Predictive Custom er Relationship M anagem ent
Transcription
Predictive Custom er Relationship M anagem ent
Predictive C ustom er R elationship M anagem ent Profiting from R eally G etting to Know YourC ustom ers H arlan C row der PrincipalScientist H ew lett-Packard Laboratories Palo Alto,C alifornia M ay 19-21,2002 M ontréal,C anada Agenda What is predictive CRM? Predictive CRM toolbox An example: Mail stream optimization at Fingerhut Critical success factors for predictive CRM Conclusion and discussion 2 Technology and applications are driving predictive CRM Computer power: Computational explosion (Moore’s Law in the new economy) 1963 CRM: 600 customers + room-sized, $1M+ computer +0.2MB, 0.5MHz CRM data sources supermarket scanners, in-house TV show ranking POS, stock trades, phones, direct mail, credit cards, prod. and service registrations, Internet transactions . . . 3 Hans Moravic, “When will computer hardware match the human brain?,” Journal of Evolution and Technology 1 (1998) Technology and applications (cont.) Product and service customization Mass production Henry Ford (1920) “Any color as long as it’s black” Mass customization IBM (1972) One box, many uses Individual customization BMW (2002) 90% build-to-order 5 Predictive CRM What do customers want? Can we identify new opportunities? And how do you use information to market and sell? Three problems: Classification of customers into groups based on similar behavior toward a given set of marketing and sales actions -customer segmentation Describing customer behavior by building a behavior model and estimating its parameters Deciding which marketing actions to take for each segment and then allocating scarce resources to segments in order to meet specific business objectives 6 Predictive CRM toolbox Dynamic segmentation Product affinity and bundling Adaptive testing Optimization of resources Dynamic pricing For more information, see H. Crowder, et al., “Predictive Customer Relationship Management: Gaining Insights About Customers in the Electronic Economy,” DM Review, February 2001. 7 Dynamic segmentation Historical basis: homogeneous clustering; static, infrequent updates On-line environment has changed the process: Immediate feedback, testing Constantly shifting segment boundries Seg. criteria, rules evolve Discrete process is now continuous Observe/ measure Segment Test Actions 8 Product affinity and bundling “beer and diapers” Customer choices can suggest affinity of products and services Fast data mining methods allow sophisticated processing of large data sets; identification of sequential affinities Mobile devices plus pervasive computing infrastructures opens opportunities for “location-aware” product and service bundles, and new branding mechanisms 9 Adaptive testing Estimating conversion rates was easier for mature products with low-variability forecasts For new, short-lifecycle products, a more rigorous estimation process is needed find the smallest sample size for which the estimate is statistically accurate within specific parameters At HP, some mobile computing products have 3-6 months lifecycle; little historical preference data Adaptive testing techniques: dynamic sample size based on recent events Bayesian approaches allow combination of data from various sources 10 Resource optimization Resources include marketing budgets, product availability and advertising space Measures of effectiveness of business objectives: profit, conversion rate, market share, customer acquisition The goal is allowing planners to analyze trade-offs for various decisions, e.g., sensitivity and investment analysis Methods include mathematical optimization, dynamic programming, meta-heuristics such as genetic algorithms and simulated annealing 11 Dynamic pricing In many areas, the Internet has started making the price-demand relationship more transparent For example, recovering historical data from online auctions such as eBay allow computation of price-elasticity relationships 12 Mail Stream Optimization (MSO) Helping direct mail retailers send less junk-mail and make more money 13 About Fingerhut 50 year history 2nd in consumer catalog sales general merchandise 14 The data 65 mil. customers on file, 7 mil. active 9 terabytes of data on customers, products and promotions Large IBM mainframe relational databases 3000 customer attributes purchase history payment history mail history demographics 100 catalog attributes content presentation mail dates 15 The business problem The business model didn’t scale increasing customer base increasing mail quantities increasing mail dates The vertical marketing model was wrong inhibited profitability ignored the customer promotions Bottomline: line: Bottom Fingerhutwas wasmailing mailingtoo toomany many Fingerhut cataloguesthat thatwere wereending endingup up catalogues theircustomer’s customer’strash trashcans; cans; inintheir excesscosts costswith withno norevenue. revenue. excess customers time 16 The solution: Horizontal Marketing -Optimize customer contacts over time promotions Segmentation Segmentation Budget Budget Allocation Allocation Customer Customer Scores Scores customers Optimization Optimization Saturation Saturation time Refs: H. Crowder, et. al., “Optimizing customer mail streams at Fingerhut,” INTERFACES (31,1) 2001 System and Method for Increasing the Effectiveness of Customer Contact Strategies, US Patent Pending Determine which promotions for each customer, NOT which customers for each promotion Optimize overall customer value, NOT value for each promotion 17 MSO predictive CRM problems Terminology Customer segmentation creating portfolios of investments Saturation mail plans and mail streams measuring the cost of junk mail Valuing customer contacts using optimization What’s my ROI for a customer? 18 MSO terminology mail plan The schedule of promotion mailings to be executed over a specified time period. For example, Fingerhut normally scheduled about 60 mailings in a three month period. mail stream For a given customer and a mail plan, a mail stream for the customer is that subset of the promotions in the plan that will be sent to the customer. For example, if a mail plan has 60 mailings, then a customer’s mail stream could be the 1st, 10th, 18th 33rd, and 59th promotions in the plan. 19 Customer segmentation for investment allocation Growth invest for the future High-value – max. investment Initial customer segmentation criteria – “RFM” • • • Recency Frequency Monetary value $ Profit • Easy to obtain this data • Provides broad investment classes • Precursor to more detailed customer purchase history segmentation $ Revenue “Up or out” min. investment 20 MSO saturation Saturation is the measure of how promotions in a mail plan adversely effects the revenue of other promotions in the plan. S Saturation is caused by: time proximity product similarity presentation sales similarity 0.25 time Sij = fraction of promotion j’s sales lost due to promotion i being mailed to the same customers. 21 What is the value of a mail stream? For a customer (segment) and a mail plan: Rp Ep S p, p' α yp expected revenue of promotion p production and mailing cost of promotion p saturation of promotions p and p’ control parameter mail stream indicator switch: 1 if promotion p is in the customer’s mail stream 0 otherwise Then the value of the mail stream indicated by {y} is ∑ (R p p − E p ) y p − α ∑ Rp y p S p, p' y p' p, p' 22 Generating good mail streams within a specified expense budget If the amount of promotion expense for a customer or segment is limited by a budget, then the following nonlinear binary mathematical programming problem produces the best mail stream for the expense budget: Determine {y} in the expression Maximize ∑ (R p p − E p ) y p − α ∑ R p y p S p, p' y p' p, p' such that ∑E p yp ≤ B p where B is the expense budget limit The Mail Stream Generation Model 23 Mail stream optimization approach Customer segmentation primary parameters: revenue and profit Saturation effects of mail plans key to eliminating junk mail Allocation of advertising resources via portfolio optimization techniques using profit expectations from segmentation Optimal mail streams from MSG Overall master mailing plan using optimization 24 The MSO solution Run once a week considers past 3 months and projects next 3 months Processing time: 12 hours The details 10,000 binary linear optimization problems (MSG) 20,000 variable linear programming problem (master mail planning model) IBM SP2 with 4 processors IBM Optimization Subroutine Library, SAS, C++ 3 year effort 25 MSO financial impact $3.5 million annual profit gain 6% advertising cost decrease 82% decrease in measured saturation of mailings 21% increase in new customer response rate First year ROI for the project Non-financial: lot of happy trees!! 26 Expense-revenue balance Where was Fingerhut? Revenue “profit” “over-mailing” “opportunity” Expense 27 Critical success factors for predictive CRM initiatives It’s about designing and adopting new business processes, not installing new technology. It’s a destination and a journey. It’s built on numerically-intensive methods. Numbers mean data. Budget accordingly. Plan a phased rollout. Small is beautiful (and manageable). Monolithic predictive CRM initiatives fail. 28 Conclusion Predictive CRM draws on a variety of disciplines in computer science -- data mining, database management and data visualization -- economics, operations research and mathematics. And, oh yeah, PROGRAMMING!! Most electronic economy businesses never see their customers; they only see data about those customers Predictive CRM is allowing businesses to increase the personalization in a previously impersonal process 29 Acknowledgement Special thanks to: Jamie Dinkelacker, HP Labs Meichun Hsu, CommerceOne Labs Shailendra Jain, HP Labs 30 Contact information Harlan Crowder Principal Scientist Hewlett-Packard Laboratories Palo Alto, CA USA harlan_crowder@hp.com 31