Comparison of time series forecasting with artificial neural network and statistical approach

Abstract

Mendel. Brun., 2011, LIX, No. 2, pp. 347-352 In this paper we concentrate on prediction of future values based on the past course of a variable. Traditionally this problem is solved using statistical analysis -fi rst a time-series model is constructed and then statistical prediction algorithms are applied to it in order to obtain future values. The time series modelling is a very powerful method, but it requires knowledge or discovery of initial conditions when constructing the model. The experiment described in this paper consists of a comparison of results computed by Multi-layer perceptron network with diff erent learning algorithms previously published and results computed with diff erent types of ARMA models. For the network confi guration an analytical approach has been applied through the cross-validation method. We performed an exact comparison of both approaches on real-world data set. Results of two types of artifi cial neural network learning algorithms are compared with two algorithms of statistical prediction of future values. The experiment results are later discussed from several diff erent points. First the comparison is focused on output precision of both approaches. The comparison consists of matching neural networks results and real values on few steps of prediction. Then the results of ARMA models are compared with real values and conclusion is made. The conclusion also includes theoretical and practical recommendations.artifi cial neural networks, time series forecasting, statistical approach, comparison studyThe aim of the article is comparison of real data forecast using statistical methods and artifi cial neural network (ANN). Traditionally, diff erent methods of statistical analysis are fi xed part of the decision process in the economical resolution (Husek, 2007) (Meloun, Militký, 2004). An artifi cial neural network off ers a kind of intelligent automation of the decision process where the variants of solutions are prepared. Also the principle of ANN is diff erent to statistical models. The ANNs are much less sensitive to input conditions. These points justify the increasing number of successful real-world applications of ANN in business.ANN is one of the Artifi cial Intelligence (AI) methods. These are used to solve tasks where the standard approach is not eff ective or impossible. The main areas of AI applications are forecasting, classifi cation and optimization. Actual applications of ANN include the usage in Management Information systems (Wenlichová, Štencl, 2009), or classification in evaluation of consumer behaviour (Wenlichová, Fejfar, 2010). In a broader application area, the implementation of various machine-learning algorithms as part in development of search engines for geodetic data (as published by Procházka (2010)).The comparison of methods described in this article is related to a previously published paper (Štencl, Šťastný, 2010), which was focused on optimization of the ANN's learning process. Both papers are part of a comparison study focused on forecasting of

How does perception on green environment across generations affect consumer behaviour? A neural network process

“…There are a number of studies that use ANN to model consumer behaviour (Deliana & Rum, ; Stencl, Popelka, & Stastny, ). This model is the fitting model for general purposes and is proven capable in estimating any continuous function.…”

“…There are a number of studies that use ANN to model consumer behaviour (Deliana & Rum, 2017;Stencl, Popelka, & Stastny, 2011).…”

Knowledge discovery on consumers' behaviour

“…The future work will be focused on new methods for regression that represent mainly evolutionary algorithms. Most of the methods found in statistical analysis are not eligible since most of the real-world problems have nonlinear character (Štencl, Popelka, Šťastný, 2011). Linear regression might be used only on short intervals of the measured data or under restricted conditions.…”