The following is from the May 2009 Drilling Down Newsletter. Got a question about Customer Measurement, Management, Valuation, Retention, Loyalty, Defection? Just ask your question. Also, feel free to leave a comment.
Want to see the answers to previous questions? Here’s the blog archive; the pre-blog newsletter archives are here.
Q: First of all thank you for your help. I have some questions I would be pleased if you answer them for me.
A: No problem!
Q: 1. RFM analysis – is it possible to use some other ranking technique rather than quintiles? Using quintiles for bigger databases will cause many tied values, isn’t it a problem?
A: Sure, you can use it any way it works best for you. There is no “magic” behind quintiles, you can use deciles or whatever works best. It’s the idea of ranking by Recency, Frequency, and Value that is the key concept in the model.
I’ve seen dozens and perhaps hundreds of variations on the core RFM model, depending on how you classify a “variation”. One change that’s common is changing the scaling, as you mention above, to accommodate the size of the database. Smaller databases use quartiles or even tertiles. Larger databases, choose the ordered distribution that meets the need.
A more common modification is to convert “M” to different types of “value” depending on the business model. Instead of Sales, people fine-tune the financial side by using Net Sales, or Gross Margin, net out discounts, etc. Or they use non-sales representations of value tuned to the business model – ad revenue per visit, total days of activity, that kind of thing.
Further, what can happen is the analyst or marketer will begin to see patterns underlying the RFM cells – in sales, response, location, merchandise, source, or some other customer variable. This leads to cross-tabbing RFM score with other variables, and discoveries are made which lead to customized versions of the RFM model.
For the most part, I envision this work really as segmentation, meaning the scoring is not really modified – it’s the population the scoring is run on that is modified. So for example, you run separate RFM scores for customers who are primarily hard goods buyers versus primarily soft goods buyers. This approach to scoring is sometimes referred to as RFM-C, where C = category.
Or for large, ongoing campaigns, you can cross-tab RFM score by source of the customer. This leads to “weighting” the value of campaigns not by Sales or Response, but the long-term profitability of the customer – you see campaign sources “clustering” in high or low RFM scores. Some campaigns generate weak customer profiles, but the volume justifies doing them, as long as they are kept “reigned in”. Other campaigns generate high value profiles who are “slow starters”, and might be killed if you only looked at Response and not RFM Score. So the scores begin to play more of a role as a “standard” way to view customer value across categories, campaigns, channels, etc.
This approach to scoring can eliminate a lot of the “gut feel” legacies that can happen in marketing and merchandising. Sure, go with your gut, but let’s use a standard way to compare the results of your gut feel and produce a “gut check” comparison.
Q: 2. I am planning to add user complaints and suggestions to RFM analysis. Each complaint will decrease the user score and then cause to organize promotions just for users who had a complaint recently. Is it a good approach to add it to RFM analysis? (I am not sure but some are using this method.)
A: I’m not exactly sure I know what you mean by “add”, but I think I get the gist of what you’re trying to accomplish. In fact, this project sounds like an example of a company actually trying to “do something” about customer engagement and experience instead of the usual navel-gazing. I have done these kinds of “apology campaigns” before and they can be very profitable, especially for most valuable or highly engaged customers.
The scores only are predictive on a single behavior being scored, so I would not involve 2 different behaviors (purchase and complaint) in the same score, since the result would be defeating to the purpose of the score. I would not “adjust” a score directly based on a different behavior; I would score this behavior separately – and then use the scores in tandem to make adjustments in execution. If you really want to use multiple behaviors simultaneously in a model, you need to move up the modeling food chain to regression.
As an analyst, you can of course “add” to the RFM scores any way you wish. You can add any characteristic as a “tag” to a score but I would not involve these characteristics in the scoring itself, unless they *are* the score. But from the perspective of a Marketing person who has to use the scoring, I would not want you to “corrupt” the scores themselves, but rather to segment by other variables and then examine and use the scores to act.
For example, if these complaints are in the customer account, you could score the customers on some other behavior such as purchases and include the RFM score in an account field, then cross-tab score to complaints. For example, “Give me every customer with a high RFM score AND at least 2 complaints”. Or lever off the complaints, “Of customers with at least 2 complaints, what are their RFM scores?”
Or, as discussed above, the complaint idea is an opportunity to create a custom RFM-style score for complaints. Recency and Frequency are still important, but there is no Monetary Value. Time frame may also be different for complaints than purchases, for example, past 30 days or past 3 months as opposed to a full year or longer. You could generate this “RF” score and then use it in combination with the RFM score to drive different messaging to people by both:
1. How engaged they are in some behavior
2. Intensity and level (overall Frequency) of complaints, where the more Recent a complaint has been made, the more likely it needs to be addressed in some way.
Customers with high scores in both areas would be both most valuable to the company in the future AND at highest risk for defection. This is, of course, an extremely valuable target from a Marketing perspective and one that should be addressed with great care. Sending these people “normal” e-mail communications, for example, is much more likely to accelerate to defection than retain the customer.
Depending on your business model, for these highly valuable and likely to defect customers, you might want to skip e-mail or snail mail and get the President of the company to phone them!
2 thoughts on “Hacking the RFM Model”
Jim – great blog. I also purchased your book and read it on vacation (believe it or not). I’m looking to apply these concepts into our products which will allow our clients to rank their customers and be more efficient with email marketing…thanks again
P.S. I’ve linked to you on my blog! Great content.
Good article here. While retailers end up spending large amounts creating and executing elaborate loyalty programs and promotional campaigns, the response may be disappointing. Accurate knowledge of customers who are more likely to buy the second time can help retailers focus their efforts on this concentrated target group and derive higher benefits from promotions and loyalty marketing. This can be done through RFM model retail analytics. More info here: http://thoughts.manthansystems.com/rfmmodel_business_analytics.php