I think you mean 1-1/e (~= 1 - (1-10^-6)^(10^6)). But yes, the idea that you can just divide a statistic sampled over thousands of events linearly to get a meaningful statistic for one occurrence of the event is ridiculous.