پيش بيني توليد ناخالص داخلي فصلي: سري هاي زماني در برابر شبكه هاي عصبي-فازي

چكيده انگليسي :

Forecasting Seasonal Gross Domestic Product Time Series versus Fuzzy neural Networks Mohsen Damroudi m damroudi@in iut ac ir Date of Submission 2011 05 01 Department of Industrial Systems Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language FarsiSupervisor Akbar Tavakoli atavakoli@cc iut ac irAbstract Gross Domestic Product GDP is one of the most important economic attributes among private andgovernment sectors Long and short term forecasts of such variable can be a basis for selection of appropriatefiscal policies and investments It is apparent that the more difference between forecasts and reality exists themore inefficient policies can become Since economic activities are mostly seasonal and have seasonalbehaviors analyzing them using seasonal data can be more efficient and accurate Of course utilizing highfrequency data has been an active field of research recently Recent progresses in quantitative models specifically in the field of forecasting has made basic changesin these models so that better results with less deviations can be achieved In this thesis it is tried to proposea short term forecasting system to forecast seasonal nominal and seasonal real GDP We hope to take a smallpart in progression of our national economy growth There are several forecasting methods which can be grouped into classical econometric techniques andmodern artificial intelligence methods In order to make a more reasonable comparison two methods werechosen from these groups one from each group In fact in this research two methods SARIMA and ANFIShave been used based on Box Jenkins and Jang methodologies For SARIMA method at first the relevant time series are being made stationary and then the maximumorder of AR MA SAR and SMA terms are calculated based on correlogram analysis At the end amongstappropriate models the best model which has made the superior forecasts has been selected It is noteworthyto mention that in order to make stationarity process correlogram analysis and Ng Perron unit root tests havebeen applied Moreover three normality Breusch Godfrey s autocorrelation and heteroscedasticity ARCHtests have been utilized for residuals Two inputs which are lags of own series have been used as inputmatrices in ANFIS models For this purpose 28 different matrices are generated with the maximum lag of 8 For different models of ANFIS 8 predetermined functions from MATLAB have been used In order to createfuzzy rules grid partition is used with the maximum 4 membership functions assigned to each input Threevarious pre processings which are Min Max Z score and Sigmoidal have been implemented for inputs ofnetwork After modeling and forecasting nominal and real seasonal GDPs quarter on quarter and also year on year growth rates were also calculated and their related errors were computed In order to calculate thedeviations two indexes MAPE and RMSE were implemented It is noteworthy to mention that forecastsbeing done are for maximum 8 seasons on a forward basis More accurately the results are calculated for 1 2 8 seasons on a forward manner but 8 steps forecasts have been considered and analyzed more than theothers Computational results for forecasting quarterly nominal GDP quarterly real GDP quarter on quartergrowth rate and year on year growth rate entirely demonstrate that considering MAPE index ANFISmethod on comparison to SARIMA is much more superior Minimum improvements in above four states are 33 in 7 steps of forecasting 59 in 8 steps of forecasting 48 in 7 steps of forecasting and 62 in 8steps of forecasting Keywords Forecasting Gross Domestic Product GDP ANFIS SARIMA Seasonal Data

دامرودي، محسن

پيش بيني توليد ناخالص داخلي فصلي: سري هاي زماني در برابر شبكه هاي عصبي-فازي

ْGDP , ANFIS , ُSARIMA , داده هاي فصلي