It is demonstrated that sampling distribution of correlation coefficientdeforms when there is autocorrelation in time series. In correlations ofsamples to the population, effects of autocorrelation and the time-lag cross-correlation are found to be superimposed on the sampling variances. Asshown both in theory and practice, correlations calculated for a series ofan arbitrary time period do not conform with those of random samples.Thus the correlation coefficients calculated for series with autocorrelationin current practices would contain higher variances than for the real ran-dom samples. It is also found that when the correlation coefficients ofsliding samples display a sine wave pattern, a U-shape probability distri-bution results, which again demonstrates the effect of autocorrelation. Animproved sampling procedure is suggested to obviate the effect of autocor-relation.