Welcome to the GitHub outpost of the recent Predictive Analytics Times "Churn 202" article by Pasha Roberts from Talent Analytics.
If you are impatient and just want to play with the graphs, you might want to just load the program in RStudio and go:
source("churn202.R")manipSim202()- Read more in the interactive model section.
As before, this program uses the fairly common ggplot2, gridExtra, and scales libraries.
To enable live interaction in the manipSim202() function, the system uses the RStudio-only manipulate library.
Manipulate is an interesting, cross-platform, powerful way to easily add sliders and other controls to R programs.
My main criticism is that it doesn't smoothly update the way that D3 does, but see below for my thoughts on issues with bringing this to the web.
To load the program, type source("churn202.R") .
If it works, it will say nothing, but you'll see code load into the variables section in the top right panel.
Default values are stored in a global-variable list named def.
To change defaults, just change the values in the source code and re-source it.
Or, change the variable during runtime.
max.benefit= year at which employee benefit reaches 100% (default 0.5)cost.ramp= slope of cost function (default 3.5)cost.scale= size of initial cost function (default 2)salary= level of employee salary (default 0.5)
These variables use the standard "shape" and "scale" formulations of the Weibull distribution. The shape tends to make the curve more or less skewed, and the scale tends to change its horizontal size.
shape.good= shape of "good fit" distribution (default 2.5)scale.good= scale of "good fit" distribution (default 1.5)shape.bad= shape of "bad fit" distribution (default 1.66)scale.bad= scale of "bad fit" distribution (default 0.33)good.bad.ratio= percentage of employees who are "good fit" (default 0.6)
max.yrs= number of years that show up on the chart (default 3)col.benefit= color of benefit linescol.cost= color of cost linescol.good= color of good linescol.bad= color of bad linescol.be= color of breakeven linescol.be.cume= color of cumulative breakeven lines
There is no way to save variables from a simulation into new defaults. This would be nice - currently I admit to writing them down. If you want to code a way to do this, please send me a pull request.
To output the histograms run runHistograms() .
The histograms seen in figures 1-3 show a sample of 1000 employees in "good fit" and "bad fit" roles.
The distribution is based from a Weibull distribution, as configured in the default variables.
Note the parameters to change the sample size, annotation, graph grobbing, and more.
To output the four plots describing the static model at the default values, run runSim202() .
This does not require RStudio.
The plots are:
Figure 4 shows the above Weibull distributions as configured by shape.good, scale.good, shape.bad, and scale.bad. Orange is "bad fit," and Blue is "good fit."
Figure 5 shows the benefit-cost curve and cutoff lines, as derived in Churn 201, and configured by max.benefit, cost.ramp, cost.scale, and salary. These are the "quantitative scissors."
Figure 6 shows the Cumulative Net Benefit curve and cutoff lines, as configured by max.benefit, cost.ramp, cost.scale, and salary.
Figure 7 shows the Expected Cumulative Net Benefit curves for "good fit" and "bad fit", plus the value of the EVH in text. These curves and EVH are derived from all above variables as well as good.bad.ratio.
Note the other parameters to change annotation, graph grobbing, and more.
Every time the model is run, it prints text to the console, such as:
Daily breakeven at 0.28, cume breakeven at 0.76
Good Fit: 48.1% net benefit * 60% weight = 28.8% overall contribution
Bad Fit: -17.0% net benefit * 40% weight = -6.8% overall contribution
Overall EVH = 22.0%
These mean what they say:
- The year marks at which daily and cumulative breakeven is made
- The net benefit of "good fit" and "bad fit" populations, times their weights
- The overall EVH - our objective function
To output the four plots with fancy sliders, using RStudio, run manipSim202().
If you don't see the slider controls, click on the little gear in the graph area.
The output is exactly as in the static model, (in fact you will see that the program just runs the static model with new parameters), so please refer to that section for tips on modifying the output.
As with the static model, text is output with every run. This way you have a running log of model results as you experiment.
We are looking at options to put this up on the web, with one of the reactive web visualization platforms. Normally we would look at a simpler, snazzier platform like D3, but as you can see in the code there are some rather intense numeric integrations involved. Maybe we leave the EVH calculation aside and show the graphs with D3?
To use R directly, it looks like we could easily wrap this in Shiny, a R-based webvis platform. But, that needs a special host with some restrictions and expensive proprietary software. We're still figuring it out, but it isn't done yet. Maybe by the time we get to Churn 204. Please let us know if you have any ideas.
If you just want numbers, and none of these heavy graphs, no problem.
From the command line, use runPredNetCume() to return the EVH value, overriding any of the defaults for the specific parameters that you want.
> runPredNetCume(good.bad.ratio=0.2)
[1] -0.03978231
To get the verbose output, specify verbose=TRUE:
> runPredNetCume(verbose=TRUE)
Good Fit: 48.1% net benefit * 60% weight = 28.8% overall contribution
Bad Fit: -17.0% net benefit * 40% weight = -6.8% overall contribution
Overall EVH = 22.0%
[1] 0.2204147
> runPredNetCume(verbose=TRUE, good.bad.ratio=0.8)
Good Fit: 48.1% net benefit * 80% weight = 38.4% overall contribution
Bad Fit: -17.0% net benefit * 20% weight = -3.4% overall contribution
Overall EVH = 35.1%
[1] 0.3505133
I tend to write this kind of article by writing programs and numbers first, then the text. In this case, the EVH concept took me well over the normal word count, therefore we threw some concepts overboard until next month.
One concept, already programmed, was sensitivity analysis.
I'll leave it as an exercise to the reader to run runSensitivityTests(), or to just wait until next month.
All I can say, is that computing power has sure come a long way since I was putting punched cards into the mainframe; this is pretty amazing.
As before, I encourage the community to engage with us on GitHub - comment, make issues, send pull requests. We do find this of great value, and hope that we can connect with others who do so as well.
Copyright © 2014, Talent Analytics, Corp. All Rights Reserved.