Chapter 3 Literature Review
3.1 On customer portfolio
3.2 On attrition
3.3 On customer value
In recent years, customer portfolio management (CPM) has focused on optimizing clients’ value to the firm. The company’s interest lies in knowing how much net benefit it can expect from a customer today. These expectations are then used to implement efficient marketing strategies to get the highest return on investment. To that end, two key metrics are estimated by firms, namely customer lifetime value (CLV) and customer equity (CE) (see chapter 2 for definitions).
According to Blattberg and Deighton, CLV can be defined as the revenues derived from a customer minus the cost to the firm for maintaining the relationship with the customer over time (Blattberg and Deighton 1996). As shown by Reinartz and Kumar, CLV modelling depends on the type of relationship a firm has with its clients (Reinartz and Kumar. 2003). In a contractual relationship, customer defections are observed which means that longer lifetime means higher customer value. Conversely, when the relationship is non-contractual, uncertainty arises between the customer’s purchase behavior and lifetime.
With the development of data collection tools, companies have lots of customer-level data (or customer transaction data) at their disposal to measure CLV (Fader and al. 2005). Consequently, different modelling approaches can be adopted in order to estimate customers’ value.
RFM (Recency Frequency Monetary) models are considered the simplest strategy to increase CLV and customer’s loyalty (Gupta, Hanssens, and Hardie 2006). It aims at targeting specific marketing campaigns at specific groups of customers to improve response rates. RFM models consists in creating clusters of clients based on three variables:
- recency which is the time that has elapsed since a customer’s last activity with the firm;
- frequency that is the number of times a customer transacted with the brand in a particular period of time;
- monetary value of customer’s prior purchases.
However, RFM models have a limited predictive power since they only predict clients’ behavior for the next period.
In their article on CLV management, Borle and Singh draw the review of more advanced modelling techniques that can be implemented to estimate customers’ value (Borle and Singh 2008). A popular method to estimate customer lifetime value is the negative binomial distribution (NBD) - Pareto (Fader and al. 2005) which helps solving the lifetime uncertainty issue. The model takes past customer purchase behavior as input such as the number of purchases in a specific time window and the date of last transaction. Then the model outputs a repurchase probability as well as a transaction forecast for each individual. In Borle and Singh’s research paper, a hierarchical bayesian model is implemented with a view to jointly predict customer’s churn risk and spending pattern. Here, the main advantage of using a bayesian approach is to give priors on CLV’s drivers. The study is base on data coming from a membership-based direct marketing company where firm/client relationships are non-contractual. In other words, the times of each customer joining the membership and terminating it are known once these events happen. Thus the implementation of a sophisticated estimation strategy is justified.
In our study, emphasize is placed on estimating the overall value of a customer portfolio. The methodology we will develop is based on a research paper written by our Econometrics teacher Alain Bousquet, whose goal is provide tools for an efficient management of a patent portfolio (Bousquet 2021). The main idea is to consider each patent as an asset with a related value which can generate income if this very patent is exploited. This modelling approach can be transposed to customer portfolio analysis with the customer’s value corresponding to the CLV and the probability of exploitation being the opposite of the risk of attrition. In this context, CLV can be estimated with techniques mentioned above. The client’s risk of churn can be modelled with duration models or machine learning techniques as evoked in 3.2.