A three parameter time constant learning curve model is applied to four case studies and the optimum parameters determined in the least sum of error-squared sense. The error residuals are then analysed using correlation, regression, and cusum techniques. A global range for learning curve scatter is derived from grouped data. One case study establishes a sinusoidal trend additional to wide-band noise. In all cases studied there is correlation between successive error residuals so that there are many less sign changes than predicted from an assumption of normally distributed, uncorrelated scatter. A simulation model is documented which generates artificial learning curves with the requisite degree of correlation as evident in terms of the number of sign changes. The results of the paper may be used to establish the a priori scatter estimates necessary in various parameter estimation techniques developed for on-line prediction of human operator future performance.