Quote Originally Posted by Blutiful View Post
You cannot look exclusively at either one or you are not grasping the entire picture.
This. SO MUCH this.

When you're looking at FFLogs in order to make determinations you have to understand that you're not asking a question that has an answer that can be given by one piece of data. You have to look at all the data, and see what the trends are, and decide whether the metric you're looking at is valid to the question you're asking.

So, let's say you're a purple parser (rarely orange) and you'll looking to decide if the job you're taking is holding your group back. Is looking at top speed runs going to help you? No. It's irrelevant to your needs. Is looking at the percentiles people using that job get going to give you information? No, that doesn't tell you anything comparing different groups. (Percentiles of performance are within a job only, not compared between jobs, obviously). Would looking at 99th percentile be helpful? Maybe, if you're actually being realistic when you think you can achieve that performance. But you know what? You probably won't, or don't have to. That's okay. That's not your goal here.

But if you start looking at the rDPS/aDPS spreads (that's all percentiles) favoring the top half of the whisker/box plot? Well now you can see 'Okay, this is where I am, and that's where I'd have to be in those jobs to get improvement. Now you're comparing where you are to where you might be on other jobs--cause a 80percenter on a monk miiiiiight not be an 80percenter on a reaper, after all!

And sometimes you just have to accept the question you're asking doesn't have an answer in fflogs. How do you rank job balance? How does SE rank job balance? Are 1% differences even significant in FFlogs? (tentative answer: probably not)

But the truth is--choosing one data point and saying 'Ah yes, I looked at the top parse and it verifies my thesis' doesn't lead to valid data. You might be right, but you might be wrong. Or you might be asking a question that doesn't count.

FFLogs does a great job, but data interpretation is hard, and the temptation to decide 'this, this one number, right here, that's all the info I need.' It rarely ever is. There's too many confounding factors, there's too much variance, hell, there's too much error margin inherent in FFLogs to come to many of the conclusions people do.