Predicting the future is done using the Messages section of an operator to the last segment of the message, regardless of whether the operator of the algorithm of mathematical calculations or electrical and mechanical devices implemented. In this regard, we have found that perfect predictor, which we first saw, exposed to two opposite types of error. Originally designed by us predicting the device could be made such that it foresaw a very smooth curve with any degree of accuracy. However, improved accuracy is achieved at the cost of increasing the sensitivity of the device. The better was the instrument to smooth the signal, the more he is caused to oscillate by small disturbances of smoothness and the longer had such fluctuations. Thus, the extrapolation good smooth curve appears to require a more accurate and sensitive device than the best possible prediction nonsmooth curve; in each case, the device would depend on the choice of the statistical nature of the predicted event. One might think that these two types of interdependent errors have something in common with conflicting tasks measuring the position and the amount of [c.53] motion considered in quantum mechanics, Heisenberg, in accordance with its principle neopredelennosti.Posle we have understood that the solution to the problem of optimal prediction just look at the statistics of the prediction of time series can be obtained, it was easy to turn what at first be difficult to predict the theory into an effective means of solving the prediction problem. Taking certain statistics for time series, it is possible to find an explicit expression for the mean square prediction error in this method, and forward at this time. And placing this expression, we can reduce the problem to finding the optimum prediction definite operator in which to become a positive minimum value depending on the statement. Tasks of this type are solved at least in well-developed branch of mathematics - the calculus of variations, and the industry has a well-developed technique. With this technique we were able to get an explicit best solution to the problem of predicting the future length of the time series based on its statistical properties and, moreover, find the physical implementation of this decision by the actual devices.