The GAP algorithm for prediction of dynamic series  includes three stages of the work: formation of the set of  basic elements (basic «cultures»), search for competent cultures and prediction of the fixed element.       Let us consider the table n*m, describing n properties xj (j=1,2…n) in m moments of time (days) ai (i=1,2...m). The most «fresh» data has an index i=1, data for the previous day – i=2 and so on until the day with an index i=m. Among k first elements of the table we will separate G elements and will call them the «basic culture» with the capacity of G. The elements of this culture v(i,j) are marked by indexes of their strings and columns.       If the index i of each element will be increased by the value q, then we will separate the culture of the same capacity and with same «architecture» as the basic one, but shifted in time by q steps into the past. Let us call such culture the «isomorphous» to the basic one. If the order of element observation in basic and isomorphous cultures will be made equal, then it will be possible to find their correlation.       Let us call the «competent» such isomorphous cultures, which provide coefficient of the correlation to the basic culture higher the given value, e.g. kor=0.85.       Let us add element a(0,j) to the basic culture. The new culture of (G+1) capacity is formed, shifting of which by q steps in time separates the culture, isomorphous to it, with an element a(q-1,j). If we will construct the equation of linear regression between G present elements of these cultures and put the element a(q-1,j), then we will be able to obtain value aq as the variant for the prediction for the element a(0,j). Using p competent cultures we will receive p variants of the element a(0,j) prediction. Let us call these variants «the predictors». The competence of predictors may be estimated by the dispersion criterion: the competence of the given set of predictors will be higher, the lower is the dispersion d of their values. Let us fix the average value of predictors a1 and characteristic of their competence h1=1:(1+d1).       For every new t-th basic culture we will find the values of at and ht. Having run over T basic cultures, we can obtain the weight-averaged value of the element to be predicted:                            a(0,j) = Sat *ht / Sht    for all t from 1 to T       According to average competence of predictors from all T cultures it is possible to estimate the expected value of prediction error  by using the dispersion of predictors: the experiments showed, that correlation between prediction error and predictor dispersion may reach the value of 0.7.