Journal of Economic Modeling Research

This study investigates predictability, chaos analysis, wavelet decomposition and the performance of neural network models in forecasting the return series of the Tehran Stock Exchange Index (TEDPIX). For this purpose, the daily data from April 24, 2009 to May 3, 2012 is used. Results show that TEDPIX series is chaotic and predictable with nonlinear effect. Also, according to obtained inverse of the largest lyapunov exponent, we are able to predict the future values of the series up to 31 days. Besides, our findings suggest that multi-layer feed forward neural network model and fuzzy model based on decomposed data, are of superior performances in predicting the return series. It is worth mentioning that, among these models, MFNN reveals the best performance.

Cigarette and tobacco products in the VAT Law is considered as one of the particular goods and in order to contorlingit’s consumption by price tools, higher tax rates than the standard rate will be levied on it. In this paper, forecasting of revenues of this tax using an approach based on the estimating of tax base has been considered. Thus the first stage, tax base (consumption expenditure) is forecasted for the period 2012 to 2015 and then tax related years by applying the tax rates, will be calculated. In this regard, Because of concerns that policy makers have access to accurate predictions of tax revenues, Supervised neural networks Method to prediction and back-propagation algorithm to train is used. The results indicate that the average annual growth of revenue from value added tax on Cigarette consumption will have 20 percent during the forecasting years.

As one of the important energy forms, natural gas consumption has an upward trend in recent years. Therefore management and planning for provision of it requires prediction of the future consumption. But many of prediction procedures are inherently stochastic therefore it is important to have better knowledge about the robustness of prediction procedures. This paper compares robustness of two prediction procedures Artificial Neural Networks as a nonlinear and ARIMA as a linear model. using resampling method to predict the monthly consumption of natural gas in the household sector. Data spans from 2001-4 to 2012-3, to train the networks, we used genetic algorithms and Particle Swarming Optimization then results were compared using 10-fold method. According to the results, the particle swarm optimization (PSO) outperforms the genetic algorithm. Then we used data from 2001-4 to 2010-3, with resampling by 2000 to predict the  natural gas consumption for the 2001 -4 to 2012-3 and to form critical values. Results show that prediction by a mixed method using ANN and PSO is more robust than ARIMA method.

In capital markets, stock price forecasting is affected by variety of factors such as political and economic condition and behavior of investors. Determining all effective factors and level of their effectiveness on stock market is very challenging even with technical and knowledge-based analysis by experts. Hence, investors have encountered challenge, doubt and fault in order to invest with minimum risk. In order to reduce cost and raise the profit of investment, determining effective factors and suitable time for sailing and purchase is one of the important problems that every shareholder or investor in stock market should consider. To reach this goal, a variety of approaches have been introduced, which are often intelligent, statistical, and hybrid. These approaches are mostly used to predict the stock price time series. Our proposed algorithm is hybrid and involves two stages: preprocessing and predictor. The preprocessing stage involves three steps: missing value, normalization and feature selection. Since there are many features in used datasets, genetic algorithm (GA) is used as the feature selection algorithm. In order to intelligent capability of Fuzzy Neural Network (FNN), this network with two structures (Mamdani and Sugeno) is used as a stock price prediction in second stage. This network is capable of extracting fuzzy rules automatically. Back propagation algorithm (gradient decent) is used for adapting all the parameters. 
Our algorithm is evaluated on ten datasets with seven features obtained from ten different companies. By comparing the simulation results of the simple and hybrid FNN network, we found that the lack of suitable feature selection algorithm will lead to high computational cost, and in many instances the hybrid algorithm outperforms the simple FNN. This results demonstrate the superiority of the hybrid FNN to the simple one. In general, since the number of Sugeno tuning parameters are more than Mamdani, its performance is better than mamdani. Moreover, our algorithm is comparable to the maximum precision rates of other approaches.