Mediumterm load forecasting, using recursive time - series prediction strategy with Support Vector Machines (SVMs) is presented in this paper. The forecasting is performed for electrical maximum daily load for the period of one month. The data considered for forecasting consist of half hour daily loads and daily average temperatures for period of one year. An analysis of available data was performed and the most adequate set of features for our model are chosen. For evaluation of prediction accuracy we used data obtained from electricity load forecasting competition on the EUNITE network. Some drawn conclusions from the results are that the temperature significantly affects on load demand, but absence of future temperature information can be overcome with time - series concept. Also, it was shown that size and structure of the training set for SVM may significantly affect the accuracy of load forecasting.