Signal detection modelling

[Question]

Hello All,

I have a built an engagement model using Logistic regression. in SAS My approach was to identify the customers my company considers engaged, and then I built a model (using both engaged and not engaged customers) and predict "engaged"..  but the most important variables are 3 months variables (time spent on site in the last 3 months, distinct devices used in the last 3 months, etc...).. I have been challenged by the business who would like to have a Dynamic engagement model, basically they want to see the scores changing quite fast when a dramatic change happens (Price change, Special Event , Xmas etc). 

At the moment, it will take the model I have , around 3 months to detect a change, which is a long period and won't be useful for the business. Could you please advise what's the approach to take in this situation? Are there any other techniques I could use here? Is Signal detection modelling , the right approach? Should I add extra variables ( dropped or increase m1 vs m2)?

Your help would be much appreciated.

Thank You

[ssoti2001]

Hi,

Interesting problem.

My take on this (based on my understanding of your problem). You can employ a technique like Shewart chart to determine the behaviour of these variables over time. You can then use time series models to extrapolate these variables into the future and also use shewarts on the forecast result. The forecast will help against the 3 months constraints you have i.e. you  can forecast into the future (assuming these variables are quantitative) and you can then determine behaviour of the forecast (as with historic) in addition you have residual analysis (Shewart on the forecast model residual, MAPE etc).

http://support.sas.com/documentation/cdl/en/qcug/63922/HTML/default/viewer.htm#qcug_shewhart_a000000...

http://support.sas.com/documentation/cdl/en/etsug/66840/HTML/default/viewer.htm#etsug_arima_syntax02...

Hope the helps.

Best

Seun Sotuminu

[AnujaNagpal]

Hi,

I have recently worked on a project where I scored the customer on their engagement with the company. My approach was a bit different from yours so it might give you some ideas on the alternate approach.

I derived the information on who is engaged and who is not based on the time spent ( in my case number of days "case ticket" was open). In other words number of days "case ticket" was open was my dependent variable.

In your case, "time spent on a website in last 3 months" would be the target and other variables like distinct devices used in the last 3 months, Price change, Special Event, Xmas etc would be input variables and you can model expected time on the website.

Now you can identify those customers whose predicted value is lower than expected and those would be "low engaged" and higher than expected would be "high engaged".

Hope that helps.