Statistic model for removing bad data Let's say I have a (ads live days, revenue) data set. 
The data set shows how much revenues each ads generates during the days it is live. 


*

*ads1 generates 100 dollars during the 5 days when it is live. 

*ads2 generates 200 dollars during the 10 days when it is live.  

*ads3 generates 300 dollars during the 15 days when it is live.


... 
so I have (100,5) , (200,10) , (300,15) ..... 
Imagine I have 100 ads in the data set and I can only launch 80 of them next month. 
I have to decide which ads perform better so I can continue launching next month. What is the statistic model?  
I was thinking using linear regression, to keep the nodes that have less variance .... but I am not sure if that is correct.
 A: Why do you need any statistics? Statistical inference lets you generalize, but no generalization is needed here (if I understand you correctly). 
What I'd do if I were you: Compute the dollars/day for each ad and pick the ones with the highest values. 
A: I am not sure to understand what you want to do because it does not seem to be a statistical problem to me. 
With the only two variables you have, I guess the performance can be calculated as revenue/nb_of_days. And you'll choose the 80 more performing adds. 
Why would you do a regression model ? What would be the predicted variable ? What would be the other variables ? 
I guess your problem is more difficult than that but you need to provide more information.
A: This is not a statistics problem. It is a dynamic optimization problem. You want to maximize your ad revenue, subject to the constraint that you only have X to spend and the placement of each ad costs Y.
Moreover, I think you need to revisit the assumption that the ads that are profitable in one month will continue to be profitable subsequent months. Aside from a few outliers, ads lose their effectiveness over time - which means that this is another constraint in your optimization problem.
If this is a real-life problem and not just a mental exercise, your best bet is to hire someone to do it. Kaggle, Innocentive, Elance are just some companies that have a database of reputable data scientist freelancers who can do this for you. (I don't represent any of the companies listed, and with some searching I think you can find others as well.)
