Joao Azevedo
March 17, 2016
820

March 17, 2016

## Transcript

1. ### Modeling Customer Lifetime Value Joao Azevedo Tech Tech Talks @jcazevedo

/ joao@shiftforward.eu March 17, 2016 Joao Azevedo Modeling CLV
2. ### Introduction CLV Models Resources Outline 1 Introduction Customer Lifetime Value

An initial formulation 2 CLV Models RFM Markov Models Supervised learning Rules of thumb Pareto-based models Spend models 3 Resources Joao Azevedo Modeling CLV
3. ### Introduction CLV Models Resources Customer Lifetime Value An initial formulation

Purpose of this talk Give a brief introduction to the current state of Customer Lifetime Value modeling. Be practical. Present mathematical descriptions only when they are important from an implementation perspective (references are provided for more details). (Hopefully) enable one to bootstrap CLV modeling on his/her own dataset. Joao Azevedo Modeling CLV
4. ### Introduction CLV Models Resources Customer Lifetime Value An initial formulation

Customer Lifetime Value The net present value of the proﬁts linked to a speciﬁc customer once the customer has been acquired, after subtracting incremental costs associated with marketing, selling, production and servicing over the customers lifetime. Robert C. Blatt, Kim Byung-Do & Scott A. Neslin “Database Marketing” International Series in Quantitative Marketing (2008). Joao Azevedo Modeling CLV
5. ### Introduction CLV Models Resources Customer Lifetime Value An initial formulation

Use cases Adjust the marketing expenditure. Target marketing activities to avoid clients leaving. Rank users for speciﬁc marketing campaigns. Joao Azevedo Modeling CLV
6. ### Introduction CLV Models Resources Customer Lifetime Value An initial formulation

An initial formulation CLV = ∞ t=1 E[ ˆ Vt] (1 + δ)(t−1) Only needs: 1 The probability of a client leaving in a given period. 2 The discount multiplier of capital for the given period. 3 Expected revenue per client. 4 Cost of relationship maintenance. Joao Azevedo Modeling CLV
7. ### Introduction CLV Models Resources Customer Lifetime Value An initial formulation

Joao Azevedo Modeling CLV
8. ### Introduction CLV Models Resources Customer Lifetime Value An initial formulation

It’s already possible to predict the future based on past data, but: All clients are treated the same. All portions of time are considered the same. Joao Azevedo Modeling CLV
9. ### Introduction CLV Models Resources RFM Markov Models Supervised learning Rules

of thumb Pareto-based models Spend models RFM RFM models present a way to group data based on the metrics: Recency: period since last purchase. Frequency: how many purchases an individual made during the observation period. Monetary value: cumulative total spent by client during observation period. Allow us to quickly segment clients. Requires at least 1 period of activity from a client (it’s possible to incorporate other features and predict the cluster the client belongs to). Joao Azevedo Modeling CLV
10. ### Introduction CLV Models Resources RFM Markov Models Supervised learning Rules

of thumb Pareto-based models Spend models Markov Models Model the discrete time periods since last purchase as states. Each state has a given probability for a client to make a purchase. Robert C. Blatt, Kim Byung-Do & Scott A. Neslin “Database Marketing” International Series in Quantitative Marketing (2008). Joao Azevedo Modeling CLV
11. ### Introduction CLV Models Resources RFM Markov Models Supervised learning Rules

of thumb Pareto-based models Spend models Supervised learning Treat the probability of a client leaving as a straightforward supervised classiﬁcation problem. Requires the deﬁnition of the set of discrete points at which the client can leave. Joao Azevedo Modeling CLV
12. ### Introduction CLV Models Resources RFM Markov Models Supervised learning Rules

of thumb Pareto-based models Spend models Rules of thumb In some cases, basic rules of thumb have a better predictive performance than some more sophisticated models: If client has not purchased in x Months then they will not repurchase. Management heuristics should always be taken into account in our suite of models and should be statistically compared to other approaches. Markus W¨ ubben, Florian V. Wangenheim “Instant Customer Base Analysis: Managerial Heuristics Often Get It Right” Journal of Marketing, Vol. 72, No. 3. (2008), pp. 82–93. Joao Azevedo Modeling CLV
13. ### Introduction CLV Models Resources RFM Markov Models Supervised learning Rules

of thumb Pareto-based models Spend models Non-contractual setting Most approaches to modeling CLV in non-contractual settings split the problem in two: 1 Probability of the user being “alive” or “dead”. 2 Probability of purchase(s). Initial proposal was on using a Negative Binomial Distribution but most modern approaches use two separate families of distributions: Pareto/NBD, BG/NBD, BG/BB... Andrew S. C. Ehrenberg “The pattern of consumer purchases” Applied Statistics, (1951), pp. 26–41. Joao Azevedo Modeling CLV
14. ### Introduction CLV Models Resources RFM Markov Models Supervised learning Rules

of thumb Pareto-based models Spend models Joao Azevedo Modeling CLV
15. ### Introduction CLV Models Resources RFM Markov Models Supervised learning Rules

of thumb Pareto-based models Spend models Spend models Integration of the RFM paradigm with CLV using a submodel for spend per transaction. The gamma-gamma model is typically used to model expenditure: the value of a transaction varies randomly around the customer mean transaction value. Peter S. Fader, Bruce G.S. Hardie, and Ka L. Lee “RFM and CLV: Using Iso-Value Curves for Customer Base Analysis” Journal of Marketing Research, Vol. 42 (2005), pp. 415–430. Joao Azevedo Modeling CLV
16. ### Introduction CLV Models Resources Lifetimes https://github.com/CamDavidsonPilon/lifetimes BTYD https://cran.r-project.org/web/packages/BTYD/index.html spark.mllib’s optimization

methods https://spark.apache.org/docs/latest/mllib-optimization.html Joao Azevedo Modeling CLV