Experimental check of algorithm GAP [1] was carried out on daily published rates of various currencies on the Moscow Interbank Currency Stock exchange. The data for the period per 545 days on 5 currencies were taken. The relations of exchange to each other were calculated and the values these cross - rates for one day forward were predicted. The experiments have confirmed expediency of use dispersion criterion for an estimation of expected accuracy of the forecast.           Was found out, that most competent stamms concentrate in immediate proximity to day of prognosis. For the same day at the forecast of different cross-rates it appear competent predictors, produced different stamms.      At the best adjustment of parameters of algorithm the mistake of the forecast for 90 days has appeared equal 2.37 %. The marks of change cross - rates are predicted with accuracy 67 %. For comparison the mistakes of forecasting on "naive" method were defined, at which was considered, that the values of rates tomorrow will be same, as today. The mistake in an estimation of rates in this case was equal 4,01 %, and the mark was guessed correctly in 54.3 % of cases. The literature: 1.Zagoruiko N.G. Selfteaching Genetic Algorithm for prognosis  (GAP).// Artificial Intelligence and Expert Systems. N160, Novosibirsk, 1997. p. 80-95.  2. Zagoruiko N.G. Metakriterium for selection of predicates in algorithms of forecasting. Proc. of 3-rd Siberian Congress on applied and industrial mathematics (INPRIM-98). Novosibirsk, 1998, Part IV, p. 95-96.