05-07-2012 05:16 AM
I want to identify those customers who have reached a certain figure of AUM(assets under management) and hereafter droped below that certain figure through a logistic regression with customers' behavior (transactions) within the 6 past months before his drop adding some individual information like marrital status as well.
General speaking, long time period will garantee a better model.
And finally I want to use this model to predict those with potential dropping possibility customers, for this purpose, the make-decision-period should be less, because when a customer fits the model well for the 6 months, it's very difficult to do the maintain work, the quicker, the higher possibility to keep those AUM.It requires a less make-decision-period.
How can I balance the contradicts? Or modify my solution to better serve the final purpose?
Can any one give a thought?Thanks in advance.
08-15-2017 10:14 AM
I'm not sure I completely understand your modeling situation but I can offer one suggestion that might be of help. When modeling things like churn or future undesirable action (e.g. taking your assets to another broker), it is not necessarily helpful to begin your target period immediately following your observation period. You need a lag period that allows you to intervene and/or take some action to try and stop the anticipated departure.
For example, suppose you have data on customers that left during the September-October time frame during a particular year (using larger target windows like this can be helpful to increase the number of events). Rather than looking at their data up through the end of August, you would consider using a lag period of 1-2 months so that you limit yourself to the data available prior to August. For example, you could try and use the data from January through June to predict a departure in the September-October time frame. This model would provide some lead time (July and August in this example) during which you could take some action to try and retain the client. You might find it works better only to use a one month lag period in which case you could use the January through July data to predict a departure in September/October.
One thing to note: When building your variables, write them as time lagged variables (e.g. use the end of the observation period as lag1 and then number them going backwards so that in the following month, you can update the values based on their relation to the current period. You would then be applying the model fit on January through July and updating the information in the lag variables to reflect information from February through August to predict a departure in October/November. In this way, you can continue to score your data as soon as information becomes available and still have time to discourage the departure.
I hope this helps!