu.P. Adaptivny’e metody’ kratkosrochnogo prognozirovaniya vremenny’x ryadov. M.: Finansy’ i statistika. 2003. 416 s.
- Sadovnikova N.A., Shmojlova R.A. Analiz VR i prognozirovanie . M.: MGUE’SI. 2001. 67 s.
- Timina I.A., E’gov E.N. E’ntropiya kak pokazatel’ anomalij vremenny’x ryadov metrik proektnoj deyatel’nosti / Nechetkie sistemy’ i myagkie vy’chisleniya. Promy’shlenny’e primeneniya // Sb. nauchny’x trudov IV Vseros. nauchno-prakticheskoj mul’tikonferenczii s mezhdunar. uchastiem «Prikladny’e informaczionny’e sistemy’ (PIS). 2017. Ul’yanovsk: Ul’yanovskij gosudarstvenny’j texnicheskij universitet. 2017. S. 47−53.