# Attrition Forecasting

I am currently trying to develop a forecast for monthly subscriber attrition that allows me to predict for a future point in time, how many subscribers I have. I have a couple of years worth of attrition data, and every month we're recruiting new subscribers.

The difficulty I am having is that if I am wanting to predict 18 months into the future for subscribers recruited this month I am having to rely on data from subscribers recruited 18 months ago and use their attrition curve to estimate how many will be around in 18 months time.

I have noticed that the month 1 attrition for new subscribers is slightly higher than it was 18 months ago. What I am wanting to know is: Is there a way I can use the most recent months attrition data as an input to my attrition curve, rather than relying on data that is 18 months old already?

I am currently doing this in excel and need to continue in excel for the benefit of my colleagues, however I am open to reworking things in R as I suspect it will allow me greater flexibility and reproduce-ability in the future.

Thanks in advance for the guidance

UPDATE: 17/06/2015 Below is a screenshot (data available here http://pasted.co/d9d6fc5e) showing what my data looks like at the moment. The area highlighted in green is what I am using to create my attrition curve (essentially at least 18 months old - this example is older) as it allows 18 months for recruits to defect. I am wondering if there is a way for me to use more recent information as an input into my attrition curve (like that data 201411 for month 1 attrition and data from 201410 for 2 month attrition).

• I may be able to help you if you post the sample data even at the aggregate level. – forecaster Jun 16 '15 at 23:16
• Even fictitious example would be helpful. – forecaster Jun 17 '15 at 1:32
• Fantastic I'll provide a solution in the next few days – forecaster Jun 17 '15 at 1:40
• Data of that form is common in some areas (effectively an "age-period-cohort" type structure). The short answer is yes, you can use the more recent data, but appropriate models depend on whether there's "diagonal" effects or not, and account needs to be taken of the fact that if there are effects in all three directions, you'll have an identifiability issue to deal with. However your data also seem to be essentially discretized survival data, which potentially narrows the scope of the models you might look at. – Glen_b Jun 17 '15 at 6:03
• See some relevant discussion here – Glen_b Jun 24 '15 at 3:15