6 TH INTERNATIONAL CONFERENCE ON APPLICATION OF INFORMATION AND COMMUNICATION TECHNOLOGY AND STATISTICS IN ECONOMY AND EDUCATION (ICAICTSEE – 2016), DECEMBER 2-3 RD , 2016, UNWE, SOFIA, BULGARIA 1 Forecasting Bulgarian Unemployment Rates Using STL in R Software Alexander Naidenov 1 , 1 PhD, Statistics and Econometrics Department, UNWE, Sofia, Bulgaria, email: anaydenov@unwe.bg Abstract. The paper presents the results from the experimental estimation and forecasting of the Bulgarian unemployment rates based on the seasonal and trend decomposition using the locally weighted scatterplot smoothing regression (STL). The reasons for the usage of local subsets of values for the curve fitting instead of a general time series model are discussed and a comparison between models is provided. Keywords. Unemployment rates, time series, STL, LOESS, local regression. 1. Introduction When there is a need for a time series analysis of some economic data (e.g. unemployment rates), especially for the purpose of forecasting, statisticians usually try to find the best model in order to predict the future in the best possible way. Because of the non-linear nature of the economic phenomena in the most of the cases the estimated model is too complex and even though it does not fully represent the data analyzed. But what if we change the point of view and estimate the regression curve by fitting a function of the independent variable(s) locally and in a moving fashion i.e. in the moving average ‘style’ [1]. Although the idea about local fitting is not a new one, the discussion has been started by Woolhouse at the end of the 19 th century [6], Cleveland [1] made some very important extensions to the existing theory which concern the weighting of the cases neighboring given estimated value in the local fitting procedure [5]. Also Cleveland [4] suggests a thoroughgoing time series analysis including not only the locally weighted scatter-plot smoother procedure (LOESS) but also the seasonal and trend decomposition (STL) for even better forecasting estimations. In the next paragraphs we explain the theory behind the STL methodology and we apply the latter to forecast one of the key macroeconomic indicators - the unemployment rate, using the number of employed and unemployed persons as input data. In order to prove the efficiency of the methodology considered, comparisons of the results from STL and some other ‘famous’ approaches, concerning the global type fitting methods, are provided. 2. Theoretical background According to the Taylor’s theorem [7] any continuous function can be approximated with low-ordered polynomials i.e. instead of using one globally fitted model (using all data from given time series) we can use series of locally fitted functions (using only a portion of the data) or so called local regressions. The latter are part of the parametric localization methods and they are distinguishable from other similarly ‘looking’ methods such as smoothing splines, wavelets and etc. [6] The LOESS type of local regression fitting is convenient in those cases where the data in a given moment of time are dependent on the data