Standardizing scores from fraud prevention systems I have two scoring systems for fraud prevention, their qualities are as follows:
system A scores 0-1000, where 1000 is the best score, and 0 is the worst, The system A set has a std deviation of 135.58 and a mean of 809.21
System B which scores from .1-99 with .1 being the best and 99 the worst.
The system B set has a std deviation 3.17 and a mean of .40
The sample sets are scoring the same information, I have 130,000 data points.
I'm trying to standardize the scores, so I can compare how the two systems think about the same information.  I started with Z-score, which seems correct, should I be accounting for the sample size in my calculations?  When I come to the correct Z scores for each data point, if I subtract one from the other will that give me a good comparison of how close they are, and thus the degree to which the two systems agree with one another?
 A: I would start by reversing the order on one of the two scores.
Perhaps let $B^\prime$ be $100 - B,$ so that high scores are best in both
systems.
If both A and B scores are close to normal, then standardizing should work fine. However, depending on how the scores are constructed, there is no guarantee of normality. If not normal, then comparing the two systems becomes more complicated. To assess normality, I would start by making normal probability (Q-Q) plots of data from each system of scores. 
If they are not normal, you may prefer to refer to the quantile (percent below) of a score. If there are very many cases in which
A and B result in different quantiles for individual data points,
you will have to decide whether A or B scores are more useful in practice,
perhaps digging a bit into the methods and criteria used in each
system.
Finally, I would want to know the correlation between A and B for
the 130,000 data points at hand. (Pearson correlation if they are roughly
linearly related, Spearman correlation or some other kind of rank correlation, if not linear.) You believe A and B are scoring the same information,
but I would want to know whether they are scoring information in functionally comparable ways.
