Does somebody knows or has experieced how to convert (for a single site) rejected traffic or congestion… etc., into revenue?
I wank to estimate what is lost due to the congestion.
Apparently this seems a simple question. But indeed it’s worthy of a full fledge data science project.
For revenue, we need the number of calls times duration of call.
Since, the call never went through, how can we estimate the duration of call.
There’s no easy way to estimate it.
One way would be to do it at subscribers level. Check out whose calls got rejected, then checkout what’s the average( or median or mode??) duration of his call. That would come from Core.
To add further complexity, we must mind that often the subscriber are enrolled to different packages, and also the particular hour you want to do calculation maybe a special offer hour etc.
Consequently, the easy way to go is to contact your IN. Check out on average how much money that particular subscriber spends when he/she calls (in that particular hour??)
To recap; so far we have seen that it’s not easy to calculate the duration of the call which got rejected, and even using the duration of call is meaningless due to various call packages.
So, apparently we can use the average value of the bill for each subscriber and sum it for all the subscribers who called and got their calls rejected.
To further fine tune your results, you may also do pattern analysis for each subscriber: since for each subscriber, there are particular hours when he/she makes long calls and there are particular hours when he/she makes short calls. Consequently, using average value of bill from IN is not logical. We have to create a model for each subscriber that tells that in this particular hour, if his call had gone through; he would have talked for this much amount of time and his revenue contribution would have been this much.
In reality, doing such calculation for each cell and repeating for every other cell for the network is overkill. It’s better to create a traffic model of the complete network and make clusters based on revenue and then calculate the difference from real value of revenue to find the loss.
To create a model, one needs to do PCA and combine RF KPIs with values from IN. Because subscribers tend to call more and longer when KPIS are good (take this as null hypothesis).
One more thing which needs to be accounted for is that how many Rejections will be converted into calls.
For example a subscriber tried to call and his call rejected and he tried again to make the same call and rejected again.
In doing so he made 10 attempts. Which actually was going to be one call only.
It’s a very complex calculation.