The paper discusses several important methodology problems relating to risk evaluation in finance. By analysing historical data in finance it is not possible to obtain results as precise as in controlled physical experiments. Moreover, the statistical methods developed for the sake of natural sciences are not always suitable for economic applications. A number of problems arise: that of ensuring the invariability of the reference category (or the homogeneity of the reality under analysis); the problem of adequate determination of the sample and population; that of ensuring a random selection of sample elements; and finally, of maintaining the stability of initial conditions. Without resolving the above issues, attempts to evaluate risk on the basis of historical data may be fraught with errors.