The following is from the August 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.
Q: First of all, thank you for the excellent book! I’m really excited about digging into our own customer data to see what we’ll learn.
A: Thank you for the kind words!
Q: However, when you’re creating the RF Scores, what is the standard timeframe you should use? I have access to about 5 years worth of purchase data – should I create RF scores based on the last 5 years, 3 years, 2 years, 6 months?
Our sales are quite cyclical, so I think the baseline should probably be at least a year, and I’m considering doing two years. It seems as though if I get too much larger than that, my results will be too watered down.
I’m also planning on generating “historical” RF scores by filtering my data to reflect the purchases only up to a certain point. So, to generate a Q1-09 score, I’d create it from sales data of Q1-07 through Q1-09. The Q2-09 score would be from Q2-07 through Q2-09, etc. Does this make sense? It will allow us to see the changes that have been happening in our company even though we’re only just now looking at the data. It will give me a picture of what it would have looked like, had I looked at it back then.
A: I think you have accurately understood the situation and have the right approach! This type of analysis is very sensitive to time frame.
There are really 2 broad types of customer analysis. There is analysis for action in the present, a Tactical approach driving towards a “we should do this now” result, and the more Strategic analysis, which is informational and says “this is what we should have done then” and / or “this is why we should make these business changes”. The shorter time frame is Tactical, the longer timeframe Strategic.
So, for example, a 2 year timeframe could give you the answer to this question: which of our best customers are becoming unlikely to buy from us again? This leads to immediate activation of some kind of marketing outreach or discount / incentive program to get another purchase from this group.
Add a timeframe that ends 4 years ago, then one ending 3 years ago, then one ending 2 years ago could highlight changes in the business over time, for example, best customers with high intent to purchase 3 years ago clustered in certain segments or SIC codes; now customers with this same definition are clustering in different segments or SIC codes. You will see migration of segment focus, if any.
Another way to think about this is time frame for the RF analysis determines sensitivity to new customers. Long time frames tend to rank customers who have been with you a long time higher than new customers; this is just a function of how the ranking methodology works – these long-term customers have had more time to increase the Frequency or Monetary component. This can mask important rankings in Frequency with newer customers, what you might call “future best customers” or “up-and-comers” who are accelerating their purchase behavior. These folks are ideal targets for soft recognition-style rewards (not discounts) – VIP treatment, bonuses, etc.
You could even use this kind of analysis to prove the strengths (or weaknesses) of the RFM methodology for your business: given an RFM score of XXX 3 years ago, what behavior did the customer engage in during the following years? Does the score in one year predict behavior the next year?
Or, perhaps rather than a ranking approach, the fixed activity threshold approach (like LifeCycle Grids) is more appropriate to our business. LifeCycle Grids are basically the same idea as RFM, only sometimes more accurate for businesses with known cyclicality; it’s easier to build that cyclicality into the model if you abandon “ranking” and use thresholds.
In fact, this idea was born from an exercise like the one you propose: let’s re-score and re-rank customers each quarter, and track the RFM score over time. Nothing wrong with this really, except there is the fundamental problem of scores changing due to outside influences, for example, a large new customer campaign.
When such a campaign is executed and then the database is re-scored, the RFM scores of customers can change even if their behavior has not because you are re-ranking a customer file that has changed in composition. Due to the new customer campaign, it is now “heavier” with Recency = 5 customers, which can push down the other customer scores even though behavior has not changed.
This is the primary reason I invented the LifeCycle Grid idea. If you use thresholds or Hurdles for behavioral segments rather than ranking, the “score” of someone does not change when the database composition changes. Someone deemed “best” and likely to buy if R = 30 days and F >= 25 purchases is still “best”, no matter how many records you add to the database. These thresholds define the customer status by putting them in a fixed position box on the Grid, not a ranking.
And that is why RFM tends to be used as the Tactical, “we are doing a campaign right now” valuation method, and LifeCycle Grids tend to be used for the more Strategic analytical exercises. However, the Grids can also be used for Tactical execution.
For example, any customer with F >= 25 over past 2 year period, who drops in R past 90 days, automatically should receive a call from their salesperson. These reports could get run on a weekly basis, and of course can be segmented many different ways depending on the population you run through the Grid. Because you’re using thresholds rather than “ranking”, a customer will appear in the Grid at the same location no matter what the size or segment of the input population.
So for example, you can run only customers who responded to a campaign and see where they end up in terms of Recency and Frequency over time. With a series of such runs, say monthly, you can create a “movie” that shows the evolution of the customers over a time frame and begin to judge the long-term effects of certain campaigns. An example of this approach is here.
Overall, I like the Grid approach much better. Not only do you avoid the “population problem” of ranking when using RFM, but you can use the same approach over and over (good for management understanding) for many different kind of analysis, both Strategic and Tactical depending on needs. You can use all kinds of visual aids such as color in the grid to represent different segments or campaigns, making presentations much easier for management to understand. Decision making with execs can be much more of a challenge when all you have is RFM scores.
All that said, RFM is still probably the easiest approach for specific, usually campaign-related tasks such as predicting campaign response or profitability. Same data but a different, more short-term oriented way to look at the world probably best kept out of the boardroom but still has a place in the analyst’s toolbox.
Hope that helps!
Questions? Does anyone think there is value in predicting which customers will become best customers and which customers are defecting, by campaign source or product line? If you knew this information, could you act on it? Would management care about this?