1530 track 1 fader_using our laptop
2 likes325 views
This document discusses the use of machine learning for predicting customer lifetime value (CLV). It argues that while machine learning is well-suited for classification and descriptive tasks, it falls short for long-term CLV prediction because it tries to explain all customer behavior patterns. Instead, CLV models should embrace the inherent randomness in customer actions. The document then presents the standard new view of using CLV models to make forecasts, and then applying machine learning to explain differences across customers based on those predictions. It provides examples of layering machine learning on top of CLV models for B2C and B2B customers.
1530 track 1 fader_using our laptop
- 1. Customer Lifetime Value in a Machine Learning World Presenter: Peter Fader Professor of Marketing at The Wharton School of the University of Pennsylvania Co-Founder of Zodiac Author of Customer Centricity: Focus on the Right Customers for Strategic Advantage
- 2. 2 Machine learning is HOT right now!
- 3. 2 Classic Use Cases for Machine Learning
- 4. 4 Use Case 1: Classification Yes or no questions or a number of different “buckets”: Which one does this person belong in? Examples Is this person going to be a good credit risk? Is that person likely to vote for our candidate? Will this customer churn or stay with us next year?
- 5. 5 Use Case 2: Description / Profiling By looking at a lot of different descriptive variables, find which ones do the best job of explaining which bucket a person should be in. Examples Will social media usage or website visitation be a better predictor of brand preference? Will hobbies or income level be a better predictor of spend in a product category?
- 6. Ideal Tasks for Machine Learning Will customer X make a purchase next quarter or not? Will customer X churn next year or not? Will customer X’s next category purchase be with us or a competitors’ brand? Does customer X tend to be a high or low spender for her purchases? When will the next purchase for customer X likely occur? How long will it probably be until customer X churns? How many purchases is customer X expected to make over the next three years? What is the expected spend (in dollars) for customer X’s future purchases? Where Machine Learning Falls Short 6
- 7. If the goal is to ask long-run questions (and that’s what lifetime value is all about) then don’t start with machine learning. What do you do instead?
- 8. Embrace the “Randomness” of Customer Behavior • Don’t try to explain away every little wiggle and jiggle in a purchase pattern • Aim to uncover the true/underlying/unobservable propensities that drive the observed behavior • Capture the inherent variation across people and over time • Result: better forecasts and diagnostics than a model that focuses on explanation 8
- 9. Looking at Regression 𝛽0 + 𝛽1 𝑋1 + 𝛽1 𝑋1+ 𝛽2 𝑋2+ 𝛽3 𝑋3 + … 9 Customer Spend ε ? ?𝑡 𝑖=
- 10. Embrace the “Randomness” of Customer Behavior • Don’t try to explain away every little wiggle and jiggle in a purchase pattern • Aim to uncover the true/underlying/unobservable propensities that drive the observed behavior • Capture the inherent variation across people and over time • Result: better forecasts and diagnostics than a model that focuses on explanation 10
- 11. Embrace the “Randomness” that Always Exists in Customer Behavior • Don’t try to explain away every little wiggle and jiggle in a purchase pattern • Aim to uncover the true/underlying/unobservable propensities that drive the observed behavior • Capture the inherent variation across people and over time • Result: better forecasts and diagnostics than a model that focuses on explanation 11
- 12. The New Standard View of CLV 12
- 13. The New Standard View of CLV 13
- 14. Where Machine Learning Fits into CLV • We don’t use machine learning for forecasts • We use it once we’ve already done our CLV prediction. Once we have these forward-looking projections, now we can put people in buckets and find how they are the same or different from each other. • Machine learning approaches explain cross-sectional differences 14
- 15. 15 Layering on Machine Learning- B2C
- 16. 16 Layering on Machine Learning- B2B 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% NAIC3 Desc. Specialty Trade Contractors NAIC3 Desc. Professional Scientific & Technical NAIC3 Desc. Chemical Manufacturing NAIC3 Desc. Transportation Equipment Manufacturing NAIC3 Desc. Other Month 3 NAIC3 Desc. Nonmetalic Mineral Product Month 2 NAIC3 Desc. Fabricated Metal Product NAIC3 Desc. Mining Except Oil & Gas CTM Size Desc. Unknown NAIC3 Desc. Merchant Wholesalers Durable Goods Quarter 4 NAIC3 Desc. Machinery Manufacturing Weekday Monday Quarter 3 Weekday Thursday Weekday Wednesday NAIC3 Desc. Unknown CTM Size Desc9 701k First Trans Units CTM Size/Desc.7 200k - 700k CTM Size/Desc.5 71k - 200k First Trans Spend First Trans Cost CTM Size/Desc. 21k - 70k Attribute Importance CTMAttributes CTM Attribute Relevance in Early Prediction of High LTV
- 17. 17 Can You Feed the Machine? A machine learning model requires many descriptive variables. • Can you really trust this data? • Can you get it over the long term, i.e. the next 5 years? • Can you link all these variables together at the customer level? • Where should these variables fit in the CLV modeling process? Location Demographics Website Behavior Social Media
- 18. RFM: The Earliest Machine Learning Example? Direct marketers in the 1960’s-70’s invented early data mining/machine learning They compared explanatory factors to see which were most predictive of actual activity (Recency Frequency Monetary Value) Let’s not reinvent the wheel—RFM works!
- 19. The right CLV platform allows organizations to predict customer behavior on an individual level and explain
- 20. 20 Explain (Machine Learning) Predict (Probability Models) Which customers will come back How long the relationship will last How much they’ll spend Which products the best customers bought first Which acquisition channel produced the best customers Which factors best predict churn