Customers Selection with RFM Skander Esseghaier Koc University Recency How long ago the customer last made a purchase? Frequency how many purchases the customer has made? (within a given time period) Monetary How much each customer has spent in total? (within a given time period) 2 1
Start with N-tile Analysis Pick a variable of interest (e.g. recency, frequency, monetary) Sort or rank the database from best to worst on that variable 3 N-tile Analysis For deciles, divide into 10 groups of equal size: top group is decile 1, next is decile 2, etc. NOTE: sometimes the top group is labeled 10 and the bottom is labeled 1. It doesn t matter as long as you know which is which! (and are consistent in whether a 1 is best or worst) You can specify quintiles (divide d into 5 groups) or any other n-tiles (divide into n groups) 4 2
RFM: Putting it all together Combine the recency, frequency and monetary information into an RFM index At least three approaches for the three magic words Independent N-tile approach Sequential N-tile approach Intuitive groupings approach 5 Independent N-tiles Recency Frequency Monetary RFM 1 1 1 111 + + = 1 to 5 1 to 5 1 to 5 1 to 125 groups 6 3
Sequential N-tiles Create quintiles (not deciles!) for recency Within each of the 5 recency quintiles, create quintiles for frequency; that is: Take everyone in recency quintile 1 Sort them from best to worst on frequency Top fifth are in quintile 1 for frequency, etc. Repeat this for each of the 5 recency quintiles Within each of the 25 recency frequency groups, create quintiles for monetary 7 RFM Cells If a customer is in quintile 2 for recency, quintile 1 for frequency and quintile 4 for monetary then RFM cell is 214 8 4
Sequential N-tiles Source: Quick Profits with RFM Analysis, Arthur Hughes. 9 Independent vs. Sequential Approach? Cell Sizes Independent approach yields unequal (often dramatically so) cell sizes Sequential approach yields more nearly equal cell sizes Implicit it Weighting Sequential approach implicitly weights Recency more than Frequency and Frequency more than Monetary 10 5
Independent vs. Sequential Approach? (Absolute) Comparisons Across Different Cells With the independent approach a 231 customer will always have spent more than a 342 customer This is not necessarily true using the sequential n- tile approach (where 1 s designate top quintiles) il 11 Intuitive Groupings Use intuitive breakpoints rather than strict n-tiles 0-6 months One-Time buyers Repeat Buyers Low $ High $ Low $ High $ 7-12 months 1-2 years 2-3 years 3+ years 6
Response Rate by RFM Cell.5.4.3 Mean RESPOND.2.1 0.0 Break - even response rate AAA ABD BAB BBE CAC CDA DAD DDB EAE EDC AAE ADC BBA BDD CBB CDE DBC EAA EBD RFM segment 13 Applying RFM: Different strokes for different folks What marketing program would you propose for these cells? R F M (where 1=best, 5 is worst) 1 1 4 4 1 1 5 5 5 1 5 5 14 7
RFM at Sears Sears RFM model has 189 cells, tracking recency on a quarter by quarter basis Found that each quarter between 20,000 and 30,000 customers acquire a Sears card, buy once, and never buy again led them to shift focus from acquisition to greater spending 14% of customers spend $2500 or more contributing 50% of total merchandise revenues over the past five years 15 RFM is used for customer selection segmenting and targeting Appropriate for existing customers It is based on the premise that past behavior is predictive of future behavior 16 8
Applying RFM: Test Campaigns Test a marketing campaign (e.g. direct mail, email, phone) Choose a random sample from the total (representing all RFM cells) How large a sample? Rule of thumb: average # responses/cell should be > 4 4/response rate = number needed to mail per cell For example, if response rate is 2%: 4/.02 = 200 (times 125 RFM cells = 25,000) Send the mailing, email, etc. Applying RFM: Test Campaigns Measure response by cell It is common to discount the test response rates to reflect the fact that most rollouts don t do as well as the tests (which tend to be carefully executed) if discounting by 15%, then a test response rate of 5.00% would be discounted to 4.25% Select which RFM cells (from full customer base) to receive offer typically select those with response rate greater than breakeven or with positive profits 18 9
RFM Summary Pros: simple, effective, intuitive, flexible, anyone can do it does not require sophisticated software or analysts Cons: limited to R, F, and M where there may be other important predictors of response which more sophisticated models could incorporate no apparent ordering of the RFM cells (i.e. is 123 better or worse than a 132?) 19 RFM is useful, but... Not very sophisticated i Clearly other variables are also important How much better would a more complex, custom designed model be? Logistics Regression? 20 10
A Few More Things on RFM 21 Extensions RFMP where P stands for product May need to sort by product category before RFM Dell Computer estimates share of wallet by customer for desktops, servers, etc. Or by brand/type ex: PC vs. MAC user - don t want to send PC offerings to your best MAC customers 22 11
RFM Migration: How FedEx uses RFM FedEx uses a variation of RFM, emphasizing monetary and frequency, to score customers for each 6 month period Recency = 2 if last purchase is past 3 months, 1 otherwise Frequency = # shipments during 6 month period Monetary = total monetary value during 6 month period Multiply to get RFM score Categorize into deciles Track migration over 4 time periods (2 years) 23 RFM Migration: How FedEx uses RFM Used cluster analysis to identify 7 groups with common behavior patterns and designed tailored marketing programs Top 10% stable customers Medium value 6 month lapsed customers Low value seasonal shippers High value growing shippers Medium value stable shippers Low value 6 month lapsed customers Low value reactivated customers 24 12
Variations Number of n-tiles may vary With relatively l small customer databases - quintiles may be too many For very large databases - 125 cells may be too few Don t have to be equal could have 5 categories for recency, 3 for frequency and 4 for monetary It isn t always RFM National Business Furniture pays little attention to frequency since furniture is a long-lasting product For some b2b markets, monetary is particularly important because of the large variations in spending between small and large customers It isn t restricted to RFM For websites could be R, F, and D (for duration of visit) Measuring R, F and M For example: M could be margin or profitability rather than spending 25 Return on Investment From customer value to firm value 26 13
The Profit Chain WHAT YOU GET FIRM VALUE WHAT CUSTOMERS DO CUSTOMER PROFITABILITY WHAT CUSTOMERS FEEL CUSTOMER EXPERIENCE WHAT YOU DO INTERNAL RESOURCES 27 CLV helps in better Customer Selection 32% customers were unprofitable - Niraj et al. (2001) based on 650 customers of a distributor Top 30% cross-selling gprospects p have predicted usage probability of greater than 80% - Kamakura et al. (2003) based on 5,500 customers of a Brazilian bank Revenue from top 30% customers based on CLV model was 33% higher than the top 30% customers selected based on RFM model - Reinartz & Kumar (2003) based on 12,000 catalog customers Profit from top 5% customers as per CLV model was 10-15% higher than the profit from top 5% customers from other models - Venkatesan and Kumar (2004) based on 2,000 B2B customers 28 14
and helps in better resource allocation Source: Thomas, Reinartz and Kumarv (2004), Getting the Most Out of All Your Customers, Harvard Business Review, July-Aug, 116-123. 29 The Benefit of CLV Models 530% % ROI Field test on 3 groups of 24,000-50,000 customers Based on Bank s heuristic Based on purchased list - Knott, Hayes & Neslin (2002) -17% -30% Based on Cross-selling model 30 15
CLV provides a good proxy for Firm Value $ Billions 18 16 14 12 10 8 6 4 2 0 15.85 14.08 11.00 5.36 3.35 0.82 1.62 1.40 1.89 2.69 Amazon Ameritrade Capital One E -B ay E *Trade Customer Value MarketValue (as of March 2002) Source: Sunil Gupta, Donald R. Lehmann, and Jennifer Stuart (2004), Valuing Customers, Journal of Marketing Research, February, 7-18. 31 Tuscan Lifestyles Targeting Customers with RFM Analysis 16
How many bought and how much? Bought from last catalog? Valid no yes Total Cumulative Frequency Percent Valid Percent Percent 94180 97.5 97.5 97.5 2371 2.5 2.5 100.0 96551 100.0 100.0 Case Summaries Dollars ordered from last catalog Bought from last catalog? N Mean Minimum Maximum no 94180.00 0 0 yes 2371 104.24 5 6249 Total 96551 2.56 0 6249 33 Decile Analysis: Response Rates Recency:.06 Response Rate by Recency Decile.05 Frequency: Response Rate by Frequency Decile Mean Bought from last catalog.04.03.02.01 0.00 1 2 3 Recency Decile 4 5 6 7 8 9 10 Monetary:.06 Response Rate by Monetary Decile.06.05 Mean Bought from last catalog.05.04.03.02.01 0.00 Mean Bought from last catalog.04.03.02.01 0.00 1 2 3 4 5 6 7 8 9 10 1 2 Frequency Decile 3 4 6 9 Monetary Decile 34 17
Decile Analysis: Average $ Order from purchasing customers Recency: Frequency: Mean Dollars ordered from last catalog Mean Dollars ordered from last catalog Avg. Purchase Amount by Recency Deci 140 130 120 110 100 90 80 70 60 50 40 30 20 10 0 1 2 3 4 5 6 7 8 9 10 Monetary: Recency Decile Avg. Purchase Amount by Monetary Dec Avg. Purchase Amount by Frequency De 140 140 130 120 110 100 90 80 70 60 50 40 30 20 10 0 1 2 3 4 6 9 Frequency Decile Monetary Decile Mean Dollars ordered from last catalog 130 120 110 100 90 80 70 60 50 40 30 20 10 0 1 2 3 4 5 6 7 8 9 10 35 Profits from mailing to all 96,551 customers Case Summaries Dollars ordered from last catalog Bought from last catalog? N Mean Minimum Maximum Sum no 94180.00 0 0 0 yes 2371 104.24 5 6249 247160 Total 96551 2.56 0 6249 247160 Gross Profit $ = 2,371 ($104.24)(.5) 96,551($1) = $247,160 (0.5) - $96,551 = $123,576.52 - $96,551 = $27,025.52 Gross Profit as % of sales = $27,025.52/$247,160 =.11 or 10.93% Return on marketing expenditures = $27,025.52/96,551($1) =.28, or 27.99% Breakeven = cost to mail/net revenue from sale = $1/$104 24( 5) = 36 18
Breakeven response rate = 1.92%.10.08 m last catalog.06.04 Mean Bought fro.02 0.00 111 133 155 232 254 331 413 435 522 544 122 144 221 243 315 352 424 511 533 RFM1 (sequential ntiles).0192 37 Sequential RFM Profitability Using sequential n-tiles approach profitable to target? no yes Total N Sum % of Total Sum N Sum % of Total Sum N Sum % of Total Sum BUYERS NUMCUST 40 40 602.00 43468 25.4% 45.0% 70 70 1769.00 53083 74.6% 55.0% 110 110 2371.00 96551 100.0% 0% 100.0% 0% Gross Profit $ = 1769 ($104.24)(.5) - 53083($1) = $92,200.28 $53,083 = $39,117 Gross Profit as % of sales = $39,117.28/$184,400.56 =.21, or 21.21% Return on marketing expenditures = $39,117.28/53083 =.74, or 73.69% 38 19
For comparison, Independent RFM Profitability Using Independent n-tiles approach profitable to target? no yes Total N Sum % of Total Sum N Sum % of Total Sum N Sum % of Total Sum BUYERS NUMCUST 44 44 571.00 41660 24.1% 43.1% 54 54 1800.00 54891 75.9% 56.9% 98 98 2371.00 96551 100.0% 100.0% Gross Profit $ = 1800 ($104.24)(.5) - 54891($1) = $93816 - $54891 = $38,925 Gross Profit as % of sales = 38,925/187,632 =.21, or 20.75% Return on marketing expenditures = 38,925/54891 =.71, or 70.91% 39 Targeted vs. Untargeted Mailing Profitability Analysis Mass Mailing Targeted: Sequential RFM Targeted: Independent RFM Number of Customers 96,551 96,551 96,551 Number Mailed 96,551 53,083 54,891 % of Customers Mailed 100.0% 55.0% 56.9% Number of Orders 2371 1769 1800 Response Rate 2.46% 3.33% 3.28% Gross Revenues $247,153 $184,401 $187,632 COGS $123,577 $92,200 $93,816 Mailing $96,551 $53,083 $54,891 Gross Profit $27,026 $39,118 $38,925 Gross Profit/Sales 10.93% 21.21% 20.75% Return on Marketing 27.99% 73.69% 70.91% 40 20