the main point here that this is not really prediction. Say after my calculations I know, that the traffic I have bought should bring me ROI 100% in a trhee monthes. I calculate this on the base of three previous monthes. But in a month I make the calculations again and understand that the traffic will bring me the ROI 100% in 2.5 monthes, so I correct my traffic accumulations coeficients. I do it month by month, so I have pretty much precise numbers. Ofcourse affiliates are mixing uo the traffic, thats doesnt matter, all your calculations are based on previous stats, so if they mix the traffic they probably did it before.
I understand what you're saying, and if aggregate data by traffic type is as deep as you want to go, that works just fine. The data is pretty smooth when you look at it all together in groups like that, and fraud/low quality does tend to stay at least a little bit stable.
What I'm talking about is increasing overall efficiency (and by extension, payouts to good, stable affiliates) by going deeper and identifying the non-performing and/or fraudulent affiliates to get them out of the equation to the greatest extent possible. Sometimes it can be done; sometimes it can't.
Suppose you know that affiliates doing media buys on various banner networks are, on average, giving you a positve ROI of xx%. That's great, and you can definitely choose to base your terms and payouts on that average figure without giving it much more thought (and only kicking people out for obvious fraud). However, if you think of each affiliate as a separate asset and evaluate them individually wherever justified by volume, you may be able to cut out the bottom 10%, 20%, whatever. By removing those with the lowest ROI, you've increased your overall ROI for that type of traffic.
So - if you find that the AVERAGE affiliate doing large media buys generates 20% of his payouts in sales during the first month, but you have a couple who are only generating 5% of their payouts in sales...you want to find out a couple things:
1 - Is the difference statistically significant?
2 - Is there anything different about the leads that can help us draw conclusions about the long-term profitability of this affiliate's leads?
3 - Is there anything about the way the affiliate is promoting the offer that's different from other affiliates using similar methods? Can we tell?
Consider a not uncommon scenario - 2 affiliates have a low back-end conversion rate in the first month. Both have similar traffic sources, but one is promoting the service honestly while the other promises free sexy videos and an unlimited free trial (vs. a more typical free trial where certain features are restricted). Also, the second affiliate is using inappropriate imagery the implies a different kind of dating site. Similarly, the second affiliate could be someone sending a mixture of fraud and real leads. The problem, though, is that you often have incomplete information and you don't necessarily KNOW what's going on with the ads, the way leads are generated, etc.
With affiliate 1 (the honest guy), you're likely to see more users creating full profiles, more logins, deeper session depth, etc. With affiliate 2, you're more likely to see a lot of people who log in just once and realize it's not what they had hoped for.
Of course, that's all complicated by the fact that some scammers will actively instruct their fake leads to fill in more than what's required. Sometimes they do a good job of making it look natural, other times they don't. I saw a guy one time who sent more than a thousand leads almost overnight, all using the password "mangoo".
You certainly don't HAVE to scrutinize things this closely, but considering the amount you can gain by removing even the obvious bottom 5% in a given category (BEFORE they run for months and months) can save tens of thousands of dollars or more, it's often worth the effort.
